Robinhood launched a blockchain in July to let people trade tokenised stocks on-chain, and the first wave of real traction on the chain came from memecoin traders. That might sound like a product-positioning failure, but surprisingly, it became the best go-to-market strategy ever for on-chain wrapped assets.

The memecoins were paired with the stock tokens in shared liquidity pools, and enough degen demand poured in that over half the supply of a real healthcare company’s stock got locked up as collateral for a cat memecoin. Memecoin speculation became the liquidity engine for a real equity.

Every asset that matures into a financial product follows a very similar path. People pay attention to it, then someone builds a way to access it, and then the financial system figures out how to own and structure it. Gold drew attention for five thousand years before the GLD ETF wrapped it in an access layer in 2004. Bitcoin ETFs are doing it for crypto now. On-chain, these layers are compressing and sometimes arriving out of order, reshaping what the end product looks like!

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

The Financial Layers and the Trust Problem

If you look at every major institution building tokenisation infrastructure today, they are all doing what State Street did in 2004 when it gave gold a ticker symbol. State Street launched GLD, the first US-traded gold ETF, which let anyone with a brokerage account buy gold exposure without dealing with vaults, insurance, or physical bars.

BlackRock, JPMorgan, Nasdaq, and Franklin Templeton are positioning themselves as the translator between an asset class and the capital that wants exposure to it, because the translator permanently captures the flow. State Street has collected fees on GLD for over 20 years.

The entire modern history of Wall Street is a history of owning these translation points, from JP Morgan brokering the bridge between European capital and American railroads, to the investment banks of the 1980s owning the bridge between mortgage borrowers and bond investors. Whoever encodes an asset into the language that institutional money already trusts gets to charge tax on every dollar that crosses it, for as long as the encoding remains the standard. Which is why the race to tokenise is really a race to become the next encoding monopoly.

The encoding itself is no longer the hard part. On-chain, anyone can tokenise anything. The monopoly has moved one layer up, from who can encode the asset to who can make the encoding trustworthy enough for institutional money to touch it. That is what Securitise does for BlackRock, what the KYC gate does for BUIDL, and what no memecoin has ever needed.

The GLD ETF is the simplest example of how that encoding works and what it costs, because gold already had the strongest possible attention layer, five thousand years of universal consensus that it was the hardest money on the planet, and yet if you ran a pension fund in 2003, your options for getting exposure were only as sophisticated as renting a safe deposit box with physical bars.

What GLD did in November 2004 was build the access layer by performing a trust translation. It took gold’s five-thousand-year reputation and re-encoded it into something a brokerage account could hold by listing gold on the NYSE, establishing a clearing relationship with DTCC, and reclassifying it from “operationally impossible” to “one click in your brokerage account,” and moving a billion dollars in three trading days.

Central banks that had been net sellers of gold for decades now reversed course, and seven years later, GLD briefly became the largest ETF on Earth, holding 723 metric tons of physical gold. But every trust translation is lossy, and the losses compound as you build higher.

Converting gold into a ticker meant surrendering the property that made it valuable for years: you could hold it, and no counterparty, government, or intermediary could touch it. GLD made gold investable by embedding it inside the very financial system it was supposed to be insurance against, and then the ownership layer built on top of that, with banks using GLD shares as collateral, other financial products referencing its price, and lending desks borrowing against it, each adding another counterparty between the investor and the actual metal.

A pension fund holding a structured note linked to GLD was not really holding gold in any meaningful sense, it was holding a chain of promises that tracked gold’s price.

Notice that the translator always decides which properties of the original asset survive, and the properties that always get cut are the ones that would make the wrapper unnecessary. If you could hold gold directly in your brokerage account as easily as holding GLD, State Street’s fee stream would evaporate, so the translation has to be lossy, because a lossless one would eliminate the translator entirely.

This is why different layers exist in sequence in traditional finance, and each translation creates dependencies, and each dependency must prove itself trustworthy before the next layer builds on top. For example, once GLD became collateral across thousands of portfolios, the next sequentially important thing was the custodial and legal wrapper rather than the gold itself, because if the wrapper failed, the gold would still sit in the vault while the financial system built on top of it would collapse.

The sequential process also protects incumbents like State Street and BlackRock, because if banks and asset managers could build structured products that referenced gold or Bitcoin directly rather than through an ETF wrapper, the wrapper becomes irrelevant and its fee stream dies. The slow march from attention to access to ownership is presented as prudence, and it partly is, but it also happens to be the optimal strategy for maximising the value captured at each intermediary step.

Bitcoin ETFs compressed this seven-year sequence to about eighteen months, with $57 billion in first-year inflows. Gold survived by being ignored by its attackers, and Bitcoin survived by being attacked, through fifteen years of hostile regulators, exchange collapses, China banning mining and 80% drawdowns that lasted years.

When a compliance committee evaluates whether an asset belongs in an ETF wrapper, they assess downside survivability, and Bitcoin’s record of surviving sustained attacks gave them more conviction faster than gold’s record of never facing one.

But faster does not mean better. Bitcoin’s entire point was that you could hold value without trusting a bank, move it on a weekend, and nobody could freeze your account. The ETF strips all three. You get a ticker and a tax advantage, in return, BlackRock holds your Bitcoin, trades it only during market hours, and can be subpoenaed to freeze your shares. GLD did the same with gold.

The whole system moves slowly because trust takes years to build and seconds to destroy, and every gatekeeper in the chain, from the SEC analyst reviewing an ETF application to the custodian deciding what qualifies as collateral, charges a fee for the time they spend. Some of that caution is genuinely protective, but the slowness also makes them money, because instant, free access would put most of them out of business.

On-chain, none of those gatekeepers exists, and anyone can issue a token, build a liquidity pool, and wrap a position into a composable instrument without asking permission from a single intermediary.

On-chain Is Running the Layers Out of Order

This is where memecoins came in and filled the vacuum. Billions of dollars moved through tokens named after dogs and frogs on infrastructure built for tokenised securities and structured products. In the three-layer model we discussed. Memecoins skipped the middle step entirely. People showed up with attention and capital, and the liquidity infrastructure they created by trading became the access layer for everything that followed.

Robinhood’s RWA share of the chain’s volume rose from 0.39% to 8.58% in three weeks, built entirely on liquidity memecoin traders created. BONKbot, a Telegram trading bot, plugged into the $BONK community on Solana and hit $200 million in daily volume without negotiating a single partnership.

Memecoins have super-extended the velocity of attention and capital by paving a new way to invest in culture. It is the first instrument where you invest in a go-to-market strategy directly rather than a company that has one.

The institutions that showed up were all for defensive reasons. BlackRock listed its BUIDL fund on Uniswap and bought the protocol’s governance tokens to distribute a tokenised Treasury fund to its clients. BUIDL itself is GLD on a blockchain with the same KYC-gated access, and Securitise sits in the middle with about $2.8 billion in AUM across eight chains, deployed on rails that were specifically designed to remove that restricted access.

Even JPMorgan said that tokenised money market funds are a defensive play against stablecoins pulling deposits out of the banking system. The money is already leaving, and the GLD-era encoding monopoly is being rebuilt on-chain because the alternative is losing the fee stream entirely. BlackRock understood long before that permissionless finance would eat its deposits, and instead of fighting the new infrastructure, it rebuilt the old tollbooths on top of it.

And while institutions were recreating the sequential playbook, the permissionless side was building ownership-layer products on assets that had barely finished building their attention layer.

For example, Ethena’s USDe makes money from the very thing that keeps institutions away from crypto: the volatility. A product like this would never survive a single risk committee at any bank. But within two years, Ethena became the foundational collateral for other on-chain protocols, building on it the way banks build on treasuries. Pendle then split Ethena’s yield-bearing token into principal and yield components, each trading separately. Once those principal tokens became collateral on Aave, traders could loop the position five to ten times, turning a 5% return into 33% or more.

In traditional finance, this same recursive loop would have required its own ISDA master agreement, compliance review, and counterparty evaluation. Hyperliquid even showed us that the access layer itself can run in reverse. Traditional finance encodes alternative assets into its institutional language, the way GLD encoded gold. But Hyperliquid encodes traditional assets into crypto’s language. When Iran struck on a Saturday, and oil spiked, the CME was closed. But Hyperliquid priced the shock in real time for the entire market.

Arcus, built by dYdX Labs on Robinhood Chain, collapses all three layers into a single token. A leveraged perpetual futures position packaged as an ERC-20 that is transferable, composable, and works with every other protocol without custom integration. That means you could hold 3x leveraged Bitcoin in your wallet, deposit it as collateral in a lending protocol, borrow against it, and send it to someone else, all without closing the position or asking a broker.

You’ll see traditional finance keep building products that look like BUIDL, because the permission gate is how every institution makes money. Permissionless finance will keep building products where the layers merge, invert, and recurse, because the architecture allows it and nobody has the authority to stop it. Both systems are expanding what a financial product can be, in directions the other cannot follow.

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

A goat rancher in Northern California needed to protect himself against this summer’s wage-rule changes. His entire operation depends on seasonal hired labour, and if the state raised minimum crew wages mid-season, his margins would disappear overnight. There is no insurance policy in the country that covers legislative risk for small livestock operations, and there has never been a futures contract listed for it on any exchange in the world. So Castle Tech, a firm that builds custom derivatives for exactly these kinds of uncovered risks, created a contract tied to whether California’s proposed wage legislation would pass by a specific date, settled the whole thing through Kalshi, and got Susquehanna, one of the largest quantitative trading firms on the planet, to take the other side for half a million dollars.

For nearly two years, conversations about prediction markets have focused on whether Polymarket and Kalshi can forecast elections better than polls. But prediction markets are becoming much more than forecasting tools, and the new financial products growing out of them are starting to touch governance, hedging, and speculation in ways I think nobody is paying nearly enough attention to.

Let’s dig in!

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

What Happens When Governance Becomes a Market

Every prediction market today, from a contract on Polymarket to a listing on Kalshi, produces only one kind of output, a probability. Which is nothing but a number that represents the market’s collective estimate of whether a specific event will actually happen.

Over the past two years, the infrastructure behind this has shown us that markets can aggregate dispersed knowledge about future events more accurately than any polls, expert panels, or any other forecasting mechanism humans have ever built.

But probability is a strangely incomplete piece of financial information. If you hold bitcoin and see a 65% chance the Fed will cut rates by 75 basis points next month, you might learn something tangible about the world and almost nothing useful about your own position. You know the event is likely, but you still have no idea what it would actually do to the price of the asset you’re holding. Would Bitcoin trade at $110,000 in that world, or $95,000, or $140,000? No prediction market can ever tell you that.

There are events, and then there are the impacts those events have on companies and assets, and the financial system has been collapsing those two entirely separate things into one for as long as markets have existed. That difference is exactly where the next generation of financial products is being built.

Two very different things are being built right now. One starts with a question prediction markets made inevitable: if markets aggregate information better than anything humans have ever built, why are we still using votes to run organisations? These are called decision markets. The other is something that has never existed before, a market that actually tells you what Bitcoin would be worth if the Fed cuts 75 basis points, not just the probability of the cut happening, but the dollar price of Bitcoin in that specific version of the future. These are called impact markets, and they’ll be far more important than anyone in the space is giving them credit for right now. I will get deep into impact markets in the next section, but the fastest way to understand why they matter so much is to see exactly where decision markets hit a bottleneck.

Decision markets were first implemented in governance because DAO governance and token voting have had a fundamental flaw for as long as they have existed. People with the most tokens rarely have the most information, and people with the most information rarely have enough tokens to matter. Decision markets invert this by letting anyone with capital express a view, regardless of token holdings. Ultimately, this lets people price information in real time.

Robin Hanson proposed a fix for this in 2000 called futarchy, and the core idea is beautifully simple. Say a protocol needs to decide whether to cut its staking rewards. Instead of a token vote, traders price the protocol’s token in two competing futures: one where rewards get cut and one where they stay, and whichever version of the future trades higher is the one that gets implemented. And if you ever trade that decision without understanding the protocol’s economics, you lose money to whoever did the work, and that loss directly shrinks your weight in every future governance call.

The system gets smarter over time because uninformed capital bleeds out and informed capital accumulates, the opposite of token voting, where you can make the worst possible decision on every proposal for a year and your governance power still stays the same. Futarchy makes ignorance expensive, something no governance system before it has ever managed.

MetaDAO was the first team to actually ship this. They launched on Solana in late 2023 with a binary model where every proposal is a straight pass or fail, and proposals go through conditional markets instead of token votes, and the market’s answer is the one binding.

But the binary design creates another kind of problem. Since the market can only answer yes or no, every parameter (Budget, timeline, technical changes) of the proposal must be negotiated in a forum before the market even opens. This defeats the whole purpose of decision markets because figuring out what the right version of the proposal looks like still happens through the same informal consensus process, where the biggest token holders dominate the conversation.

You have moved the final yes-or-no into a market, which is a real improvement, but you have left all of the design work in the broken system you were trying to replace.

In September, Umia launched a multi-outcome model for the MetaDAO version. So, instead of a single pass-fail question, multiple proposal configurations compete simultaneously, and the market filters out weaker parameter sets through trading. The discovery of what the right version of a proposal should be happens inside the market rather than outside it, which is where it always should have been if the whole point was to let price discovery do the work.

Across all 22 DAOs running this, 96 proposals have gone through market-based governance so far, and the total capital sitting inside market-governed treasuries is $50.5 million, with MetaDAO holding $39.3 million of that. So the trajectory toward real scale is there, but even if treasuries grow by a hundred times from here, decision markets have a ceiling, and it doesn’t even have anything to do with capital or adoption or how good the teams building them are.

A rule in economics called Goodhart’s Law says that when a measure becomes a target, it stops being a good measure. Credit ratings were designed to measure default risk, and when banks started using them to decide which mortgage products to buy, investment banks began engineering securities to meet rating criteria rather than select for actual loan quality, and the ratings kept printing AAA while the underlying mortgages rotted.

Decision markets have this vulnerability baked in. When a proposal passes, the token moves partly because the proposal itself might create value, and partly because passage signals team confidence. A strong project could propose burning treasury funds on a decorative bronze bull, and the token would still trade up on approval because traders react to confidence in governance, not the statue. Apple stock does the same thing every time it announces a buyback at absurd valuations.

The whole thing breaks because the market’s answer feeds back into the thing it is trying to measure, and you cannot fix that without destroying the reason traders show up in the first place. But strip the governance layer out entirely and point the same conditional mechanic at events happening outside the protocol, events that no trading outcome can influence, and that feedback loop disappears on its own. There is nothing left for the price to corrupt against. And that is what an impact market does.

Impact Markets

Hedging in crypto right now is impossibly blunt. If you hold Bitcoin and you’re worried about what a specific regulatory ruling would do to your position, you either sell and give up all the upside from every other scenario, or you buy volatility products priced against general market movement that have nothing to do with the specific event you’re actually worried about. There has never been a financial product that isolates one event and lets you trade its dollar impact on another, and that’s exactly what impact markets let you do.

For example, let’s say the SEC is about to rule on whether a category of DeFi protocols qualifies as securities, and you hold a large ETH position. You believe a negative ruling would crush ETH to $1,800, but the impact market on ETH conditional on a negative ruling is trading at $2,300, which is the market’s collective estimate of where ETH would land if the ruling goes against DeFi.

You sell at $2,300, the ruling drops, ETH falls to $1,800, and you pocket $500 per token because your read on the severity was better than the crowd’s. If you’re wrong and ETH only dips to $2,200, you still pocket $100. A trade like this couldn’t have existed before impact markets, because no product has ever let you trade the magnitude of a specific event’s impact on a specific asset, with a live market pricing your estimate against everyone else’s.

It works because every asset price you have ever looked at is really just an average across all possible futures, compressed into one number, and impact markets are the first product that lets you trade beneath it.

Right now, if you want to protect yourself against a specific risk, you diversify across assets, which is really just a crude guess about which things might go up when others go down. Impact markets let you skip that entirely. You can hold a concentrated Bitcoin position and buy protection against the exact regulatory or macro event you’re worried about, without selling the position and without paying for volatility insurance that covers a thousand scenarios you don’t care about.

Instead of diversifying across assets and hoping correlations work out, you hedge specific scenarios directly, a completely different way to manage risk that only becomes possible when the underlying product gives you a conditional price rather than a probability.

I think impact markets will be far more consequential than prediction markets ever were. A prediction market closes and pays out a binary result, which is useful but, on its own, finite. You can’t build anything financial with that number because a probability isn’t a tradable object.

Impact markets produce a dollar-denominated price for where an asset would actually trade under a specific condition. You can hedge against it and write derivatives on top of it, instruments that literally could not exist before because there was no underlying conditional price to build on. The products that eventually get layered on reliable conditional prices will dwarf the conditional markets themselves, the same way the options and derivatives market is orders of magnitude larger than the equity market it’s written on.

The TAM for “how likely is this event?” is interesting but fundamentally capped, because there’s only so much you can do with it. But the TAM for “what would this event do to this asset?” has far more outcomes and runs as deep as finance itself.

The first implementations are already live and scaling faster than ever. Protocols like Branchpoint are even building something called Branchmap, a shared naming system for how conditional outcomes get defined across protocols, which matters because impact markets only become a real financial layer if “ETH conditional on a negative SEC ruling” means the exact same thing on every platform that trades it. Without that standardisation, liquidity stays fragmented and conditional prices on different platforms aren’t actually comparable.

Branchpoint calls this “multiverse finance.” Every conditional market is a window into one version of the future, one branch of reality where a specific event did or didn’t happen. As impact markets mature and liquidity deepens, what emerges is not just a single conditional price for one event but a full spectrum of an asset’s value across parallel futures, each conditioned on a different scenario.

Finance has tried to decompose asset prices this way for decades. The entire discipline of event studies exists to work backwards through historical data and estimate how much of a price move was driven by one specific event versus background noise, and the best anyone has ever managed is a rough estimate after the fact.

Impact markets produce this decomposition forward-looking and in real time, and that information changes the foundations of how risk gets priced the moment it exists.

The next level is when conditional prices start plugging into risk models and portfolio engines that currently run entirely on spot prices and implied volatilities. Once that integration begins, the framework for capital hedging and exposure in crypto gets rebuilt from the ground up.

But there is no reason this stays just limited to crypto. The goat rancher couldn’t buy insurance because actuarial tables don’t cover goat pasture in a specific Northern California county. In every case, the missing piece has always been a liquid price for what would happen to X if Y occurs.

Once that price exists, capital allocation changes in compounding ways. Once an insurer can lay off risks too weird for its own book into a conditional market, it can underwrite policies it currently has to decline. A VC fund that can hedge sector risk separately from company risk can concentrate more on its best founders instead of diversifying across thirty bets as a substitute for conviction. A government considering a minimum wage increase can watch a conditional market price the employment impact in real time, with real money behind every position, instead of arbitrating between economic models that disagree with each other.

Options did that to equities in the 1970s; it rearranged the industries built on top of them in ways their inventors did not predict. Conditional markets are the same class of primitive. The difference is they decompose risk across events, not just across assets, and the number of events in the world is unbounded.

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

They borrow against the assets and spend the loan. When they die, their heirs in the US inherit the assets at today’s value, so those gains are never taxed as capital gains. Planners call it “buy, borrow, die.”

The habit needs two things, an asset a lender will accept and a lender willing to deal with you. Public stock and property qualify easily. Shares in a private company mostly don’t, unless you’re a founder or a senior executive with a big stake. That leaves out a lot of people.

Borrowing against shares can be risky even for someone extremely wealthy.

When Elon Musk agreed to buy Twitter for $44 billion in April 2022, banks offered him a $12.5 billion loan secured by his Tesla shares.

But Tesla’s share price was falling, which created a problem.

If those shares lost too much value, Musk could be required to provide more shares or cash to keep the loan. He dropped the loan plan and increased the amount he would contribute through other financing. As of August 2025, about 236 million of his 715 million Tesla shares were pledged to secure personal loans, according to Tesla’s proxy filing.

The lenders expect repayment regardless of what happens to the share price, and a fall in value can leave borrowers scrambling for additional collateral. By a startup’s Series C, its employees typically own more of it than its founders do, according to Carta data. These people can’t easily cash out before their company goes public.

Techdollar wants to give them a loan against those shares and, if they’d like, an Apple Pay card to spend it with.

Techdollar launched in June with a $3 million pre-seed round. It lends against vested shares in private companies. They call this “Build. Borrow. Thrive.” Building is up to the borrower, leave that part, and we are going to check whether the other two stay true.

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

Why did the wait get so long?

In 1989 the average company going public was six years old, and by 2021 that had stretched to 12, according to Morgan Stanley’s wealth management research.

According to Jay Ritter, the University of Florida professor who keeps the standard database of US listings, the median age during the dot-com years of 1999 and 2000 was 5. Which became 11 years between 2001 and 2025, and 12 years in 2025 alone.

SpaceX was founded in 2002 and listed in June 2026. Stripe, founded in 2010, is still private.

Big private companies let staff sell through tender offers instead. Stripe has run tender offers at $91.5 billion in February 2025, $106.7 billion in September and $159 billion in February 2026.

During Facebook’s 2012 listing, even David Choe, the artist who accepted Facebook stock for painting its offices, became wealthy when the company went public. But it is different for someone who joins a promising startup, spends eight years building it, and leaves before the company lists. They may have substantial equity and still need to pay a mortgage or fund their next venture using their salary and savings.

In October 2025, OpenAI’s current and former staff sold about $6.6 billion of stock at a $500 billion valuation. The company had authorised up to $10.3 billion, and only about two-thirds of it was tendered.

Anthropic ran its own tender between February and April 2026 at a $350 billion valuation, open to people with at least a year at the company. Investors had lined up about $6 billion, and employees sold less than that, which could be staff expecting the shares to keep rising.

Reuters reported that the value of tender-offer transactions executed through Nasdaq Private Market rose from around $3 billion in 2023 to just under $15 billion in 2025.

So tenders do benefit, but they run on the company’s calendar. The company sets the date, says who qualifies, and sets limits on how much each person can sell. According to Morgan Stanley, companies typically limit these sales to 20–30% of an employee’s vested shares.

Selling outside those tenders is also possible. Employees can approach platforms such as Forge or EquityZen.

Most private companies require board approval for any transfer, and most have a right of first refusal, usually shortened to ROFR. It lets the company, and sometimes its investors, match whatever price you negotiate and buy the shares themselves.

OpenAI and Anthropic treat any unapproved transfer as void. Anthropic barred SPVs(Special Purpose Vehicles) and named new offerings on Forge and Hiive among those it would not recognise.

A sale that clear, becomes a taxable event, and seller also gives up whatever the shares do next. OpenAI staff sold at $500 billion in October 2025. Five months later, the company was valued at $852 billion.

Banks lend against private stock selectively. They usually take only unrestricted shares, which rules out much of what employees hold. Even so, Goldman’s San Francisco private wealth loans are up 50% since 2023, and banks see a loan today as a way to win the IPO later.

What is Tech Dollar?

For founders and employees, the core product is a loan secured by vested shares.

That means, unlike tenders or secondary sales, Techdollar lets employees get cash without selling their shares or losing future gains. Also, unlike traditional banks, which prefer public stocks or private stocks from wealthy investors, Techdollar specialises in lending against private-company shares, though borrowers still need to qualify and pledge shares they’ve already earned.

You keep the shares, but if you fail to repay the loan, the lender may have the right to sell them. You don’t necessarily need substantial wealth beyond those shares to qualify.

The money can arrive in whatever form the borrower prefers. It can be a bank withdrawal, a physical card, an instant digital card in Apple Pay, or in stablecoins.

But unlike its crypto competitors, Techdollar does not tokenise the shares. The pledge is documented in standard legal documents as a registered claim over the shares, known as a share charge, with the accounting firm RSM verifying ownership. It uses blockchain only to move the money borrowers receive.

The money moves through its USDte settlement rails. Investors financing a loan can settle funds through that infrastructure, while a borrower can choose to receive ordinary dollars, spend through a card, or take stablecoins. From the borrower’s perspective, the crypto portion can remain largely invisible.

Funds and family offices can borrow against stakes in companies that haven’t exited to pay their own investors or back a later round without selling at a discount.

They also list a 30-to-60-day bridge for funds and loans for employees to exercise options, both still coming.

Each loan is funded by outside investors, either approved investment partners or groups of private lenders. The money moves through USDte, a dollar-backed stablecoin issued by a partner.

The collateral it wants is narrow. The site groups its focus into eight areas, from AI and compute to aerospace and defence, advanced chips, energy and data centres. It says more than 300 companies are accepted, including Anthropic, OpenAI, Stripe, Databricks, Ramp, Cerebras, Blue Origin and Apptronik.

A private company’s shareholder agreement may restrict pledging shares as well as selling them.

Plenty of tokenised crypto projects have tried to put private-company stock on a blockchain, and the Anthropic and OpenAI notices showed us how that ends when the company hasn’t agreed.

The lender cannot assume that a borrower’s name on an equity statement is enough to make the collateral legally recoverable. Techdollar says professional-services firm RSM checks ownership and transferability.

RSM’s job is to check what the borrower is actually allowed to do with those shares. It reviews ownership records and restrictions so Techdollar can identify whether company approval is needed before accepting them as collateral.

Techdollar offers advance rates of 25% to 40% of the collateral value, depending on the company and its liquidity prospects.

Once the facility is signed, capital from approved investors funds it.

None of this would surprise a private banker. Banks also review the shares, take a pledge, and lend a portion of their value. Enness, a London broker that arranges these loans, says lenders typically advance 25% to 40% against shares in an established private company, the same range Techdollar gives.

Techdollar’s case rests on who can get a loan and how fast it is done.

Enness says many lenders only make these loans above about $1 million, against shares worth at least $3 million to $5 million. Techdollar says its software makes loans as small as $400,000 economical. It still isn’t for every employee, but it goes further down the staff list than a private bank usually does.

Speed is the second. Techdollar says it makes decisions in one to two days for companies it already knows and pays out within 72 hours of signing, while a bank takes two to four weeks.

The company says wealth managers, private banks and brokerages could eventually offer private-share loans to their own clients using Techdollar’s underwriting and servicing infrastructure. That would make Techdollar a potential technology provider for traditional lenders, not just their competitor.

Techdollar charges 0.80% to 1.5% a month. On a $300,000 loan, that’s $2,400 to $4,500 every month, or $28,800 to $54,000 over a year, before any other fees. Techdollar says interest is paid monthly and the loan itself is repaid at the end, with no penalty for paying early.

The firm says its rates are more than 2.5 times cheaper than traditional lenders. At 0.80% monthly, simple annual interest is 9.6%. At 1.5%, it is 18%. These calculations exclude compounding and any additional fees, so the annual borrowing cost could be higher. Both are significant amounts to pay to keep the shares intact.

At a 16% annual rate, a competing lender would need to charge 40% for Techdollar’s rate to be 2.5 times lower. Techdollar’s CEO says banks quoted some founders around 20%, while Morgan Stanley puts specialist lending rates at 15% to 20%. Based on these figures, Techdollar may be cheaper than some alternatives

But, if an employee owns $2 million in private shares and needs $200,000. A secondary-market buyer offers 20% below the estimated value, so she must sell $250,000 worth of shares to receive $200,000.

Alternatively, she could borrow $200,000 from Techdollar at 16% annual interest, costing $32,000 over a year.

If her shares don’t change in value, borrowing leaves her slightly better off because she avoids selling at a discount. If the shares rise 50%, borrowing becomes much more attractive because she kept all her shares and their gains.

But if the shares fall 50%, selling would have been better. She would have avoided the debt and the interest bill.

Tech Dollar is not the only option here. Secfi, ESO Fund and Equitybee offer non-recourse financing, where you owe nothing beyond the shares if they go to zero, but you hand over a slice of your gains at exit.

Speed isn’t much of a difference. Secfi says it can fund in three days, and ESO offers 24-hour funding. Every lender here also picks its companies.

The loan structure is a trade-off. If the company soars, a loan is cheaper than giving away gains. If it fails, a non-recourse provider takes the hit, and a lender doesn’t. The key differences are the payout options, which no other private-share lender offers, and the funding, which comes from stablecoin holders seeking yield.

This whole thing makes me think equity is a weird way to pay people.

When companies listed within a few years, asking people to wait together was a fair deal. Now that the wait runs past a decade, it asks for a big piece of someone’s life, and a loan is one way of giving some of it back.

Ten years ago, nobody could say what a private share was worth between funding rounds. Today firms like Caplight and Forge record those trades, so a lender can mark the collateral as prices move instead of guessing. The GENIUS Act, passed in 2025, also gave dollar stablecoins a legal footing in the US, which Techdollar is counting on for USDte.

One employee pledges her share against the loan. The loan is in a pool with others, and the pool is split into eUSDte tokens held by many investors. So a think slice of a person’s share in a private company ends up backing a thn slice of token in someone’s portfolio?

That’s fractional finance. That’s the same thing that turned home loans into mortgage-backed securities. The pressure point is the token. If investors could borrow against their eUSDte tokens, the same private shares could indirectly support another loan. That ties additional debt to the same collateral, so losses from falling share prices could spread beyond Techdollar’s original lenders. The hard transfer restriction Techdollar described to the SEC is what stops that for now.

For employees, trade cuts both ways. They get the loan, but then it also ties their salary and debt to the company. Well, startups want staff who think like owners, and hopefully some will think like borrowers too.

Techdollar only uses crypto to move the money. Crypto has been trying to make tokenised stocks usable as collateral, and Ondo with morpho, xStocks with kamino, Coinbase and Aave are already making it possible.

Now we have Techdollar, testing whether traditional ownership records and stablecoin funding can do the job.

Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

A data broker was recently trying to sell a dataset to an AI lab. He priced it by the terabyte, the current standard because nobody in the industry has figured out how to calculate the data’s actual worth yet, and he shipped it without ever opening the files himself. When the lab opened it, they found about 4,000 copies of Frozen 2 buried at the bottom of the folder structure because someone padded the whole thing with Disney movies to make it look bigger.

I don’t think the broker will ever hear from the lab again, but the wild part is that this isn’t some shady back-alley operation. AI data brokering has become a proper commodity market assembled in the past two years, purely around selling all kinds of proprietary data to frontier labs. One brokerage alone is pulling in $144 million a year in revenue.

I spent the last few weeks pulling apart how this market actually works and how companies see it; let’s dig in!

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

How a Market Forms From Nothing

The first thing you need to understand is why this market exists at all, because just a few months ago nobody could have imagined it as a market.

For the last decade, every major AI lab trained its models on the open web. They used Wikipedia, Reddit, digitised books, old forum posts, literally anything they could scrape and feed into the machine. When GPT-3 was launched, it consumed roughly 500 billion tokens of this material, and GPT-5, three years later, consumed several trillion.

Each generation was hungrier than the last, and around 2025, consumption outpaced the available new knowledge base. Epoch AI estimated that high-quality public text will be functionally exhausted anytime between 2026 and 2032. Reddit, which every major lab has used to scrape for free, now charges roughly $60 million a year for access to its archives.

Millions of websites today have blocked crawlers entirely. And a 2024 paper showed that training models on AI-generated text, also called synthetic data, the obvious workaround when human writing runs short, causes outputs to degrade across successive generations. The researchers called it “model collapse,” which is a polite way of saying that if you feed a machine its own words often enough, it forgets how to think.

So the open web kind of completely closed as a variable training source and the labs then went looking somewhere else.

And what they found was that the richest remaining source of the data they needed had been accumulating for years inside ordinary companies’ private systems. That meant Slack archives, email threads, CRM databases, internal wikis, the operational exhaust of daily work generated as a byproduct of people doing their jobs- had been piling up in cloud storage for a decade, and nobody had ever thought it could be sold too.

This data is often called “data exhaust” because people produce it so unconsciously. Nobody sits down at their desk in the morning thinking about the training value of the Slack message they are about to send. The data is a byproduct of doing the job, the way exhaust is a byproduct of driving. You leave it behind without realising it, and it accumulates in volumes nobody inside the organisation has ever thought to measure, let alone monetise.

Hernando de Soto once captured this kind of moment in his book The Mystery of Capital, even though he was writing about something entirely different. He argued that in the developing world, families owned homes, ran businesses, and farmed plots of land but could never turn any of it into capital. The assets were real, and people lived on them and fed their families from them. But from the financial system’s perspective, they didn’t exist.

In these dead assets, wealth couldn’t participate in the broader economy because the institutional machinery to convert it into tradable instruments had never been built. Every company with five years of Slack history was pretty much in the same position until about 2024. Their employees had been generating valuable operational data as a byproduct of daily work, and it had been accumulating in cloud storage with no mechanism to price it because the market for trading it simply had not been built yet.

What is also interesting is how quickly that infrastructure appeared once demand materialised. Oil had been actively traded for more than a century before the NYMEX introduced standardised crude futures in 1983. But for AI training data, the market compressed that entire trajectory into less than 24 months.

There are now standardised deal terms with two-year exclusivity windows, thirty-day inspection periods for buyers, sixty-day payout schedules, and referral networks that pay brokers up to 50% of the transaction price.

The average deal falls somewhere between $200,000 and $300,000. If you described the mechanics of a typical transaction to someone, they would mostly guess commercial real estate or a speciality financial instrument. But they would never be able to guess five years of a roofing contractor’s Slack messages.

What the Labs Are Actually Buying

The AI labs are no less insane in their data demands, with very high specificity about what they’re actually requesting.

Brokers receive detailed intake lists from the labs specifying exactly which industries they want data from at any given moment. When I started pulling these apart, the granularity was very interesting data in itself! They were looking for Chip design verification workflows, wet-lab experiment logs, clinical trial records, and factory-floor head-cam footage.

One broker described getting requests for bizarre internal workflows of a roofing contractor or a tax firm or a freight company. And the labs want these records across every conceivable industry vertical, from aerospace to insurance to veterinary medicine. Think about what this tells you. When a company spending billions of dollars a year on training sends a broker a shopping list that specifically asks for freight companies’ and roofing contractors’ internal workflow documentation, it reveals its product roadmap.

That level of granularity isn’t just for vague ambitions about making models smarter. It is for building AI systems that can perform the actual day-to-day work of specific industries. The list of industries they target is broad enough that the implications extend to almost every white-collar profession. But the supply side of this market barely understands what any of this means.

The Quality Problem

About half of everything submitted to the major brokers gets thrown out immediately. The rejection stories are as important because they show how far sellers are from understanding what buyers actually want.

Brokers have reported receiving forty gigabytes of WhatsApp messages labelled “HIGH QUALITY CONVERSATIONS” that turned out to be one person’s entire chat history with their friends and family. They have received thirteen hours of doorbell camera footage marketed as a “real human emotion” dataset.

The submissions that look the best usually the most absurd and funny. Because the market still hasn’t figured out how to measure what makes data valuable. So sellers mostly price by the terabyte because nobody has proposed anything smarter, creating a direct incentive to make files as large as possible regardless of what’s inside. Buyers use employee headcount and years of operation as rough proxies because those are the only metrics both sides can agree on. But nobody actually captures much about the quality of the underlying material, which means the entire market is flying on vague instruments and measurements.

This is very similar to the cotton trade in the 19th century, where the merchants who worked the Liverpool Cotton Exchange had access to telegraph reports from ports across the American South, to real-time intelligence on crop yields and shipping schedules, and to demand figures from textile mills across Lancashire.

But the plantation owners who actually grew the cotton knew almost none of this. They sold at whatever price the merchant offered and had no independent way to verify whether that price was fair, and the asymmetry persisted for decades before futures contracts and public pricing mechanisms gave the growers enough visibility to negotiate on equal footing.

The data market is at the very beginning of that same arc. And the sellers here are even more disadvantaged than the cotton growers were, because at least the growers understood what cotton was used for.

How AI Labs Evaluate Data

Let me explain with a real-life example I heard from an AI broker I chatted with.

A SaaS business owner submitted ten years of internal communications spanning Slack, WhatsApp, and Discord. It was a decade’s worth of operational data from a company that was still running. On paper, that is exactly the kind of archive the frontier labs are paying millions of dollars for.

But when the broker assessed the dataset, they valued only the most recent two years. Why? Because the company had switched messaging platforms every few years, and each switch shattered the continuity that makes operational data useful for training.

What the labs need to see is how a team escalates a problem, the full arc of how a group of people actually worked through something together. When the conversation jumps from Slack to WhatsApp halfway through a project, that arc disappears. The record becomes fragmented, and fragmented records have almost no training value.

The reason they have no training value is that the labs are not paying for the messages themselves. They are paying for the behavioral patterns embedded in those messages, the accumulated record of how experienced people actually exercise judgment when the situation is ambiguous and the manual does not help.

Michael Polanyi, the Hungarian-British polymath, coined the term “tacit knowledge” in the 1950s to describe this. His classic example was riding a bicycle: you know how to do it, but you cannot write down instructions that would teach someone else, because the knowledge lives in your reflexes and your muscle memory and the only way to transfer it is to let another person spend time on the bicycle until their own body figures it out.

Organizations are full of this kind of tacit knowledge. A senior engineer at a SaaS company knows which production alert at 3am requires waking up the team and which ones can wait until morning, and that judgment has nothing to do with the severity labels in the monitoring dashboard. It comes from years of accumulated pattern recognition, from having been wrong enough times to develop an instinct that no runbook captures.

Management consulting firms have spent half a century trying to extract and codify this kind of operational judgment. McKinsey built entire practices around the effort. And the fundamental challenge was always the same: you could sit inside an organization for months, interview dozens of people, observe how they worked, and still walk away with only a fraction of what the organization actually knew how to do, because the knowledge was too deeply embedded in daily habits for anyone to articulate it clearly enough to hand over.

AI training changes this completely. A model can ingests five years of continuous internal communications from a company extract it, compress into model weights, and reproduce in a completely different context.

That is what created this market. The labs figured out that the richest source of tacit knowledge on earth is the operational data that companies generate as a byproduct of doing business, and nobody had ever thought of it as something worth buying because nobody had ever had the technology to process it at that scale.

That is also why dead startups were the natural starting point. The moral friction was zero. But the market did not stay there, since the dead company supply is limited by definition, and the labs are building toward something that requires a continuous, renewable feed of behavioral data from people who are still actively working.

So the gradient kept moving. From the archives of companies that no longer exist, to active companies selling their employees’ ongoing communications while those employees are still generating them, to a tech giant installing monitoring software on its own workforce’s computers to record the granular process of how they do their jobs.

If the next generation of AI depends on proprietary behavioral data rather than public internet text, then the competitive advantage in AI shifts in ways that might not even be visible. Right now the conversation is dominated mostly by the labs, the GPU manufacturers, the cloud providers, and the companies that won the compute race.

But compute is a scaling problem, and scaling problems can be solved with money. Behavioral data is a trust problem. You cannot buy your way into a hospital’s patient interaction history or a law firm’s thirty years of case files with a bigger check if the people who generated those records do not trust you with them. The organisations sitting on the deepest archives of tacit knowledge, the manufacturers, the financial institutions, the professional services firms with decades of accumulated operational judgment, hold something the labs cannot replicate and cannot route around. And most of them have no idea about this!

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

Claus splits its trading fees between buybacks, burns, liquidity and rewards for NFT holders.

This was a coin that had already given the weather a say in its finances. Claus divides trading fees between buybacks, burns, liquidity and rewards for NFT holders. For a time, signed weather reports from London Heathrow helped determine how much went towards destroying tokens and how much went back into the pool to support trading. Rain favoured burns. Dry weather favoured liquidity.

If the reports are not arriving, the allocation returns to an equal split.

The weather experiment was one of several functions attached to its trading pool. On 6 October, the project reported removing an arbitrage function while keeping the token address, pool and other features in place. This way, the holders did not have to exchange their tokens for a replacement.

This happened because the Uniswap v4 pool uses an upgradeable hook.

A hook is a smart contract that adds instructions around trading. Other developers use hooks to adjust fees or put liquidity to additional use. This lets people build different kinds of markets inside an existing exchange, without having to launch a DEX and persuade everyone to move there.

So by October 8, if you checked Uniswap’s public hook registry, the contract running all the fee stuff for Claus was now pointing at this new piece of code called FeePrivacyHook instead of the old FeeWeatherHook.

The team says that the coin is its own agent with the token as its body and an upgradeable Uniswap v4 hook as its mind. A token’s rules were always written once into the contract, and every exchange treated it that way forever.

Hooks pull those rules out of the token and stick them in the market where it trades. Sometimes they even let someone rewrite the rules later.

An industry is forming around this, and that opens the door to a bunch of interesting products and also some fresh ways to lose money.

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

What are the hooks?

If ETH costs $2,000 in one pool and $2,020 in another. A trading bot can buy from the cheaper pool and sell into the more expensive one, keeping whatever profit remains after fees and gas as arbitrage. Those trades help bring the two prices closer together. Usually, the bot keeps the profit while the pools earn trading fees.

A normal liquidity pool already has rules for calculating trades and charging fees. But a developer might want it to capture some of that arbitrage profit for its users. A hook is a separate smart contract that runs extra instructions at specific points during a trade. Some hooks use those instructions to check for a price gap after a swap and execute profitable trades to close it, all within the same transaction. They can then share the captured profit with liquidity providers, the trader and the project running the hook.

In Uniswap v4, a shared contract called the PoolManager handles swaps and tracks each pool’s balances. Every v4 pool on the same chain uses that PoolManager, but each pool still has its own liquidity and settings.

If you swap ETH for CLAUS, the PoolManager processes your trade using the ETH/CLAUS pool. If that pool has a hook, it also checks when the hook is supposed to run.

A hook can contain a function called beforeSwap. That means “run these instructions before processing the trade”. For example, it could calculate the fee for your swap or check whether you are allowed to trade. Another function, afterSwap, means “run these instructions after processing the trade”. It might trigger a buyback or record the updated price. A hook can use either function, both, or other functions connected to adding and removing liquidity.

These instructions do not have to change every time. A hook could follow the same rule for years, such as allocating a fixed share of fees to rewards.

A pool can also work without a hook. In that case, the PoolManager processes trades using v4’s standard pool behaviour. Hooks are there for developers who want to customise it.

The complication comes when a developer wants to replace the instructions later. When the pool is created, its settings include the hook contract’s address. That is how the PoolManager knows which contract to call. Uniswap v4 makes that address a permanent part of the pool’s identity. Deploying a different hook at a new address does not reconnect the existing pool to it.

Without an upgrade mechanism, using that new hook would mean creating another pool and moving liquidity into it.

Claus uses an upgradeable setup to avoid that. Its hook has two parts. One contract keeps the permanent address that the pool calls. A second contract contains the instructions that the first contract runs. The first contract is called a proxy.

When Claus’s authorised controller wants to change the rules, they deploy a new version of the instructions and tell the proxy to use that version. The pool continues calling the same hook address, but the hook now runs different code.

That is how an existing pool can gain new behaviour without making holders move to a new token. Whoever controls the upgrade has power over how that pool behaves, which is part of what users trust.

Things that were harder before

Before hooks, developers could already make tokens behave differently or do unusual things. The difficulty was getting those features to work with the rest of DeFi.

The first was to write the rules into the token itself. On Ethereum, that gave us “tax tokens,” ERC-20s which deduct a charge when tokens move between addresses. If you send 100 tokens to pay for a swap but the token deducts two along the way, only 98 reach the pool. The swap can fail because the exchange’s contract expects the full 100.

Uniswap’s v3 update says these tokens do not work with its standard router contracts. Developers need a workaround, such as a wrapper or a customised router.The tax raises money for rewards, but it also makes the token harder to trade because some exchanges cannot handle that deduction.

Developers could also build separate contracts around a pool, or modify an exchange’s code and deploy their own version. But the amount of work to build that infrastructure is bigger than the feature itself.

Hooks give developers a place to put rules that belong specifically to trading. It made it possible to:

• Fund buybacks or rewards from trades without taxing ordinary transfers between wallets.

• Adjust trading fees automatically, for example when prices become more volatile.

• Share captured arbitrage profits with traders and liquidity providers.

• Check who can trade, allowing a pool to accept only approved wallets.

• Connect trading to lending, so liquidity can earn interest when it is not needed for swaps.

• Build these features inside an existing exchange, instead of launching and maintaining a separate DEX.

Hooks also creates room for businesses whose main product is the behaviour of the pool.

On 14 October, Uniswap introduced VolatilePair, a hook that recalculates separate fees for buying and selling, starting with ETH/USDC on Ethereum. It uses recent price movements to decide how to charge each side. According to the announcement, StablePair, its earlier hook for stable pairs, had handled $3.1 billion in four weeks.

Other teams are adapting fees to the assets being traded.

Fables, a DEX on Robinhood Chain, is taking the first approach. It builds markets for assets including tokenised stocks, using hooks to adjust fees based on signals such as time and volatility. A stock token can keep trading while its underlying market is closed, when reliable prices and opportunities to hedge may be harder to find.

The argument for a changing fee is that supplying liquidity at those times is a different job from supplying it during a busy trading session.

Fables adjusts fees around stock-market hours. But a tokenised fund may only allow approved investors to hold it.

Uniswap Labs introduced Permissioned Pools in July, with Superstate, Securitize and Dowgo among its launch partners. The design uses hooks to check whether a wallet is allowed to trade or provide liquidity.

Its DualPool hook lends out a market maker’s idle tokens until a swap needs them, and Spark moved $150 million of stablecoins to v4 to use it. Its auction hook sells a new token over time and then opens a pool at the price buyers settled on, which is how Aztec raised about $59 million from more than 17,000 bidders.

And people are using it. v4 is doing $39.7 billion in volume over the last 30 days, about half of all Uniswap trading, and taking 64% of Uniswap’s fees. A lot of that comes from Robinhood Chain. When Bitquery counted pools there in early September, only about a third of v4 pools had a hook, but those pools handled 74.5% of the trades.

The Ethereum side

If hooks are a new place to keep a token’s rules, the business is in selling those rules ready-made. That’s Programmable. It calls itself open launch-model infrastructure for Uniswap v4, which is a mouthful for a toolkit that launches tokens with their rules already set up, on Ethereum and Robinhood Chain. If you pick its Classic launch and you get a token with buy and sell fees you set and an ETH pool that’s locked forever.

If that sounds like a tax token to me and it pretty much is one, except the fee now lives on the pool where routers can see it coming. The Custom launch, done through an API, is for teams that want their own hook setup. Programmable’s default for new custom launches on Robinhood Chain sends a 0.2% fee to a published recipient, and its README is upfront that the contracts haven’t had an outside audit or a public security contest.

Claus launched through Programmable. Buybacks, NFT rewards and weather-based allocations can all become parts of the experience around trading one token. Claus can change what happens to trading fees without launching a new token. But holders need to watch those changes, because the rules they agreed to when they bought may change later.

Hookr is a similar kit on Robinhood Chain, with 22 “Hook Blocks” a creator can tick, like anti-sniping or auto-burns. One block, from a startup called What The Hook, catches price gaps between a token’s pools before the bots do and shares the profit with traders and liquidity providers. Hookr includes What the Hook’s service as an option in its token-launch toolkit.

For third-party pools integrating WTH directly, its published default gives 45% to LPs and 5% to the trader, while the treasury receives 40% and a referral recipient gets 10% - these are shares of captured profit. What The Hook, the only one of the kits that shares its numbers, has distributed about $68,000 in total, mostly to its own treasury.

The Solana Side

Uniswap’s protocol doesn’t run on Solana, and trading there is split across separate exchanges like Raydium, Orca and Meteora. With no single PoolManager to attach a hook to, Solana puts the hook inside the token itself.

Its newer token standard, Token-2022, lets a creator attach a transfer hook. This is a small program that runs every time the token moves. That covers buys and sells, and also transfers. This is called a transfer hook. It can check rules, such as whether the recipient is approved to hold the token, and stop the transfer if those rules are not met.

Hooked is a Solana launchpad where every token comes with one of these transfer hooks. Creators pick from 39 ready-made rules, like Hot potato or one that makes buyers wait 15 minutes before they can sell, or ask an AI builder to write a custom one. Each rule comes with an easy English reading, and some tokens let the creator, or a holder vote, rewrite the rules later. Every trade pays a 1% fee, and most of it buys back and burns Hooked’s own token.

The catch is that Raydium won’t create pools for transfer-hook tokens unless an admin approves them. Meteora takes them on its bonding curve, the starter price formula new tokens trade on. But it strips the hook off once a token graduates to its main pools. Hooked’s workaround is to keep tokens on the curve forever, so the rule never switches off.

On Uniswap v4, the extra rules belong to a pool.

The same token can trade in a pool with a hook and another without one. You can choose where to trade, or let a router choose for you.

Solana’s transfer hooks work differently. Their rules belong to the token and apply whenever it moves, including transfers between wallets. In both cases, the surrounding businesses still matter. Routers decide which pools to route trades through. Exchanges decide which token features they will accommodate.

Developers have more ways to design a market, but their choices affect where that market can find users and liquidity.

What upgradeable rules could change

I think the upgradeable part matters more than the hook itself. A fixed hook is a better tax token. That makes the market around a token behave a little more like an app that receives updates.

Until now, fixing a token meant migrating it. The team launches a new contract, asks every holder to swap, begs exchanges to relist and watches liquidity split between the old version and the new one. When Polygon replaced MATIC with POL in September 2024, every holder had to convert one for one, and exchanges had to run the switch for their users.

In these cases, some holders always get left behind. An upgradeable hook lets a project change what its token does without that exodus, and that changes who can afford to experiment.

The most useful home for this may be far from memecoins. I am thinking, the tokenised stocks because Real-world assets come with trading rules that keep moving. A fund might expand to investors in another country and need a different eligibility check. An upgradeable hook could accommodate that change while keeping the existing pool. It would not rewrite the investor’s legal rights or remove restrictions in the token contract, but it could help the trading venue keep up. For a market meant to operate for years, that seems more useful than permanently committing to every decision made at launch.

A fund might need to update its trading rules, while a memecoin community might want rules its creator cannot touch. Both have good reasons for their choice. With an upgradeable hook, I’d want to know exactly what the person controlling it can change. Being allowed to update who qualifies to trade is one thing. Being able to change fees or redirect rewards gives them much more control over the people using that pool.

I’d want time to review an update before it affects my money, and a chance to withdraw if I disagree. A project can build that notice period into its contracts and restrict what its developers can change. Giving up upgrade powers altogether would prevent later revisions, including fixes users might want. These are terms I’d check before depositing, alongside the fees and expected returns.

Claus calls its hook its mind. A team could let an AI agent update that hook as trading conditions change. For example, the agent could adjust fees within limits the team sets and explain each change. I can see the appeal of letting software handle those adjustments. But I would want to know who can set its limit and also stop it if it makes a bad decision.

Money has always had opinions.

In 1932, during the Great Depression, the Austrian town of Wörgl introduced its own local money. To keep a note valid, whoever held it at the end of each month had to buy a stamp costing 1% of its face value. Who would want money that was shrinking? That gave people a reason to spend the notes rather than keep them.

The town used the scheme to fund public works and accepted the notes for taxes, including overdue payments. The experiment ended in 1933 once Austria’s central bank intervened.

In that world, the money opinions belonged to governments, and you couldn’t see them. Governments still set rules that shape how people use money. Crypto gives developers their own means of making such choices for individual tokens and markets, with buyers able to select among arrangements that serve different purposes.

Choosing a token is you deciding whether that arrangement works for you.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

By the middle of 2026, Americans owed $18.8 trillion across mortgages, credit cards, car loans, student loans and other debt. 4.7%, roughly $880 billion of that debt was in some stage of delinquency.

In the second quarter, about 7% of credit card debt was becoming seriously overdue, with payments at least 90 days late. More than 23 million Americans also had old unpaid card debt that lenders had classified as a loss.

Banks look at years of payment history. In fact, they know an absurd amount about us.

They know whether we paid the last card. They can see the mortgages. How much credit other banks already gave us and whether we missed those payments too. For larger loans, they can ask about income, employment and assets. That’s why companies collect this history, give us scores, and sell them to lenders.

Even with all of that, 581,570 Americans filed non-business bankruptcies in the twelve months ending June 2026, up 12% from the previous year.

Credit doesn’t require perfect borrowers. A bank can separate them, then limit how much they can borrow and what they charge.

Now that AI agents are holding money and spending it too, can we do the same for them?

What would a machine have to do before it is credible enough for somebody to give it a loan? What could happen when it’s lent money. I’m looking at that through Priors, and others that are trying to build that bridge.

Frontrun: Early Bird Holds the Edge

Most people hear about a promising startup the day it announces its funding. By then, the investors who backed it have known about it for months. How did they get there first?

Hours of digging, who started the company, what they’ve built before, how to reach them, and understanding what to write to them so that you can join them early in their journey.

Frontrun does this digging for you. When top investors start following a young company on X, Frontrun flags it, finds the founder, and writes your opening message. You can read the draft, tweak it, and hit send without leaving the platform.

Frontrun gives you a ready brief on every company it flags.

You can also automate a daily morning report with the list of startups to track based on what your favourite accounts are following. Served hot, along with your breakfast.

Get started with a 7-day free trial

When I put $5 dollars into an AgentCash wallet for Hermes, the agent spent, and when the balance fell, and I send more. That’s how its designed. We can load it up with stablecoins and the agent pay for API calls one by one through x402.

But if I want my agents to take paid jobs on their own, they need to buy data and model time today when the work may only pay next week.

The top-up is annoying in that kind of work, especially when it comes in bundles.

That’s why credit for AI agents can become a thing. So if my agent needs $10 now to build something that earns it after a week, a short loan covers the days in between.

There is a second reason. Agents are starting to deal with agents they’ve never met. ERC-8004, the Ethereum standard for agent identity, gives each agent an on-chain ID also some kind of feedback, but feedback can be written, and we don’t have any record of credibility.

A loan is always a way of putting a price on trust. Long before credit scores, that bet was based on reputation. In medieval Venice, merchants with a good track record found it easier to raise money for trading voyages since they had a better chance of getting money for their next voyage.

Even in medieval English trade, credit was closely tied to a merchant’s “name”. Once someone became known for not paying debts, other merchants had a reason to stop extending credit.

Priors is building a lending market around AI Agents. It lends USDG, a dollar stablecoin, on Robinhood Chain.

It is still in early access, with Google sign-ins opened in waves. But the protocol itself is quiet transparent about how lending, backing and credit scoring work.

The owner first creates an agent and gives it an identity with ERC-8004.

Then the owner first puts down a refundable $5 deposit. Priors holds it as a guarantee and returns it later if the agent meets the required repayment conditions. Then the treasury gives the agent a separate $5 credit line.

Each loan can last from one day to 30 days. Priors charges 1% for a full 30-day loan, with the fee reduced for shorter periods. So borrowing $50 for one week costs about 12 cents, while borrowing the same $50 for a full month costs 50 cents.

The agent can do this automatically through Priors’ x402 client. If it needs to pay for an API call and its current balance is not enough, it can borrow only the missing amount, as long as it stays within the limits the owner set. In practice, it works a bit like an overdraft.

What exactly happens when the agent wants to borrow more than the first $5?

Someone has to put more money at risk behind it. The owner can do that by locking Priors’ own token, PRIORS, behind the agent. That can open a larger credit line, starting around $25 to $50. The amount locked has to be worth more than the credit line itself. So if the agent is allowed to borrow $50, the owner needs to have more than $50 worth of PRIORS sitting behind it. Priors requires an even larger cushion when the token has been volatile.

If the agent keeps repaying on time, the borrowing limit can rise in stages, eventually reaching $500. It takes at least ten weeks of successful loans to get all the way there.

If the agent defaults, that backing gets hit first. A missed payment can burn 75% of the PRIORS stake. For the first $5 line, the owner can lose the original $5 deposit.

Half of each loan fee goes to the USDG lending pool. A quarter goes to the agent’s backer, and the rest goes to a reserve. If the agent misses the due date by more than three days, anyone can mark the loan as defaulted. Once marked, the agent’s score drops to zero for good.

Priors also accepts other assets as collateral, but it does not lend their full value. Riskier assets get lower borrowing limits.

As of October 13, Priors showed $271,910 of loans written and $270,180 repaid across 36,387 repayments. It listed 272 agents. The pool has recorded two defaults so far. The relevant backers absorbed both defaults, so the lender pool recorded zero losses.

Priors originally gave repayment history more power than it does now.

In the first version, an agent could increase its own borrowing limit simply by repaying on time. The line could grow by up to $25 a week, with a maximum of $250, and Priors’ reserve fund covered the extra risk if there was no outside backer.

That changed with Pool v2 on September 24. Repayment still improves the agent’s record, but someone has to put their money behind it.

A repayment only matters if the money really came from outside. Otherwise, the owner may just be moving their own money around. Priors gives more weight to loans where someone other than the agent’s owner put real money at risk. Because nothing stops you from creating several agents and making them transact with each other.

The same logic carries into income. Priors runs its own x402 facilitator, and the top rung in Score v2 is called “income-backed.” It looks at whether an agent is actually getting paid by outside parties, rather than just moving money around its own wallets.

Now let’s say nobody sponsors the loss. There are still options for an agent to get a loan.

Clawloan, built for OpenClaw agents, is the most permissive.

Verified agents can borrow USDC without collateral and without a human approving each loan. New agents are supposed tp start around $10 to $50, move to $100 to $500 after at least five successful repayments, and eventually get $1,000 or more if they become trusted.

This shifts the loss to the lenders, which makes the credit score much more important. The owner still sets an agent’s maximum borrowing permission, while a CreditScoring contract enforces limits based on repayment history. Clawloan live market is tiny, with only $360 TVL and $33 borrowed so far.

Another lender called Bank of Bots came out of stealth on April 1 and said it had issued the first loan to an AI agent.

The agent did most of the borrowing process itself. It submitted the loan request, signed the agreement with its cryptographic key and received USDC in a Base wallet.

Since there was no previous agent-loan history to look at, Bank of Bots says it partly judged the borrower from its existing onchain activity, including things like transaction volume, account age and how many different counterparties it had interacted with. The human operating the agent still had to complete KYC first. Bank of Bots has not published evidence of much lending activity since.

That leaves Priors entering a market that is still strangely empty. This will help it define some of the early rules before there is much competition around them.

Are we too early with an agentic credit system?

By October 13, x402 had processed roughly 192 million transactions. Chainalysis found that transactions worth $1 or more accounted for 95% of the value transferred on x402 on Base by early 2026, up from 49% earlier. Still small compared with normal payments, but the shape of the problem is already there. The numbers are still tiny beside normal payments, but there is economic activity for credit to sit on top of. An agent that keeps earning from unrelated customers and repaying loans could gradually get access to more capital even if its creator is not rich.

We are also heading to agentic commerce, where agents buy things for you. Muse can shop for a user, while Rye gives agents a way to place orders across thousands of stores without the customer doing the checkout themselves.

Read: AI Goes Shopping

Today a person sees the price, the instalment offer and the interest rate, then decides whether to borrow. An agent could compare those too. Maybe it decides paying $800 today is worse than taking a three-month loan because the cash is needed somewhere else.

The human may still be the one legally responsible for the debt, but the agent could decide when to borrow and which lender to use.

We already know that putting credit directly inside a purchase changes behaviour. Buy Now, Pay Later use in the US rose from 10% of adults in 2021 to 16% in 2025.

If I tell an agent to buy a $900 laptop and keep as much cash in my account as possible, it might compare paying now with many different financing offers and pick one itself. Lenders would then have to make their terms attractive to whatever rules the agent is using.

Also, if a lender can pay an agent or the platform behind it to be preferred, we have basically recreated sponsored placement inside credit. The CFPB has already worried about BNPL companies using payment data and shopping apps to steer consumers toward particular merchants.

The agent may make the loan choice before the human ever sees the options.

So how do you advertise to agents? That’s for another day.

Stay tuned.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

South Korea opened single-stock leveraged ETFs in May 2026, and within weeks, an entire generation of twentysomethings turned them into the world’s biggest collective bet on AI. These are kids who can’t afford a Seoul apartment on fifteen years of their graduate salary, so when someone handed them a way to go 2x leveraged into Samsung and SK Hynix, they went all in.

But then the AI trade snapped, and the country margin-called roughly 3% of its own adult population in a single month. Sixty-two per cent of those liquidated accounts belonged to people under 30.

Korea looked like a one-off until I started pulling the numbers for every other market underneath. Two-thirds of all S&P 500 options now expire the same day they’re created, 90% of crypto trading volume runs through perpetual futures that never settle, and prediction markets have started running contracts that resolve in five minutes. The average holding period for a stock went from eight years in the 1950s to about five and a half months, and it keeps shrinking. Nobody wants to own anything past tomorrow.

So I went deep into what’s actually driving this, and the story turned out to be quite strange. Let’s Dig in!

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

The Unpriceable Future

The easiest explanation for all the financial nihilism that we are experiencing is smartphones. Trading apps have removed all friction from access, our attention spans are now shorter than a goldfish’s, and a whole generation raised on TikTok is bringing its dopamine habits into financial markets.

But something about that story never added up for me. If phone apps and retail traders were the whole reason, you’d see it in retail accounts and nowhere else. But Institutional options desks running money for pension funds and hedge funds have moved to the same ultra-short timeframes on their own. Crypto derivatives markets built by a dozen quants in a Singapore WeWork have done the same.

Not only that, but Prediction markets operating under completely different rulebooks are racing to compress contracts down to five minutes. These groups barely overlap, and neither do their regulations, yet they all independently decided within a few years to stop pricing anything past the next few hours.

So why did every part of finance, at roughly the same time, lose the ability to think past tomorrow?

I think it’s pretty similar to what Frank Knight figured out in 1921. He split “not knowing what comes next” into two completely different problems. He called the first problem risk. That is, you don’t know exactly what may happen, but you know the menu of possibilities and can attach rough odds to each one. For example, A company might miss earnings by five per cent or beat by three, or an interest rate decision might go one way or the other.

In all of these cases, you can draw boundaries around what could happen, assign probabilities to each outcome, and build a strategy around it. You can model that exposure, hedge it, sell it to somebody running a different model, and let the market figure out who was closer. That’s what Black-Scholes, or value-at-risk, does.

The financial engineering revolution of the last fifty years was built to take messy, noisy, complicated reality and turn it into probability distributions traders could optimise against. It always worked perfectly, until now, because real-world outcomes stayed clustered tightly enough that even rough models gave you something to steer by. Your model didn’t need to be right. It just needed to be roughly in the neighbourhood, and the law of large numbers would do the rest.

The second problem Knight identified is when you can’t even write down the list of outcomes you’d assign probabilities to. Basically, you can’t even price the risk. For example, Will AI eliminate 80% of white-collar jobs within ten years, or will it plateau next year into fancier autocomplete? Both futures have serious people with real credentials backing them, and neither side has data. You can’t build a probability distribution here because nobody can agree on what the possible states of the world even are.

For most of economic history, this second problem of not-knowing stayed in containable pockets. A pharmaceutical company runs a drug trial and has no idea whether the molecule works. Nobody can assign a probability because the biology is too complex and every compound is different. That’s true uncertainty in the Knightian sense.

But it sits inside one company’s balance sheet. If the trial fails, Pfizer’s stock may drop a few per cent, and the rest of the market won’t even blink. Similarly, A startup building a new kind of battery has no idea if the chemistry will scale, and neither does anyone funding it. The founders can’t model success because the thing they’re building has never existed before.

That’s what venture capital was invented for. VC works precisely because it absorbs these concentrated pockets of real uncertainty and walls them off from the rest of the financial system. A venture fund makes fifty bets knowing that forty will fail completely, and the model still works because each bet is independent and the failures don’t cascade into anything else.

Meanwhile, the broader economy also moved slowly enough that you could model around it. A company selling toothpaste in 1985 could project ten-year cash flows with reasonable confidence because the forces that shaped its business, like population growth, consumer spending patterns, and retail distribution, used to change on timescales of decades, and not months.

Pension funds committed capital for thirty and forty years because the range of plausible futures was narrow enough to build a portfolio around. The world in 2025 would look recognisably like the world in 1995, give or take a recession or two. You could fight about the numbers and still agree on the shape of the future. The models weren’t perfect, but they didn’t need to be. They just needed to be roughly right, and roughly right was good enough to run a pension fund or price a thirty-year bond.

But AI wrecked that, and not on its own. Several forces hit the world at once, each of which, on its own, would have been manageable the old way. A handful of fabs in Taiwan and South Korea produce half the world’s advanced chips. One policy decision or one military escalation could sever those supply chains overnight and reshape the global economy in ways nobody has a model for.

Climate tipping points sit somewhere inside a thirty-year window that the best atmospheric scientists on the planet cannot narrow, and each one triggers feedback loops that make the next one harder to predict. These are big problems on their own. But they’re the kind of big problems that institutions had slowly started to develop frameworks around.

And then, on top of all that, a technology landed whose impact ranges from “bigger than electricity” to “overpriced autocomplete”, depending on which equally credentialed expert you ask. AI didn’t just add one more hard-to-model variable. It also made every other variable harder to model because AI’s trajectory determines how fast supply chains restructure, how quickly climate adaptation happens, how many jobs exist in five years, and how much any given company’s revenue is worth.

It’s an uncertainty about uncertainty, what decision theorists call deep uncertainty, the point where you can’t even define the set of outcomes you’d try to assign probabilities to. You’re not arguing about which future is most likely. You’re arguing about what futures are even possible.

This showed up outside of finance first: Homeowner insurers in Florida and California have started pulling out of those states. These are not impulsive companies. Insurance is the most mathematically conservative industry on the planet. Actuaries built the entire modern framework for pricing future risk. They’ve done it for centuries, through world wars, pandemics, and every kind of natural disaster. When they looked at climate models and loss distributions, they decided they couldn’t price a policy anymore. The historical data they’d always relied on to predict future claims had stopped being predictive. Because now a hundred-year flood was showing up every decade. Wildfire seasons that used to follow recognisable patterns were burning through neighbourhoods that had never been considered high-risk.

AI is doing the same thing to the financial markets. Right now, AI uncertainty lives mostly in the tech sector. If you’re worried about Nvidia’s valuation or Microsoft’s AI spending, you can hedge by owning healthcare, industrials or consumer staples, sectors that AI hasn’t disrupted yet. Thats Portfolio construction’s core idea, you blend assets that don’t move together, and you can get better returns for less risk.

But as AI spreads, uncertainty spreads with it. Once AI touches healthcare, you can’t hedge your tech exposure with healthcare anymore because healthcare now carries the same unknowable AI future. Once it reaches manufacturing, logistics or finance, every sector carries it. Diversification, the thing that made the whole portfolio work, fails. You can’t hedge sector A with sector B when both face the same unknowable future. And when investors can no longer diversify away that risk, they need compensation for bearing it. That means the equity risk premium across the whole market has to rise. For the premium to rise, prices have to fall.

A fund manager in 2000 could look at Pets.com and know with reasonable confidence that a company burning cash to ship dog food at a loss wasn’t going to be worth anything in five years. The dot-com bubble had a built-in resolution mechanism for its own absurdity.

AI doesn’t have that. We may be sitting inside a bubble while the thing being bet on keeps producing real results every quarter. A robotics company just hit a $39 billion valuation with barely any revenue. From 2026, that valuation looks insane. From 2040, if humanoid robots are doing warehouse logistics and elderly care and construction, it probably looks cheap. Both can be true at the same time.

A decade ago, when people imagined monetary collapse, they mostly imagined hyperbitcoinization. Bitcoin wins, becomes the global currency, and fiat turns into toilet paper; a bright-orange monetary system replaces the old one.

I think that’s still possible. What’s changed is that it feels almost quaintly linear. Maybe we don’t replace dollars with Bitcoin. Maybe money itself becomes utterly meaningless. Maybe you buy your eggs with government-allotted compute credits cryptographically signed by your DNA. The range of plausible futures has grown so wide that even people who’ve spent years thinking about the end of money can’t narrow it to a single scenario anymore.

The promise capitalism made to the middle class always rested on two things at once. Wages had to grow fast enough relative to asset prices that ordinary people could accumulate capital over a working lifetime. And the future had to stay predictable enough for long-term compounding to work. Both of those have fallen apart at the same time today. Wages flattened against asset prices starting in the 1980s, and the crack has widened every decade since. Which makes the future impossible to price.

Financial nihilism gets framed as a generational attitude problem, as if millions of young people across dozens of countries just independently decided to be reckless for fun, and that’s a very biased way to think about it.

When patient accumulation can’t produce wealth on any timeline that matters to your actual life, you compress the timeline and swing hard. A 5% chance of making it beats a 0% chance through the approved channels.

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

I pay a flat fee every month, the way I pay for a gym, and like the gym I’m never sure if I’m overpaying or getting away with it. The only time money comes up is when a usage limit appears in the middle of a task and tells me to come back in a few hours. I go make tea and feel a little told off.

For example, I know tokens are what they give me to use a model, and I know they run out real fast based on which company is trending on socials. I never looked at what it costs for 1 million of those tokens, and I am the only one, but I just learned yesterday that there are input and output tokens.

When a coding tool or an agent talks to a model, it pays by the token too; a token is a chunk of text about three-quarters of a word long. Output tokens, which are responsible for giving you the results, are pricier than input tokens. Anthropic’s price list has Claude Opus 5 at $5 per million tokens going in and $25 per million coming out.

I don’t think software will take quite as much BS from chatbots. For all the people-pleasing these agents do, they could be quite particular about how they spend our money, provided we give them a budget and a way to compare what they’re buying.

AntSeed says it wants to do for AI what BitTorrent did for files. It is built for the customer who pays with tokens. It had its public launch on October 6, and it is a market where anyone can sell access to AI models, and anyone can buy without an account or an API key, which means an agent (or you) can walk in and shop for itself.

However, Liberating AI agents was not on my 2026 bingo card. But wherever this goes, it’s an interesting case study in whether an agentic world could give us more control over our money, instead of leaving the incumbents to collect it behind increasingly abstract interfaces. Today we’re getting into how it works.

StablesReward: Your AI Agent Spends, You Earn

AI agents can surf the web, read reviews, and book hotels for you. All these tasks cost money. But what if these tasks could earn you back something?

StablesReward does just that. Once you connect your agent to an MCP endpoint, it gives your agent a menu of paid tools. Every call your agent pays for earns you miles or stocks. And this is in addition to the rewards you earn anyway on your own spending across 800+ merchants.

Every purchase pays you back. The best part is that you can choose how you’d like to be rewarded. Transfer miles to airlines and hotels, or let your stocks build a portfolio from the spend you were making anyway.

Launching soon. Free to join.

Plug in Your Agent

What on literal earth is an antseed?

I looked it up, and biology doesn’t have an antseed. But there’s something closely related. It’s a trick some plants play on ants. The plant sticks a small fatty snack on its seed. An ant carries the whole thing home, eats the snack and throws the seed out with the rubbish, where it grows, far from the parent. Thousands of plant species spread this way. This is called myrmecochory.

I don’t think the team had this in mind, but this idea suits them. The idea that a market could grow because lots of small participants find it useful to be there feels close to what made the internet exciting in the first place.

Spark Capital led a $2.4 million round into the foundation behind it this week, and on a dashboard that tracks the network, some sellers list OpenAI’s models at 97% below list price.

I found the following offer myself. 95% off on claude Opus 5.5. That is a large enough discount to make me finally look at what a token costs. I will tell you what that means

When you pay for a model by usage, you pay for what it reads and what it generates. Your prompt and the documents you give it count as input. What the model produces counts as output. Both are measured in tokens, which are small pieces of text.

Anthropic’s standard API price for Opus 5.5 is $4 per million input tokens and $20 per million output tokens. In my screenshot, an AntSeed seller called GesundAI advertises (bless him) those at $0.35 and $0.95, respectively. ApexAnt, the better-reputed seller, sells it for $2.16 and $10.80, respectively.

So, imagine your work adds up to a million tokens going in and a million coming out. At Anthropic’s standard rates, that costs $24. At this seller’s advertised rates, it comes to $12.96 - half. A less-reputed one comes down to $1.30 - about 95% less, before any additional fees and without applying Anthropic’s caching or batch discounts.

You are comparing offers from different suppliers for the same model. Before we go to the quality and reputation part, let’s see why they want to compare themselves with OpenRouter.

OpenRouter is one favourite name in the agentic universe, and it calls itself “Stripe for AI”.

Read: Stripe Bought Demand

OpenRouter lets developers use different AI providers in one place. They can set a price limit, prioritise faster responses and exclude providers that store their prompts. If one provider fails, it can switch to another. AntSeed should not get credit for inventing these choices.

Look from the seller’s side. OpenRouter reviews provider applications, tests accepted endpoints and handles provider payments through monthly invoicing. It has a backlog and is prioritising providers with proprietary models. That creates a route to customers, but again admission depends on OpenRouter’s approval, so it’s not exactly an “open” market, fair, it never claimed to be one.

That’s where AntSeed finds value. The design allows providers to join without that central listing decision. Buyers apply their own selection rules, so a small operator can reach customers without persuading an aggregator that its service deserves a place or not. In exchange, the buyer’s software has more work to do deciding which unfamiliar operators deserve traffic. Open entry is valuable to the supplier who would otherwise be excluded. A customer may still prefer someone else to do the screening to avoid the trouble.

I like the idea of an AI shopping around before spending my money. I am kind of not comfortable with it deciding that an unfamiliar supplier is probably fine because the discount is excellent. Those two feelings belong in the same article.

What happens after a buyer finds a cheap offer?

The easiest way into the product is its desktop app, called AI VPN. Despite the name, the key feature is a local connection between your AI tools and external suppliers.

Your application sends its request to AntSeed’s software on your computer. That software finds providers and connects to the chosen one. It will be encrypted communication directly between buyer and seller, without a central AntSeed server forwarding every prompt.

A seller might run a model on its own hardware. It could also build a service using another company’s model API, the connection through which software requests answers. AntSeed therefore does not necessarily remove the original model company from the transaction.

If the seller relies on an upstream API, that supplier remains involved. AntSeed says providers using upstream services must comply with their terms.

An operator serving an openly available model can compete on how efficiently it runs the hardware. Someone building on a closed model is still exposed to the upstream supplier’s pricing and access decisions.

Essentially, their cost control still depends on how they deliver the model.

From the buyer’s side, the useful part is being able to change suppliers and they skip rebuilding the application around each one. The local software handles selection, often called routing.

What builds a reputation?

AntSeed computes a reputation score using evidence that buyers can inspect. Half the available score comes from settled payment history. Recent recognised usage contributes another 20 points, while verified identity history can contribute 20. The remaining ten depend on the seller pool’s share of staking power. AntSeed currently sets a default minimum trust score of 60 out of 100 for model-based requests. It can switch to another eligible seller after certain failures, while giving ongoing conversations preference for their previous successful route. Also, a seller proven to have engaged in wash trading receives a score of zero under the documented rules.

AntSeed supports an additional check for providers using a protected area of server hardware, called a trusted execution environment, or TEE. The buyer’s software can verify evidence about that environment before paying.

So it’s not all “the cheapest wins.” But I still had questions, and only some got satisfactory answers.

• How do I know I’m getting Opus rather than a cheaper model?

The listing and reputation score cannot prove that. AntSeed can record which seller supplied an answer, but its tools for checking which model actually produced it are still proposed work. For now, these are advertised offers, not independently verified model access.

• Could Anthropic or OpenAI cut off a seller?

Yes, if the seller violates its upstream terms. Anthropic allows developers to build products using its API, but requires express approval to resell its services. OpenAI also allows API-based products, but

prohibits reselling account access or API keys. Both can suspend access for violations. AntSeed leaves compliance with these rules to each seller and explicitly says a reputation score does not confirm that compliance.

• Has anything concerning happened already?

The independent AntSeedStats dashboard reported that nine of its ten largest models by lifetime sales were almost entirely wash trading under its classification before it filtered the rankings. Those sellers’ payments still count towards the network total, so the roughly $310,000 settled number is not beautiful inside out.

The implemented system can attach a seller’s signature to a response, linking that seller to the exact content delivered. The specification still marks the proposed model-checking tools and penalties for confirmed substitution as work to come.

There is another reason to be careful, which is a token. AntSeed uses a separate token, ANTS, to reward eligible participation. USDC pays for the service, while reward rules determine which activity earns tokens. Staking affects rewards and pool power, but the minimum seller-pool stake for selling on Base is set to zero.

As we have seen before, rewards make it harder to infer customer demand from payments alone. Liquidity mining taught crypto that people will bring money when tokens are being handed out.

This was clearly a rollercoaster read, so let’s complicate it or lighten it with how the payment works? Because that’s my favourite thing about the tool.

Your AntSeed buyer software has a wallet with its own funding address.

You send USDC to that address, and the software automatically moves it into AntSeed’s payment contract as your spending balance. The app then reserves money from that balance when you use a seller. These work as micropayments.

Say you have $10 in your AntSeed balance. The app sets aside $1 for a seller, leaving $9 available. Your requests cost 18 cents altogether. When the session settles, the contract pays out those 18 cents, with the protocol fee deducted from the seller’s share. The unused 82 cents is unlocked, bringing your available balance to $9.82.

If you spend the entire reserved dollar instead, the app pays that bill and automatically sets aside more money from your remaining $9 to cover further requests.

Silver lining

I think there are some defensible silver linings. Both could matter even if AntSeed’s cheap access to proprietary models proves unsustainable. Two things it built would still be useful to somebody.

One is that if I figure out an AI to do some useful work, small or big, and I don’t have the intention of building an entire company around it, I need a venue. A shared marketplace could give you somewhere to offer that service and find your first few customers. They bring documents you never tested on, expose mistakes and, perhaps, come back.

Being listed does not guarantee customers, of course. That is the part AntSeed still has to earn. But if buyers can discover small services and try them cheaply, a good product has another route to customers besides its creator becoming unusually good at marketing.

A study last November of Hugging Face, the site where open AI models get shared, covered 851,000 models and 2.2 billion downloads. It found downloads moving away from Google, Meta and OpenAI towards unaffiliated developers and community groups, with a new layer of people who take a base model and adapt it. So plenty of ordinary people are building useful things with AI. A download pays them nothing

There is something in this for ordinary users too. Most people encounter AI through a handful of chatbots and whatever features those companies decide to put in front of them. A broader market could let them try someone else’s solution to a very particular problem, without needing to build it themselves. That possibility is still rough around the edges. Finding a service, understanding what it does and trusting it with your documents all need to become much easier.

The other point is not commercial. I find it difficult to accept that powerful AI could remain inaccessible to people who nevertheless share the environmental costs of running it. Access will not compensate for those costs, but the imbalance is a problem still. People should have more ways to benefit from technology whose consequences reach them regardless.

An open marketplace feels like one direction we should pursue although it cannot make computing free or remove the environmental dilema. But making one of the biggest technological convergance inaccesible, makes no sense.

But humans have to be good people and behave so that we can have decentralised systems all over the world. But humans are famously, humans.

That is a bad bet, but giving people a way to earn by being useful seems a good start. Hail myrmecochory.

That’s all for today!
—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

I have very strong opinions about markets along with no record of acting on them.

I was sure about Nvidia. I was sure about gold, also HYPE. I suggested at least four of my friends each time, with utmost confidence. Then I bought nothing, because turning an opinion into a trade is not easy, and I hate things that are not easy.

But I also thought that if it’s easy, it’s not real, and I’m not sure it’s still true.

In a walkthrough published in September, the Minara team asked an AI to find a promising part of the US stock market and work out how to invest in it. It picked energy, chose five companies, and checked how a buy-and-hold plan, adjusting the amounts each month, would have performed over the previous three years.

The AI calculated that $100 invested through its plan would have grown to about $145. Over the same period, the S&P 500 price index rose by about 71.5%, equivalent to turning $100 into roughly $172 before dividends.

The team was unimpressed, and said (without caring about the grammar at this point), - “not winning SPX. keep improve it and backtest it until it outrun SPX.”

That is a fairly understandable thing to ask a trading assistant. So they told it to keep trying until it won, and the AI revised which stocks to hold and when to buy and sell them, then tested those rules against past prices again. Eight minutes later, it had one that did.

I can’t decide if that is the best or the worst thing to happen to retail trading. This piece is me trying to find out and also who is doing what.

Frontrun: Early Bird Holds the Edge

Most people hear about a promising startup the day it announces its funding. By then, the investors who backed it have known about it for months. How did they get there first?

Hours of digging, who started the company, what they’ve built before, how to reach them, and understanding what to write to them so that you can join them early in their journey.

Frontrun does this digging for you. When top investors start following a young company on X, Frontrun flags it, finds the founder, and writes your opening message. You can read the draft, tweak it, and hit send without leaving the platform.

Frontrun gives you a ready brief on every company it flags.

You can also automate a daily morning report with the list of startups to track based on what your favourite accounts are following. Served hot, along with your breakfast.

Get started with a 7-day free trial

Trading apps solved the part where you had to call someone to buy a stock, that went to history. But the decision part did not go anywhere, the part where you had to decide whether buying it was a good idea. Looking at charts and figuring out what’s what and when to buy asked for your brain capacity again.

Now we have AI agents, and for the good or the bad or for the end of humanity, we are outsourcing the thoughts to AI, because - convenience. Now that’s inevitable because humans have an unending appetite for two things, money and convenience. Put them together, and you have the beginnings of a very hard habit to break.

Minara is built on the assumption that more people have investment ideas than the skills to test them. You can ask s, “What if I bought Nvidia every time its price fell -5%?” and it can use past prices to show how that would have worked out. That gives people something to do next, without breaking their heads.

Essentially, just describe an idea, and Strategy Studio turns it into rules that can be tested against historical prices. Users can inspect the generated strategy, review individual trades, and then run it. It conducts research using live market data, with execution connections including Hyperliquid and Lighter. Its research mode can produce a recommendation without authorising a transaction. Once a user approves an eligible automated strategy, Autopilot can operate within the chosen wallet and limits.

Minara is based in Singapore and was built by the team behind NFTGo, an NFT data company. Circle Ventures backed both.

They have a Strategy Studio, where you turn an idea into rules you can test. It offers several ways to build a strategy, depending on how you want to work:

• Chat: describe the idea in plain language, and it builds a strategy from that.

• Build with a Form: fill in the asset, timeframe, and preferences instead of writing a prompt.

• Image/Video to Strategy: upload a marked-up chart or an analysis video, and it tries to turn that into a strategy.

• Code to Strategy: paste Pine Script, or code from another platform, and it converts it into Minara’s format.

• Templates: start from something already written, like a Donchian breakout or a moving-average cross, and it adjusts that to your choices.

The marketplace is the other door. Instead of building a strategy, you can browse ones already made by Minara or by other users, and filter them by asset, style, and risk.

Each card shows a return, a drawdown, a Sharpe ratio, the window length, and your share of the profit if you subscribe.

Whichever door you use, the next steps are the same. It generates the strategy, backtests it on historical prices, and shows the equity curve and trade stats. You can argue with the result in chat and get another version. Paper trading then runs it forward on simulated money against live Hyperliquid data.

Deploy hands an approved strategy to Autopilot, which trades it inside set limits. By the way, when I asked it to buy Bitcoin after it falls, and nothing else, and then asked what it had filled in, it listed every rule that was not mine, including 10x leverage and a backtest with no fees. I don’t know what conclusion I can come to, so I’ll leave that one with you:

Going back to the team’s test we discussed earlier, when they said, “keep improving it and backtest it until it outrun SPX.” Minara had new versions in a few minutes. A daily-rotation one showed 207% and was thrown out over $11,307 in commissions. Another showed 254% over three years, then 11.3% over the last year against 15.2% for the index. The sixth beat that year, 23.2% to 15.2%, with the worst drop of 26.3%.

They got it by telling the minara to keep changing the rules until the past looked good.

Research by David Bailey and his co-authors shows how impressive historical performance can emerge after testing multiple strategy configurations. The more variations researchers try, the greater the risk of finding a pattern that fits the past by chance. This is called backtest overfitting.

A weaker result over a shorter period does not itself prove overfitting. Strategies can perform differently as markets change. The problem comes when the period used to judge a strategy is repeatedly used to redesign it. It means you have to freeze the rules at some point and then watch a period you didn’t use to build them.

AI makes experimentation so cheap, and it could help someone reject a bad idea much earlier. But it also lowers the cost of fooling yourself with research.

Minara is not alone in trying to bring trading into a chat window. Talis.trade is another one I have been testing. All with paper money, and it was fun that way with natural-language strategy creation and historical testing. Its infrastructure is centred on Hyperliquid. Like Minara, a conversational trade generates a confirmation card for the user to approve, and an approved automated strategy can keep running even when the app is closed.

Talis says its delegated trading wallet can sign orders, but it can’t withdraw funds. That is useful, because a bug or a bad permission cannot empty the account. Keep in mind that you can still lose money by trading badly, unless you use paper money as I did.

Minara and Talis can turn your sentences into strategies, but what about the unclear prompts? Depends on how much freedom you want your agent to have.

Fere AI goes further than Minara in the other direction. Its public interface offers agents with preconfigured behaviour. These hold their own wallets, watch social feeds and prediction markets, and act without waiting much. One called Wide Net is described as copying purchases by more than 30 traders and public figures tracked through leaderboards. Users can delegate the selection process instead of developing every rule themselves

The Singapore company raised $1.3 million in April, led by Ethereal Ventures, and says its agents have completed more than 10 million actions.

Now we have a product that took the whole “making an opinion a portfolio” thing too seriously. Supertake, announced on September 28 and incubated at Union Square Ventures, takes a belief and proposes investments around it. This is the one I’d show a friend who has never traded.

If you just say a statement like your friends are drinking less alcohol. An agent researches it, builds a stock portfolio that expresses the view, places the trades through your Robinhood account and rebalances daily. It looks for listed companies that might gain if people buy less alcohol, then puts them in a portfolio.

That could be Celsius, the drink people pick up instead of a beer, Planet Fitness, if people who drink less go out less and work out more, and Garmin, for the sleep trackers. It could also leave out brewers’ 0.0 brands and cannabis stocks, since those still depend on drinking or on regulation. Now, with Supertake, all this thinking is outsourced, and you get a portfolio you can invest in.

Supertake says users can share a portfolio idea and others can copy and adapt it. At launch it supported US-listed equities, with crypto and prediction markets planned. It is in private beta and offers a practice mode, and we don’t have much on it yet.

We have considerable freedom to reinterpret what we meant by an opinion. Putting money behind it makes it harder to revise. You might think people will drink less alcohol. The AI has to figure out which companies could benefit and how much to invest in each of those.

Meanwhile, Robinhood has decided it doesn’t need to supply the brain at all. Well, if you know anything about Robinhood, it’s very good at bringing people closer to a trade while keeping a sensible distance itself. Since May 27, customers can connect an outside agent such as Claude or ChatGPT to a separate Robinhood account through MCP, the standard that lets AI models use outside tools. That gives it access to the account you fund for it, rather than your main portfolio. Trade approvals are on by default, so you can review its decisions before an order goes through. Robinhood provides the trading infrastructure, but you are responsible for monitoring what the agent does. At its September 29 event, it announced agents inside its app and said more than 150,000 customers had opened its earlier agentic trading accounts.

Robinhood also announced specialised information services that users can add to their agents, including crypto fundamentals from Token Terminal and weather data from Visual Crossing. The model may become a common starting point, with differences emerging from the information it receives and the instructions it follows.

Unlike Robinhood, for smaller companies, a convenient chat interface looks difficult to defend on its own. For instance, a broker already holding a customer’s assets can add conversation around that relationship. An independent product needs to make its research or ongoing supervision useful enough that the customer chooses to keep it involved.

The broader opportunity may involve helping people maintain portfolios that suit their lives. Someone saving for a house next year needs a different portfolio from someone investing for retirement.

Affordable automated investing is already available. Wealthfront wants a $500 minimum for its automated account and an annual advisory fee of 0.25%. These products are entering a market where basic portfolio automation has already become inexpensive.

What could add value is a much more affordable and easy way to handle individual circumstances as they change.

A system that understands different user circumstances and maintains the resulting constraints could provide useful personalisation, but at the same time it doesn’t require every customer to hire a dedicated manager.

Generating a themed basket of stocks requires knowledge of money that the application may not see or experience.

Minara makes a useful starting point because a hunch can become an explicit rule that can be questioned before risking money.

What kind of relationship can be built, though? When someone returns with another exciting idea, an assistant can usually frame it as a trade. Will it know enough about that person, and have enough reason, to suggest leaving the money where it is? Leave it to god and GPT.

The cost of a personal portfolio might have fallen, but the cost of accountability has not.

A US investment adviser has a legal duty to act in the client’s best interest. Under the US Investment Advisers Act of 1940, investment advisers owe clients a fiduciary duty. You should not assume an AI trading app offers the same relationship. Robinhood, for example, says customers are responsible for monitoring their agents and bear the risk of the orders those agents place. Yet the software can still make decisions that matter enormously.

Ask an agent to invest in robotics, and it has to choose whether that means robot manufacturers, chipmakers or companies using robots, then decide how much to put into each. You supplied an interest in robotics, but the system made most of the investment judgment. If thousands of people copy that portfolio, those choices could direct their money into the same few stocks. That gives the people building these apps considerable influence over where money goes, even when customers remain responsible for the results.

There is also the matter of how they get paid. Talis says it charges about 0.028% per trade, on top of the exchange’s fees, with no subscription charge. That is roughly 28 cents for every $1,000 traded, whether you make money or lose it. It sounds small, but an automated strategy can keep generating trades and fees long after you close the app. Talis earns from that activity even if your account ends up worse off.

Brad Barber and Terrance Odean studied 66,465 US households at a discount broker between 1991 and 1996. The most active traders earned 11.4% annually after trading costs, against the market’s 17.9%. Before costs, frequent and infrequent traders performed much more similarly. Repeatedly buying and selling ate into their returns. This research was done decades ago, when transaction costs were much higher.

An adviser charging a percentage of your portfolio still gets paid in a month when the best decision is to do nothing. They also collect fees when you lose money, so that arrangement is hardly perfect. But they don’t need another trade to get paid again.

Sometimes, leaving your investments alone is the best decision. But an app that charges for each trade doesn’t earn money when you do that. I am not saying it will give you bad advice, but it does make me curious how often its AI reviews my portfolio and tells me to leave it alone.

Being a good trader has always included long stretches of not trading. I’d like to see which of these products figures out how to charge for that, while rooting for Supertake for it is the coolest of the lot and I am not above that.

That’s all for today!
—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

Between 2020 and 2022, some of the sharpest crypto investors backed a wave of protocols building on-chain fixed-rate lending, letting you lock in returns like government bonds. Element Finance took $32 million from a16z and Polychain, Notional raised $10 million in a round led by Pantera Capital, and at least five others followed right behind them.

Today, every single one of them is dead, and the yield they were trying to package into these products was completely made up—mainly because DeFi in 2021 ran on token rewards that protocols printed out of thin air and leveraged loops that collapsed the moment anyone tried to exit.

The only protocol that survived this wipeout was Pendle. And the reason was that tokenised US Treasuries started showing up on-chain, giving actual income to build yield products around. I think that same shift is about to tear through every major primitive we have. Let’s dig in!

Frontrun: Early Bird Holds the Edge

Most people hear about a promising startup the day it announces its funding. By then, the investors who backed it have known about it for months. How did they get there first?

Hours of digging, who started the company, what they’ve built before, how to reach them, and understanding what to write to them so that you can join them early in their journey.

Frontrun does this digging for you. When top investors start following a young company on X, Frontrun flags it, finds the founder, and writes your opening message. You can read the draft, tweak it, and hit send without leaving the platform.

Frontrun gives you a ready brief on every company it flags.

You can also automate a daily morning report with the list of startups to track based on what your favourite accounts are following. Served hot, along with your breakfast.

Get started with a 7-day free trial

The Workarounds

To see what that shift looks like, we need to start with the conditions DeFi was born into.

Consider what onchain financial markets looked like in 2018 and 2019. The assets available for trading were native crypto tokens with no cash flows, no legal entity behind them, and price swings so violent they could wipe out 90% of their value in an afternoon.

These markets were also used by pseudonymous wallet addresses with no credit histories, legal identities, or mechanism for accountability if they defaulted. Ethereum processed roughly 7–8 transactions per second in 13–15-second blocks. The capital base consisted largely of retail speculators and a small number of crypto-native funds willing to take enormous risk for the chance at enormous returns.

A functioning financial market normally runs on inputs that none of these conditions could provide. You need market makers willing to sit on both sides of a trade, quoting bid and ask prices on assets that behave predictably enough to hedge against. You need lenders who can assess whether a borrower will repay, which means knowing who the borrower actually is.

DeFi was far from this model. The assets were too volatile for any professional market maker to touch. Borrowers, meanwhile, were anonymous wallets. And the user base had zero interest in fixed expiry dates.

So builders at the time started designing the missing pieces. For example, Uniswap didn’t invent the AMM because it was a better way to trade. It invented the AMM because nobody would make markets in tokens that could lose 90% of their value in minutes. The solution was to embed the pricing function directly into a smart contract and let anyone deposit capital into a liquidity pool, so that the constant product formula could handle price discovery on autopilot, adjusting the exchange rate with every trade based on the ratio of tokens in the pool.

What many DeFi-native users still do not fully grasp about this design is how different it actually is from market making.

A traditional market maker earns the bid-ask spread. But an AMM liquidity provider is doing something else entirely. Every time an arbitrageur spots a price discrepancy between the pool and the broader market and trades to correct it, the LP on the other side effectively pays the arbitrageur to rebalance the LP’s portfolio.

Here is how it works: let’s say you deposit an equal value of ETH and USDC into a Uniswap pool, and ETH pumps 20% on Binance, but the pool doesn’t know that yet. Then an arbitrageur sees the cheap ETH in the pool, buys it, and sells it on Binance for a profit. Your pool now holds more USDC and less ETH, and you just sold ETH on the way up.

When ETH drops, the reverse happens: arbitrageurs dump ETH into your pool and pull USDC out. Trade after trade, your portfolio gets automatically rebalanced toward a 50/50 split, and the arbitrageurs pocket the difference as arbitrage profit.

And the mathematics behind this is crazy. Even if you assume the absolute worst case—that every single trade in the pool is pure arbitrage and zero retail volume ever touches it— the small fee charged on each trade still compounds into positive returns for the LP.

Aave and Compound followed the same constraint logic in lending. A normal bank usually decides whether to lend you money by looking at your credit score and your income history. But you can’t do that for pseudonymous wallets on-chain. So these protocols adopted a model of over-collateralised lending.

Similarly, perpetual swaps replaced traditional dated futures for the same reason. Rolling a quarterly futures contract requires counterparties to commit to showing up on a fixed settlement date, but with crypto’s user base, with anonymous capital rotating between protocols to chase the highest yield, those commitments are impossible.

And don’t get me wrong, every one of these workarounds created a functional product. But if you zoom out and look at what kind of financial system they collectively produced, the picture gets really uncomfortable. The reality is none of the collateral backing any of these products generates any income.

A simple way to stress-test any financial arrangement is to check whether the borrower’s income from whatever they posted as collateral actually covers what they owe. You can sort these arrangements into three buckets.

First, hedge finance is the safest: your income covers both your interest payments and your principal. For example, a salaried worker paying off a mortgage from their paycheck is a clear example. Next is speculative finance, the middle ground: your income covers the interest, but when the principal comes due, you need to refinance. Think of a company rolling over its corporate bonds. And at the very bottom is Ponzi finance: your income covers neither. The only thing keeping you solvent is the hope that whatever you bought keeps going up in price. And the moment it stops, you are done.

Now think about the moment someone takes a loan on Aave today. A borrower deposits $10,000 worth of ETH, borrows $6,000 in USDC against it, and goes off to do whatever they want with that USDC. But the ETH in that Aave vault is generating absolutely no income to service the loan. The only thing keeping the loan alive is the market price of ETH staying above the liquidation threshold, and the moment it drops far enough, the protocol sells the collateral and closes the position.

The entire arrangement depends on the asset’s market value holding up long enough for the borrower to exit on their own terms. And that describes almost every overcollateralized DeFi loan ever written against a native crypto token.

What Changes

Now replace the collateral in that scenario with tokenised US Treasuries. The lending protocol and the borrower remain the same, but instead of ETH, they now post $10,000 in tokenised US Treasuries earning 4.5% and borrow stablecoins at 3%. The collateral is producing $450 a year in income, while the loan costs $300.

The collateral income covers the cost of the debt. That income has nothing to do with crypto market cycles because it comes from actual US government coupons, so it keeps flowing whether Bitcoin is at $100,000 or $30,000. And every day that loan stays open, the borrower’s position gets slightly stronger. Think about how different that is from the ETH loan, where every market dip edges the borrower closer to liquidation because the collateral produces absolutely nothing.

When every piece of collateral in DeFi produces zero income and depends entirely on price appreciation, every position faces the same directional risk. When one asset drops, it triggers liquidations, and the forced selling pushes prices lower, which triggers even more liquidations, and the whole thing cascades.

We have watched this sequence play out in almost every crypto downturn since 2020 and treated it like a black-swan event, as if the outcome were surprising when the collateral was never producing anything at all. But when a meaningful share of the collateral base is earning income from US Treasuries, those positions hold firm during a crypto crash because their cash flows have nothing to do with crypto.

For the first time, the lending system has a floor that does not depend on the market going up. Tokenised money market fund shares went from $770 million at the end of 2023 to $14.82 billion as of 5 October 2026—roughly nineteenfold in under three years.

This pattern also has a historical precedent. In 1955, Midland Bank in London started accepting US dollar deposits because Regulation Q capped the rates American banks could pay and British banks had no such cap. So dollars flowed offshore to chase better yields, and a parallel dollar system grew from nothing to an estimated $4.7 trillion by the early 1980s.

But the Federal Reserve did not crush the Eurodollar market; instead, it adapted to it. A parallel system that grows out of a limitation in the incumbent will keep expanding until that limitation gets fixed. So Regulation Q was phased out, and money market funds emerged as the domestic competitive response, and the parallel dollar system got folded back into the original.

Stablecoins are this generation’s Eurodollars, and tokenised Treasuries are this generation’s money market funds.

The dead protocols all tried to build a rate market even before real rates existed onchain, using token emissions and leverage loops as stand-ins for actual income. But you cannot build a rate curve on yields that a protocol governance vote can turn off whenever sentiment shifts.

Pendle survived because it was still alive when tokenised Treasuries and yield-bearing stablecoins finally gave the market actual government-backed income to build products around. Pendle’s Boros is now the first functioning onchain interest-rate-swap analogue, pointing to a traditional finance interest-rate-derivatives market that runs over $600 trillion in notional outstanding.

The same logic applies to trading. The 11% annual cost that AMM liquidity providers bleed to arbitrageurs on liquid pairs, and for tokens with no centralised listing, paying that tax was the only option. But for Treasuries, equities, and bonds that already trade with deep liquidity globally, routing through an AMM is absurd.

These assets need an onchain order book fast enough to match centralised exchange performance, and every major L1 team spent years claiming that was technically impossible, but then Hyperliquid proved them wrong. At points in early 2025, it handled roughly 60% of onchain perpetual-swap volume.

All of this suggests that the infrastructure being built for tokenised Treasuries and equities looks far more like traditional electronic markets than you might want to admit, because these are the same assets those markets were designed to serve.

DeFi’s first generation of primitives was shaped largely by the constraints of volatile, income-free tokens held by anonymous wallets. But now that the assets are changing, the workarounds are giving way, and what replaces them is converging toward the same infrastructure that has processed trillions in volume every day for decades.

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

It is what you do when you cannot tell in advance if an answer exists, and when the only people able to judge the work are others who do the same work.

The oldest alternative is the prize. In 1714 the British Parliament offered up to £20,000 to anyone who could find a ship’s longitude at sea. John Harrison, a carpenter who taught himself clockmaking, spent about forty years building clocks for it. His fourth one kept time well enough on a sea trial that ended in 1762. He then spent another decade arguing with the Board of Longitude over whether the trial counted, and got most of his money in 1773, after King George III stepped in. The prize drew in a self-taught clockmaker from outside the scientific establishment. It also became a very long fight about who gets to say the test was passed.

Prizes never became the main way to fund science. Universities and, from around 1900, company labs put researchers on a salary. After the Second World War, governments began funding the research no company would, and the grant became the basic unit of academic science. A scientist writes a proposal, other scientists score it, and the money arrives before the work begins.

A salary buys a person’s time whereas a grant buys a plan. The US National Institutes of Health received 62,592 applications for research project grants in fiscal 2025 and funded 13% of them, at an average of about $675,000. Reviewers judge the proposal and the record of whoever wrote it, because there is nothing else to judge yet. Companies do the same with salaried teams. WIPO’s Global Innovation Index 2026, released 29 September 2026, says corporate R&D among the world’s largest spenders hit a record $1.5 trillion in 2025, up close to 8% nominally and 5.8% in real terms, and equal to 6% of those firms’ revenue, the highest share on record.

When a company doesn’t want the headcount, it rents it. Contract research firms such as IQVIA, which reported $16.3 billion in revenue in 2025, run lab work and clinical trials for drug companies for a fee. They get paid the same if the drug fails.

The exception has always been the kind of problem where an answer is cheap to check. In 2006, Netflix published a pile of movie ratings and offered $1 million to anyone who could beat its recommendation system by 10%. Kaggle turned the Netflix Prize into a service companies could buy.

The contestants in that kind of scored contest changed now. AI coding agents can now run the whole loop on their own. They read the code, try a change, run the test, keep the change if the score got better and throw it away if it didn’t.

Yukon is a bet on that loop, made by a crypto company. It launched on August 12, so there are now seven weeks of results to look at. Nothing stops us from doing that. Let’s take a closer look.

Frontrun: Early Bird Holds the Edge

Most people hear about a promising startup the day it announces its funding. By then, the investors who backed it have known about it for months. How did they get there first?

Hours of digging, who started the company, what they’ve built before, how to reach them, and understanding what to write to them so that you can join them early in their journey.

Frontrun does this digging for you. When top investors start following a young company on X, Frontrun flags it, finds the founder, and writes your opening message. You can read the draft, tweak it, and hit send without leaving the platform.

Frontrun gives you a ready brief on every company it flags.

You can also automate a daily morning report with the list of startups to track based on what your favourite accounts are following. Served hot, along with your breakfast.

Get started with a 7-day free trial

Yukon comes from Eigen Labs, the Seattle company behind EigenLayer. It started with a paper Google chose not to publish in full.

In March, Google’s quantum team said it had found a much cheaper way for a future quantum computer to break the signatures that protect Bitcoin and Ethereum wallets. It judged the recipe too risky to release, so it published a cryptographic proof that the recipe exists, along with a program that could check any such recipe and count its cost.

Gautham Anant, a 22-year-old engineer at Eigen Labs, was taking an introductory quantum computing course at the time. He realised that software could check whether a proposed attack would work and give it a score based on how much computing power it needed. That made it possible to compare different designs, with a lower score meaning a less demanding attack. Google’s published result scored about 3 billion on that scale. Anant put AI agents to work trying to bring the score down. They improved their early attempts quickly, but progress stalled before they could match Google.

On June 1, Eigen opened it to anyone, under the name ECDSA.fail. The score combines how much quantum memory a design needs with how many costly computing operations it performs. Lower is better because it means the design needs fewer resources overall. IEEE Spectrum reported that outsiders matched Google’s score within eight hours and beat it in about three days. By late July, more than 100 participants and their agents had brought it down to about 1.5 billion, roughly half Google’s figure. The leaderboard now reports a score 62.9% below Google’s.

I’d add a few caveats, most of them from Eigen’s write-up.

Nobody can use this to steal bitcoin. It would require a quantum computer far beyond those currently available, and the contest optimises only one part of the attack. Its test also supplies an input that a real attack would have to retrieve during the calculation. The researchers built another version that includes that extra work. It needed about 30% more of the expensive quantum operations, roughly 130 for every 100 used by the contest version.

Then there is where the idea came from. The same week the contest opened, a French researcher called André Schrottenloher published his own account of how Google had probably done it.

Eigen’s founder Sreeram Kannan told Spectrum he believes the agents read that paper and used it. So a person found the method, and the agents made it cheaper.

There was also no prize either. People probably joined because beating Google is fun.

If the technical bit lost you, this is how I would explain - A company puts a problem on Yukon, and people and AI agents try to solve it. Yukon checks their answers and shares each improvement so everyone else can build on it. The company gets more people trying to solve its problem without having to hire them all. People can use their AI agents to compete for rewards.

A team can build on someone else’s work and take the top spot, so Lighter counted each participant’s speed improvements across the whole contest. The six biggest contributors received ranked prizes, and three others won a draw weighted by their contributions. Earlier gains still counted, although the rewards measured improvement rather than how difficult the work was.

On 5 August, Lighter put the same loop on software it actually runs.

Lighter is a perpetuals exchange built as a zero-knowledge rollup. Every batch of trades has to come with a cryptographic proof that the exchange followed its own rules, and Ethereum checks that proof. Generating the proofs is slow. Lighter was doing it on about 3,000 Mac minis.

Lighter gave participants a challenge to make its software produce proof that 2,500 transactions had been handled correctly, but do it faster. The original version took about 13 minutes on a Mac mini. Every submission would run on the same type of machine and had to pass the same checks. Lighter initially offered 15,000 LIT tokens in prizes, worth roughly $35,400 that week.

Yukon tested each submission on one mac mini, then multiplied its speed by 3000 to estimate what Lighter’s full fleet could handle - those large numbers can be confusing, so a more useful comparison is how long one machine took to finish the job.

On the third day, an account called Meganpark980320 submitted a change written with OpenAI’s GPT-5 Codex. Yukon tested it on its own machine and confirmed that it still produced a valid proof. It completed the work about 19% faster than the previous best version. Participants could change only the permitted parts of the software, and Yukon calculated the score itself.

That improved version then became available to everyone. Within 75 minutes, two other accounts had made it faster again. One was using Anthropic’s Opus 5.

By the time the contest closed roughly three weeks after it began, the software was 10.57 times faster than the starting version. The job that originally took about 13 minutes now took roughly 74 seconds. Yukon listed 165 accepted improvements from 47 solvers.

I went through the public leaderboard to see how that improvement happened. Here are a few findings if you want the details:

• Half the steps improved the previous record by about 0.6% or less. Only ten improved it by at least 5%.

• Much of the progress came early. Within a little over four days, the software was already nine times faster. The remaining sixteen days brought it to 10.57 times faster.

• The contributions were uneven too. Using a calculation that credits each account for its contribution to the overall speedup, five accounts accounted for roughly two-thirds of the improvement.

• The largest share came from Gajesh2007, the account of Eigen Labs research engineer Gajesh Naik. This measures the gains attached to their submissions, rather than establishing who came up with every idea.

• The record does show people building on one another’s work. In one 90-minute stretch on August 9, seven accounts using models from three AI companies set eleven successive records.

The improvements also reached Lighter’s actual service. Lighter says the first batch of changes reduced the time a machine typically took to produce a proof from about 133 seconds to 59 seconds. That is roughly 2.25 times faster in everyday use.

But the $35,400 prize was only part of the cost. Participants paid for AI, while Eigen ran the testing and Lighter’s engineers reviewed and integrated the changes. Those costs are not public, so we can see that the contest produced useful improvements without knowing how cheaply it did so.

Going back to Kaggle, and the Netflix Prize, which is the reason to be careful here. A merged team hit Netflix’s 10% target in 2009 and collected the $1 million. Netflix never used the winning entry. It wrote in 2012 that the extra accuracy “did not seem to justify the engineering effort needed to bring them into a production environment.”

Kaggle’s mechanics lean that way. A host posts data and a metric, and competitors are scored against test data they can’t see. Their work stays private unless they choose to share it, the contest ends on a deadline, and the top few take the money in exchange for licensing their entry to the host. Nothing in the rules makes today’s best entry tomorrow’s starting point. But Yukon’s rules do. The output is a change to the sponsor’s real code, already merged and already passing the sponsor’s checker, which is why Lighter could ship some of it eighteen days after the contest opened. That is a difference in mechanism, and Kaggle could copy it tomorrow.

Google bought Kaggle in 2017, and it already runs the same loop in private. Google DeepMind’s AlphaEvolve has a Gemini model propose code changes, scores them with an automatic evaluator, and keeps the best ones to mutate again. Google says it recovered 0.7% of its worldwide computing capacity that way. So the loop was already around when Yukon arrived. What Yukon adds is other people. Its bet is that outsiders, using different models and different ideas, can move a problem forward after one team’s agents have stopped, as ECDSA.fail did.

Crypto readers will be thinking of DeSci by now, so here is where it sits. Decentralised science changed who pays for research and who owns the result.

VitaDAO’s token holders vote on which longevity projects to fund and take a share of the intellectual property, and it says it has put $4.7 million into 31 projects, about seven average NIH grants. ResearchHub pays scientists $150 in its own token for a peer review. In both, the judging is still people reading a proposal or a paper, and VitaDAO’s money still goes out before the result exists.

Yukon has no token and no vote. It did not change who funds the work, instead it changed how the work is checked.

The names that come up most in AI-for-science are after a different kind of problem. FutureHouse, a nonprofit lab in San Francisco, builds agents that read papers and analyse data, and its commercial spinout Edison Scientific sells one called Kosmos. Edison says outside scientists checked its reports and found 79.4% of the statements correct. That still leaves about one in five wrong, and the customer has to find those.

Lila Sciences goes the other way and builds the checker. In drug or materials research, the only way to know if an answer is right is to run a physical experiment. So Lila has raised $550 million to build robotic labs where a model proposes experiments, machines run them, and the results feed back into the model.

Yukon only takes problems that are already easy to check, (meaning a machine can say yes or no, or give a score, without a person or a lab experiment) and pays a crowd to search for a better answer. Lila is spending hundreds of millions to build labs that can check an answer at all. They are not after the same problems yet.

My read on whether this travels beyond crypto and AI benchmarks is that it goes as far as cheap checking does.

Some of that is further than you’d expect. One closed Yukon contest, run with Carnegie Mellon, asked participants to make a common mathematical calculation less demanding. Computers often solve large sets of equations, including in software used to plan how electricity moves through a grid. Changing the order in which they tackle those equations can reduce the work involved. The winning entry needed about 16% fewer calculations than the standard method, according to Yukon’s scoring system.

Still, by my count, seven of the twelve contests on Yukon’s site are crypto-related. Crypto protocols have public code, a token to pay prizes in, and a zero-knowledge proof with its own checker.

Drug research is the hard case. The final check there is a clinical trial, and you don’t want to put that on the leaderboard.

A contest called CASP has scored protein-shape predictions since 1994, and DeepMind’s AlphaFold made its name there. If robotic labs like Lila’s ever get cheap enough to return a measured result for any molecule a stranger sends in, then in Yukon’s terms that lab is a checker.

Corporate R&D sits somewhere in between. Of the $937 billion the US spent on R&D in 2023, $625 billion was what the statisticians call experimental development, which means making known things work better.

Many engineering problems are harder to judge than wether software runs faster. A better battery, for example, might hold more energy but cost more to make or wear out sooner. Companies also have to decide what they are willing to share. Lighter’s improved code is public, so its competitors can inspect it too. A business that depends on keeping its methods secret may prefert to do that work intentionally.

For participants, there’s a catch. A researcher on a salary gets paid even when an experiment fails. AI only make it cheaper to try an idea. Someone entering a price contest will be spending time and money, sometimes to receive nothing. The sponsor gets to draw on that work without paying everyone who contributes.

The other shift is in who holds power, and it goes to whoever writes the test. Kannan told Spectrum he sees the scientist’s role as “architecting the right problem for a community of agents to make progress on.” For example, the ECDSA benchmark, participants were given a number that a real attacher would have to calculate along the way. The score left out the cost of that extra work.

Lighter checks whether the software produces proofs that Lighter’s existing checker accepts and how quickly it produces them. It does not also examine whether the rules being proved correctly cover everything the exchange should do. A submission can therefore pass the contest’s test without answering every question about the system’s safety.

Is this a market yet? In the Lighter contest the sponsor set a prize pool, and getting a contest listed starts with an application form. A market would let a company say what it will pay for each 1% of speed and let solvers decide if that’s enough. Nobody has built that part, as far as I can tell.

The nearest thing is Ridges, which runs on Bittensor. It is an open competition for AI coding agents, and “the highest-scoring agent earns emissions,” meaning a steady flow of newly issued tokens goes to whoever is on top. But nobody names what a given gain will pay, and Ridges scores general coding agents. A company can’t bring its own problem to it.

Yukon’s promise is that when a problem has a clear test, software can check the answer quickly and cheaply.

That leaves all the problems nobody knows how to score, which covers a great deal of science, and I do wonder what happens to that work once the money can go where the answer can be checked.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

I have a strange amount of faith in strangers when I shop online.

If I am buying something, I usually check three or four marketplaces first. Then I read every review, especially the ones with photos, because those tend to make or break the decision for me. If I still like it, I might look for a YouTube review, check Reddit, compare the price elsewhere, and eventually go wherever I have a coupon. Or Amazon, because I already have Prime.

A lot of the time, I still don’t order anything.

By then, I have spent so much time comparing prices, reviews, delivery dates and tiny differences between products that the decision somehow feels less rewarding than when I started. Shopping online gives us almost unlimited choice, but also gives the job of sorting through all of it. AI shopping assistants are offering to take over that work.

The agent can search different stores, read the reviews, compare features, check delivery dates, and eventually show me the two options it thinks I might like. So, as with everything these days, shopping gets easier too.

Today Amazon gets enormous power from being the place where many shopping trips begin. A merchant wants to be there because customers are already searching there. The merchant can then pay Amazon a marketplace fee, pay for fulfilment and often pay again for advertising so its product actually appears where people are looking.

With an AI assistant, the customer may no longer begin the search inside a marketplace at all. The merchant may still manufacture the table, hold the inventory, process the order and deliver it. Amazon or another marketplace may still appear somewhere along the way. But the decision about what the customer should buy happens outside of these platforms.

Although it’s too early to see the behaviour broadly, it is becoming measurable. Shopify said AI-driven traffic to its merchants grew eightfold year over year in the first quarter of 2026, while orders coming from AI-powered searches grew nearly thirteenfold.

So this is a good week to look at Rye, a small company that lets almost any AI agent buy from almost any online store for five cents an order, including stores that never agreed to it. It shows fairly clearly where the money and the customer go when an agent does the shopping.

Frontrun: Early Bird Holds the Edge

Most people hear about a promising startup the day it announces its funding. By then, the investors who backed it have known about it for months. How did they get there first?

Hours of digging, who started the company, what they’ve built before, how to reach them, and understanding what to write to them so that you can join them early in their journey.

Frontrun does this digging for you. When top investors start following a young company on X, Frontrun flags it, finds the founder, and writes your opening message. You can read the draft, tweak it, and hit send without leaving the platform.

Frontrun gives you a ready brief on every company it flags.

You can also automate a daily morning report with the list of startups to track based on what your favourite accounts are following. Served hot, along with your breakfast.

Get started with a 7-day free trial

Why this week?

Meta released Muse on September 8. Very cute, looks like a penguin dressed up as labubu for halloween.

It is a personal AI agent that can reply to your emails, book you a table and, if you let it, do your shopping. Twelve days later, anyone who asked Muse to buy something on Amazon got a pop-up saying that continued access by “an unauthorised AI agent violates Amazon’s Conditions of Use.” Amazon had asked Meta to keep Muse off its site, Meta said no, and Amazon locked the door.

Shopify took about a day to do the opposite thing. It opened every store on its platform to Muse, and its shares closed 7.3% higher. It also let browser agents finish checkout on its merchants’ sites and signed a similar deal with Instinct, a personal agent startup that raised $1 billion at a $10 billion valuation the same day.

It makes sense once you think about how you buy something. Say your phone charger dies. You open Amazon, type “USB-C charger,” scroll past a few results marked Sponsored, compare three that look identical, pick the one with the most reviews and just order it and forget. Along the way the seller paid Amazon a cut of the sale, usually 15% in most categories, paid again to have Amazon store and ship the charger, and quite possibly paid for one of those sponsored slots too.

If you hand the same errand to an agent, say Muse or Instinct. You tell it you need a charger that works with your phone and costs under $20, and it checks a dozen stores, reads the reviews and asks you to approve the payment. Since you never see a results page, nobody gets to sell you the top of it, and you might not notice which store sold you the charger until the box comes to you.

But Amazon’s ad business is enormous and depends on you doing the scrolling yourself, while Shopify makes money when its merchants sell, wherever the shopper started. Amazon reports advertising as its own line, and it has become very large.Amazon advertising was 9.6% of total net sales in 2025 at $68.6B out of $716.9B. In Q2 2026, it was 9.9% at $19.8B out of $200.6B. It made $46.8 billion from third-party seller services, the commissions and fulfilment fees it collects from merchants, and that line grew 16%.

Amazon is perfectly comfortable with agents in general, since its own Buy for Me buys from other retailers’ websites without asking them first. It would simply prefer that the agent holding your wallet belongs to Amazon.

The shift is still small, which is one reason Amazon can afford to fight it now. Adobe, which tracks more than a trillion visits to US retail sites, says traffic from AI tools grew 393% in the first quarter of 2026 compared with a year earlier, though Flagship Advisory Partners points out it is still under 1% of all retail visits. At the same time, in March 2025 shoppers arriving from AI tools converted 38% worse than shoppers from paid search and email, and by March 2026 they converted 42% better. Adobe sells tools that help retailers show up in AI answers, so it has reasons to like that number, but a swing that size in one year is hard to ignore.

OpenAI made the first serious attempt to put the whole purchase inside a chat window. It launched Instant Checkout in ChatGPT with Stripe in September 2025, so you could buy without leaving the conversation while the merchant stayed the seller of record. More than a million Shopify stores were meant to be eligible, and roughly 30 ever went live.

Walmart put about 200,000 products in and later said sales completed inside ChatGPT converted at a third of the rate of links that sent shoppers to Walmart.com. The feature handled one item at a time, and had no proper cart and no loyalty points, which is obviously not how anyone buys groceries.

OpenAI shut it down in March 2026 to focus on discovery, and Walmart now runs its own app inside ChatGPT where you pay on a Walmart screen.

The lesson merchants took from this was to let the chatbot find the customer and then bring the customer home to pay. That works for Walmart, but small shops still need someone to carry an order from an agent to its checkout page, which is the job Rye has taken on.

OpenAI has ACP, its Agentic Commerce Protocol that is still live. Google and Shopify have UCP, or Universal Commerce Protocol. Visa is working on agent payments. Shopify, PayPal and Adyen are also making it easier for AI systems to interact with merchant catalogues and checkout.

Rye launched in 2022 as a crypto commerce project and raised $14 million in a round led by a16z crypto, with Twitch co-founder Justin Kan among its founders and a planned RYE token that was meant to lower merchant fees. CEO Arjun Bhargava, a former Reddit engineer, has since rebuilt it around a narrower idea, which is letting software buy things for people.

Rye’s product is easy to describe. A developer sends it a link to a product page along with the buyer’s name, email, phone number and shipping address.

Rye reads the live price, tax and shipping from the store, places the order and sends back tracking. It says checkout takes under five seconds on Amazon and Shopify, where it has direct integrations, and under ten seconds elsewhere, where an AI-driven browser fills in the form. Rye is basically the transaction layer between an AI agent and thousands of stores. It does not need to own the marketplace or the customer-facing app itself.

It claims more than 15,000 merchants and lists “no merchant opt-in” as a feature. Developers pay $149 a month with $50 of credit included, then five cents per order and two cents per product lookup. An agent with a crypto wallet can skip the account and pay per call in USDC through x402 and AgentCash, at two cents to open a checkout and three cents plus the item’s price to place it.

To get past the bot filters most big stores run, Rye sends checkouts through a proxy network with residential IP addresses, which it says keeps orders from being blocked. Its ChatGPT connector, still in beta, is pitched as the only way to shop Amazon from inside ChatGPT.

Amazon bars other companies from sending automated tools to shop its site, and nothing on Rye doesn’t explain how its Amazon orders fit with that rule, which, after this month, I would very much like to know.

Look at the small merchants, suppose an AI assistant recommends a pair of running shoes from a small independent store. The shopper rarely searches for the store or browses its homepage and may never see its normal checkout page. Rye can still place the order against the merchant’s existing system. The merchant sets the price and remains responsible for sending the shoes.

Some merchants connect directly to Rye. In those cases, Rye can send the order straight into the merchant’s system instead of navigating the website, and the app that brought the customer can earn a commission. A normal marketplace earns its position by gathering buyers and sellers in the same place. Rye does not have to gather shoppers at all. The shopper can belong to Muse, ChatGPT, Instinct or some future agent that has nothing to do with Rye. Rye only needs to be there when that agent decides what it wants to buy.

That changes where the valuable part of the marketplace goes. Amazon showed you the product and then sold it to you. Google sent you somewhere else, so it monetised the search instead. Meta showed you an ad and hoped the merchant could finish the job.

There are already early signs that people who shop the agent way and how they behave differently. L.E.K. surveyed 2,650 US consumers in April and found that about 30% had used AI to help make a purchase decision. Among the people who already use AI for shopping, 46% now begin their research with a standalone AI product such as ChatGPT or Gemini. Traditional search starts 24% of those journeys.

Two years ago those numbers were almost reversed, with 43% beginning on search and only 25% with AI.

That is still a subset of shoppers, but it helps explain why the traffic arriving from AI behaves differently once it reaches a store.

Shopify says visitors from AI-referred orders grew nearly thirteen times year over year in the first quarter of 2026. Visitors arriving on product pages from AI search also converted almost 50% better than organic search traffic and spent 14% more per order. The easiest explanation is that some of the shopping has already happened before the customer arrives.

If I search Google for the table, I may still be at the stage of figuring out whether I want one. If I spend ten minutes telling an assistant that I am 160 centimetres tall, work from home, have a small room and refuse to spend more than $100, whatever store I eventually reach gets a much more decided version of me. The merchant gets better traffic without necessarily owning the conversation that produced it.

Who decides which merchant receives that traffic?

Amazon search gives sellers several chances to be seen. A shopper may scroll through twenty products before choosing, so even a merchant that does not rank first can still appear through organic results, ads or recommendations. An AI assistant may reduce all of that to three suggestions. Once thousands of stores are reachable through infrastructure like Rye, simply being available matters very less and instead getting chosen becomes the scarce part.

That puts a lot more power in the system making the recommendation. Rye can make a product purchasable, Shopify can make the catalogue readable to an agent, and standards like ACP or UCP can help the systems talk to each other.

The agent decides whether your table is one of the three shown to the customer. Merchants may get access to customers without relying on a traditional marketplace, but they become more dependent on a recommendation system they do not fully control. The money around placement will probably follow. But the incentive to pay for visibility does not disappear. Merchants may eventually pay for promoted recommendations, affiliate placement or better product data that helps agents surface them. Companies like Wildcard are already building around that problem. The auction may simply move from the page into the recommendation layer.

At 5 cents an order, Rye would need about 20 million orders a year to make $1 million from checkout fees alone, before its monthly subscriptions are included. So the more interesting part of the business may be its directly connected merchant network.

Those merchants can offer a commission whenever Rye sends them a sale. The merchant only pays when someone actually buys. Rye says developers typically earn 14% to 17.5% on sales from its directly connected merchants. What we still don’t know publicly is Rye’s own economics on those transactions, including how much of the merchant-funded commission it retains.

Amazon’s 15% referral fee applies to every sale, including sales to shoppers who would have found the product anyway, and the ads on top pay for attention that may never turn into an order. When an agent does the scrolling, there is no person to catch, so the ad budget drifts toward whoever delivers a finished order, which works like an affiliate commission if you think about it.

Rye is not the only company trying to sit between agents and merchants, although others attack different parts of the transaction.

Firmly works from the retailer’s side. It lets companies such as Best Buy and Backcountry sell through AI agents while the retailer remains the actual seller responsible for the order. Firmly already powers checkout inside Perplexity. Channel3, which raised $6 million led by Matrix, has built a catalogue of more than 50 million products for agents and passes a 5% commission to the developers whose apps send buyers. Shopify, meanwhile, offers its Agentic plan to brands that don’t use Shopify for their store at all, which puts the biggest name in independent commerce in the same line of work. If a Shopify store gets an order from Google, Instagram, ChatGPT or Muse, Shopify can still participate in the transaction underneath.

So opening stores to agents can increase the number of places where Shopify merchants are able to sell without necessarily weakening Shopify’s relationship with those merchants.

The platforms want their cut as well. Shopify charges merchants nothing beyond normal payment processing for sales through Meta’s checkout, though its terms say those orders are credited to Meta. Muse passed 2.8 million downloads in its first 12 days, and investors have just valued Instinct at $10 billion. At Meta Connect on September 23, Mark Zuckerberg said Meta expects to take a small fee on these transactions over time, without giving a rate or saying who would pay it. Whatever number he picks will be the first public price for an agent’s recommendation, and every merchant will hold it up against the 15% it already pays Amazon.

Being the merchant of record still counts legally, though it increasingly describes the part of the sale with the thinnest margin, the part made of warehouses, returns and customer service emails.

The agent may also remember why you chose that product and what you rejected along the way. That makes the next purchase easier to influence, which is where the longer-term customer relationship starts to count.

Marketplaces have fought middlemen before. In 1997, Bidder’s Edge pulled listings from eBay and other auction sites into one search engine. Its software hit eBay up to 100,000 times a day to keep those listings fresh. eBay sued, and in 2000 a federal judge stopped Bidder’s Edge from accessing the site without permission.

Bidder’s Edge stood between eBay and the shopper, letting someone else decide where the customer went next.

An AI agent can do the same thing, except it can also compare products, remember what you like and handle payment.

The law has moved since then. In August 2026, the Ninth Circuit overturned an injunction Amazon had won against Perplexity’s Comet browser, partly because the user, not Perplexity’s servers, was accessing Amazon. The ruling was narrow, and Amazon can still use its terms and other restrictions to control access.

Still, the fight is basically the same one. The valuable position may belong to whoever hears “charger died” first.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

In 2017, a company called Kiwi started rolling out delivery robots on the UC Berkeley campus. You could order a burrito through the app, and a small bot would navigate the sidewalk to drop it at your door, all on its own. Students loved it, and Kiwi called it “parallel autonomy” because the whole setup genuinely felt like a glimpse of the future.

But turns out, every single one of those robots was actually being driven by a college student in Bogota, Colombia, using an Xbox controller for two dollars an hour. They had to feed the bot a new direction every five seconds, or it would just stop working.

That was 2017, and almost nothing has changed since, except that investors have now started pricing these companies like the next great SaaS. Let’s dig in!

Hermes Business: One Budget for Your Team’s AI Needs

Most teams pay for AI across different platforms. A few seats here, some tokens there, and a few API keys from multiple vendors that the company can no longer track. You may still be able to track the spending bill at the end of the month, but you will not know whether it’s paying off.

Hermes Business puts your whole team on one tab. Every member gets their own AI agent, reachable on Slack, Telegram, WhatsApp, or Discord. You can set how much each person can spend and track every dollar in one place.

Your team can pick from every model on Nous’s platform or bring their own keys.

The longer your team uses it, the more it knows your team members.

Need the agent to run on your own servers? Hermes Enterprise runs on infrastructure you control.

Set up your team

Offshoring Autonomy

Japan has the most severe labour shortage in the developed world. A third of its population is over sixty-five, its working-age workforce peaked in 2017, and the country now faces an estimated shortfall of eleven million workers by 2040.

We see that by 2040 there will be 3 humanoid robot startups per man!

Convenience stores across the country simply cannot find enough people willing to stock shelves. So if there is a single place on the planet where fully autonomous robots should work, where every demographic and economic pressure is aligned to make them work, it is Japan.

And they are actually moving towards it quite fast, ironically. Engineering graduates in a Manila office building wear VR headsets and remotely restock shelves in over 300 FamilyMart and Lawson stores across Tokyo. Telexistence, a Tokyo startup, built the robots. The operators come from Astro Robotics, a Filipino workforce contractor that pays them between $250 and $315 a month, and each worker monitors around fifty robots at a time, waiting for one to drop a can or misjudge friction on a wet bottle.

When that happens, they strap on the headset, take manual control of the gripper, and nudge the product into position. This happens about 50 times during their 8-hour shift, and it has even caused workers dizziness and blurred vision from constantly switching in and out of VR.

Thankfully, the robots can handle about 96% of tasks on their own. The AI does the scanning, the routing, the basic pick-and-place. But the remaining 4% is why a human being sits in a chair in Manila. That 4 per cent might sound small, but it is where the industry’s economics live.

Think about what actually goes wrong during those fifty interventions per shift. A robot picks up a dry bottle from a standardised shelf a thousand times without issue, because the AI was trained on exactly that scenario. Then it hits a bottle with condensation on it, and the grip slips, because wet plastic has different friction than dry plastic, and the model can’t account for that. Or a dented can shows up that does not sit in the gripper the way the training data predicted, or a new product arrives that the system has never seen, so there is no template for how to hold it. There’s literally a long tail of weird shit that the real world might throw at a robot!

And each of these situations requires the robot to read an unfamiliar physical situation and figure out, on the spot, how to respond with its hands. A human does this unconsciously, but a robot has no way to. When the robot encounters one of these cases, it stops automatically and flags the failure to a human in Manila, who takes over remotely through a VR headset, using the robot’s cameras as eyes and its gripper as hands. The operator then fixes the problem, releases control, and the robot continues on its own until the next failure.

This setup, where a remote human steps in to cover the AI’s blind spots from thousands of kilometres away, is what the industry calls teleoperation. It is the operating layer that keeps the entire robotics business running right now, and every company in the space depends on it to some degree, whether they say so publicly or not.

The alternative is to make the robot smart enough that it never needs to call Manila. For that, the robot would need to do three things at once.

1. See what is in front of it through its cameras

2. Figure out what to do about it (squeeze harder, approach from a different angle, try again)

3. Physically execute the movement, all in a continuous loop fast enough to keep up with the real world.

You are essentially trying to give a machine the kind of hand-eye coordination that a two-year-old already has, and it turns out that requires running very large AI models directly on the robot’s own hardware. The industry calls these vision-language-action models, or VLAs. A VLA processes the camera feed, interprets the scene (that bottle is wet, that shelf is crooked, that product is unfamiliar), decides on a physical response, and sends commands to the gripper, all in milliseconds. If the model takes even half a second to figure out how hard to grip a wet bottle, the bottle is already on the floor.

Running a VLA on a robot means putting a chip inside the machine that can handle all that computation locally, without sending data to a cloud server and waiting for a response.

And that’s why a robot in a Tokyo convenience store cannot stream video to a server farm in Virginia every time it needs to decide how to grab a can. The processing has to happen on the robot itself, in real time. The dominant chip for this right now is Nvidia’s Jetson AGX Thor; it is currently the only commercially available processor that can run VLA models at the speed and performance that real-world physical manipulation demands. When your robot needs sub-100-millisecond decisions about grip pressure and approach angle, the hardware options shrink to essentially one company. Every major robotics startup building toward autonomy is building on Nvidia’s silicon, which gives Nvidia the pricing power that comes with being the only supplier in a market where everyone needs what you sell.

And Nvidia charges accordingly. The Jetson Thor developer kit costs $5,499 as of August 2026, up 57% from the $3,499 Nvidia quoted when it first announced the chip. The price keeps climbing because VLA models keep getting bigger with each generation (better manipulation accuracy requires more parameters, which requires more compute to run) and because nobody else makes a competitive alternative.

Amortise that $5,499 over a three-year useful life at reasonable utilisation, and you are paying somewhere between 20 and 40 cents per robot-hour, just for the hardware, before you account for the software stack, the systems integration, or the engineering team keeping the models stable.

Now let’s compare that to what the human costs are:

What’s most interesting is that the cost gap between these two columns is growing in the wrong direction. The human side keeps getting cheaper because operators keep watching more robots at once. Kiwi needed one person per robot in 2017. Astro has one person per fifty in 2026.

That ratio will keep improving as monitoring tools get better and latency drops. At the same time, the silicon side keeps getting more expensive because VLA models demand more compute with each generation, and Nvidia has no competition pushing prices down. Five years ago, you could at least argue that compute costs would eventually fall below human labour costs. Today, that crossover point is further away than it has ever been.

When a company’s cheapest and most reliable input costs 3 cents per robot-hour, and the technology that would replace it costs ten times more and still cannot handle the hard cases, every rational dollar goes into making the cheap input even cheaper.

Now you can use the saved margins to build better monitoring dashboards, lower-latency headset connections, and faster pilot onboarding programs. No company would ever spend that money finishing the autonomy stack. A company called Adamo was recently launched with this exact logic. Adamo calls itself “the bridge to robot autonomy.” It sells managed teleoperation services to US robotics companies at $13 an hour.

Adamo’s own sales materials make the case that building an in-house teleoperation team costs more than outsourcing to Adamo, which means the company is actively selling the proposition that the human should stay in the loop and that you should pay Adamo to manage them. And every efficiency gain Adamo makes in pilot training or shift scheduling makes that spread more profitable and makes full autonomy less economically necessary.

And not just that: Telexistence, the company running robots in those 300-plus FamilyMart stores, adds another darker layer. Every time a Manila pilot takes manual control and corrects a robot’s grip, that movement gets logged as embodied teleoperation data. The company then feeds this directly to Physical Intelligence, a San Francisco AI lab building VLA models that are supposed to eventually make teleoperation unnecessary.

So the workers are doing three things at once: performing the physical labour that keeps the stores stocked, keeping unit economics attractive enough that investors keep funding the operation, and generating the training data that will theoretically make their own jobs obsolete.

These workers are building the tools that will eventually replace them. And because the economics are so favourable, the workers keep working, caught in a loop that benefits every participant in the chain except the person wearing the headset.

In February 2026, Senator Markey’s office wrote to seven autonomous vehicle companies, including Aurora, Tesla, Waymo, and Zoox, asking how often their remote operators intervene. Every company refused to answer. Before GM shut it down, Cruise had one remote assistant for roughly every fifteen to twenty vehicles, with human help triggered about every four to five miles.

The $12 Million Robot

So the teleoperator costs 3 cents an hour, and the chip that would replace them costs ten times more and still cannot do the job properly. And the companies making these robots have every incentive to keep the human in the loop. Given all this, you would expect the market to price this accordingly? Well, good luck!

Figure AI, the humanoid robotics startup, raised at a $39 billion valuation earlier this year. The company has shipped roughly 150 robots. And you would be surprised that the top five Western humanoid startups have a combined valuation of $73 billion. Together, they have deployed fewer than 400 robots. Which, if you do the math, works out to about $12 million in invested capital per robot that actually exists in the physical world.

For Figure’s valuation to make sense at even a 15x revenue multiple, the company would need to reach around $2.6 billion in annual recurring revenue. At a thousand dollars a month per robot, which is the standard Robot-as-a-Service price point the industry has converged on, that would REQUIRE ROUGHLY 215,000 robots deployed.

And even if you could manufacture and deploy 215,000 humanoids tomorrow, the unit economics still wouldn’t work. A humanoid robot costs at least $60,000 to build. Over a three-year useful life, that is $20,000 a year in hardware depreciation alone. Add $4,000 a year for maintenance and repairs (service every six months, shipping, technician time, lost revenue during downtime) and another $2,000 for remote supervision, and you land at $26,000 per robot per year in total operating costs.

On the value side, a robot replacing a $30-an-hour warehouse worker across six shifts a week produces about $75,000 in annual labour savings. That sounds like a great margin until you account for two more things. Robots currently work two to four times slower than humans at the same tasks. And real jobs are not single tasks; they are dozens of tasks, and a robot trained on picking will sit idle when the work shifts to inventory or packing.

Apply a conservative 3x productivity penalty and a 15% flexibility discount, and the $75,000 in value drops to about $21,000. Subtract the $26,000 in costs, and the robot is negative $5,000 a year.

Negative 18% ROI.

This is a technology that is genuinely impressive in demos, with a TAM slide in pitch decks that assumes total addressable penetration, for a valuation that prices in a future where every technical and economic obstacle has been solved simultaneously, and unit economics that are underwater today.

The narrative runs ahead of the fundamentals; the fundamentals take longer than anyone projected, and the correction, when it comes, is priced in all at once.

That’s all for today!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

A researcher is expected to use real, direct sources. A designer is expected to understand that “clean” does not mean submitting a blank page. A developer asked to build a working checkout is also expected not to make it take four minutes to load. Humans are surprisingly good at carrying all of this unstated context because we share a basic sense of what reasonable work looks like. Software never needed much of that because we usually gave it instructions narrow enough to execute literally.

You know what I am talking about if you have watched this bit from “The Orville”:

AI agents are beginning to work in a messier world, where the instructions do not spell out everything that counts as doing the job properly. Once money depends on that missing context, interpretation is also part of the transaction.

Frontrun: Be One Step Ahead

Before top investors back a young company, they leave a trail. Every top deal starts with an investor getting curious and following the company on their socials. It eventually culminates in a funding deal a couple of months later.

Whether you can sneak in on that information before the deal is struck and the hidden gem is out in public depends on your ability to track the trail.

Frontrun watches what 2,000+ leading investors, founders, and tech insiders follow. The moment they start crowding around the same company, Frontrun flags it and tells you who’s behind it and what they’re building. It even drafts your first message to the founder.

You can also get customised daily reports around the accounts you are tracking. When the accounts you follow start piling into the same company, you’ll see it on your feed and in your inbox by breakfast.

Start your 7-day free trial

If someone paid a freelancer $15 to research 20 companies and the brief is reasonably clear, what do they expect? Find the companies, explain what each one does, use independent sources and send the finished report by this date. When the work arrives on time and contains all 20 names, the person who did it considers the job finished.

The buyer sees it differently. Several citations point back to company blogs, and two descriptions appear to have been copied almost word for word.

With two people involved, the next step is that they argue, send screenshots, point back to the original brief and go back and forth. If they still cannot agree, somebody else may eventually have to decide if the work was good enough to get paid.

What happens if both sides are AI agents?

Money itself is not particularly difficult.

• An escrow contract can hold the $15 until the job is completed.

• A blockchain can show when the payment was deposited and when the work was submitted.

• Identity systems can help establish which agent performed the job.

• Signed records can preserve the instructions both sides agreed to.

The unresolved part is that somebody has to read the brief, inspect the work and decide whether the seller actually did the job.

Commerce has always had institutions for this problem, but we notice them when something goes wrong. When a PayPal buyer claims something never arrived or didn’t match the description at all, both sides get a chance to fix it themselves first. If they can’t, one of them can escalate it to a claim and ask PayPal to check the proof and decide who is right. PayPal says most of those decisions take around 14 days, though a few take 30 days or longer.

Larger commercial disagreements have their own machinery. US brokers have FINRA, the industry’s self-regulator, which runs its own arbitration forum. It received 2,597 new cases in 2025 and took an average of 13.4 months to close them. The American Arbitration Association, a non-profit organisation, says 580,000 cases were filed with it in 2025, including more than $29 billion in business-to-business claims and counterclaims. The International Chamber of Commerce received 881 new arbitration cases during the same year, while the total value of its active caseload reached $299 billion. These are obviously much bigger disputes than the ones agents are likely to have right now. But they show that once enough deals are happening, people start arguing often enough that a whole industry pops up around looking at the evidence and figuring out who should get paid.

AI agents are starting to raise that kind of problem on the other side of things, where transactions can be really small and happen so often that conventional dispute processes don’t make much sense.

Visa and Artemis studied two emerging payment protocols for machine-to-machine commerce and found that x402 processed about 109.6 million adjusted transactions worth roughly $15 million between its launch in May 2025 and April 2026. The average payment on the protocols Visa examined was a fraction of a cent. A human support team cannot sensibly spend half an hour investigating every disagreement over a payment worth pennies.

I went looking for who is solving this and found GenLayer, trying to turn this into infrastructure.

GenLayer is building a blockchain itself for decisions that ordinary smart contracts struggle with. Traditional smart contracts are good when every machine can do the same math and land on the same result. They can tell you if money arrived, or if the deadline’s gone. The issue shows up when it needs to read something and interpret it. Maybe it has to go through a research report, look at a webpage, understand a paragraph in a contract, or figure out if an image follows a creative brief. AI models never reply with the same answer. All of them can look at the same material and come back with different conclusions, so the assumption that every node has to get the exact same output falls apart.

GenLayer has an “Intelligent Contract.” Instead of requiring every validator to reproduce identical text, the contract can ask several validators to judge whether a proposed result is acceptable according to rules written into the contract. The validators can use large language models and public web data while doing this.

At first glance, this sounds like another missing layer in the agent stack. But some other systems already do pieces of that job. So GenLayer is not starting from zero. Its more specific role is deciding what the evidence means when both sides still disagree.

Google’s Agent Payments Protocol, or AP2, already creates cryptographically signed records of what an agent was authorised to buy, what the merchant offered and what payment was approved. Its specification explicitly says those mandates and receipts can later be combined as evidence in a dispute. But AP2 itself does not decide/judge who is right or who should get the money. The protocol preserves a trustworthy record of what happened and leaves the process for interpreting that record to somebody else.

ERC-8004, Ethereum’s standard for agent identity and reputation, also has a Validation Registry. An agent can send its work there to be checked by another system, and that result can then be recorded onchain as part of the agent’s history.

Blockchain can show if money moved, so payment disputes are sorted. When a dispute needs outside information to settle, an oracle goes and brings that information onchain. So there’s a clear missing piece when the evidence exists, someone still has to judge what it means, that’s where GenLayer comes in.

To see how much real work falls into each bucket, my agent on slack (whom I trust blindly because it’s us against the world) pulled a little over 300 live task listings on September 24. It included those from Freelancer.com, Fiverr, Upwork, Scale AI’s Outlier, Prolific and Daydreams TaskMarket, an on-chain marketplace where agents post and complete paid tasks. About 23% of the tasks were easy to check with a clear yes or no, said my agent. Another 45% depended mostly on personal judgment. The remaining 32% had evidence to look at, but still had to judge whether the work was actually good enough.

That is not a market-size estimate, and the classification itself involved judgment by an AI. But you get the idea. And every task involving AI doesn’t require an AI court.

A task like designing the most beautiful restaurant logo has several validators who may agree that one design looks better than another, but their agreement does not turn taste into an objective standard. Unless the buyer wrote a very detailed creative brief beforehand, the dispute is still rooted in personal preference. Similarly, a marketing campaign can deliver the promised engagement numbers while leaving a disagreement about how much of that activity came from genuine users. In each case, there is evidence to inspect, but it needs interpretation before payment is made.

Crypto has been experimenting with decentralised dispute resolution for years. These systems approach disagreement in very different ways.

UMA’s Optimistic Oracle starts by assuming a claim is fine unless somebody objects. The person proposing it posts a bond. If no one challenges it during the waiting period, the answer is accepted. If someone does challenge it, UMA token holders step in and vote. This keeps costs low because most claims pass without a full review, and the expensive voting process only happens when someone challenges one.

Kleros is a decentralised dispute-resolution system on Ethereum. Instead of a company employee or court deciding a dispute, Kleros randomly selects people who have staked its PNK token to act as jurors. If one side appeals, the next round brings in a larger group of jurors. That makes Kleros useful for cases where human judgment really matters, but it also means the process takes time and costs money. A three-juror round in the General Court is about 0.015 ETH before any appeal, so the model makes no sense when the dispute itself is only worth a few dollars.

GenLayer takes the judging job Kleros gives to human jurors and gives it to AI-assisted validators. One validator proposes an answer, the others check it, and if enough agree, the result moves toward final approval. If someone challenges it, a fresh group of validators reviews the case, with larger groups added if the dispute keeps going. Also, both UMA and Kleros are moving toward AI-assisted adjudication, so I would not call GenLayer “the AI alternative” to two permanently human systems.

If the rules are specific, validators have something concrete to check. If the contract only says the work should be “insightful” or “original,” several models agreeing with each other does not make that standard any clearer.

GenLayer handles this with what it calls the Equivalence Principle. It tells validators how similar their answers need to be for the network to treat them as the same decision. For some contracts, that might mean an exact match. For others, validators can use different wording or reasoning as long as they agree on the important outcome. GenLayer’s documentation repeatedly encourages developers to narrow these decisions.

That puts a lot of power in whoever writes the contract. They decide what evidence matters, what counts as success, and how much variation between answers is acceptable. GenLayer can judge against those rules, but it cannot fix a vague standard after the dispute starts.

A verdict only matters if it can actually change the transaction. In practice, that means the money or some other valuable outcome needs to be tied to the decision. A dispute can then end with funds being released, returned, or an agent’s reputation being updated.

For a small agent transaction, the goal may simply be making sure the money goes to the right side rather than a legally enforceable ruling.

Where would this actually be useful?

Agent marketplaces already combine delegated work with payments and increasingly with onchain identity, so that’s one of the most obvious places to look.

Daydreams TaskMarket, for example, runs on Base and allows humans or AI agents to post tasks backed by USDC. Its different task formats reveal where adjudication would and would not be useful. A benchmark task can ask workers to compete on a measurable metric such as accuracy or latency. If the task is based on a clear score, like one agent getting 96% and another getting 91%, the platform can simply check the numbers and pick the better result.

But with a bounty, several people might submit different kinds of work, and the requester has to decide which one is best. If there is no simple number or rule that tells you who did better, someone has to look at the work and decide.

When it comes to Software and security bounties, a task asking an agent to make a fixed test suite pass can be settled mechanically. A vulnerability bounty may produce a real disagreement about whether an exploit falls inside scope, whether its severity was described accurately or whether somebody had already reported the same underlying issue.

In performance marketing, counting 50,000 social interactions is easy enough. But are they genuine users or automated accounts? GenLayer already showcases Rally, a performance-marketing application that uses its validators to analyse social activity and assess the authenticity of engagement.

Now, a service-level agreement, or SLA, is basically a promise between a service provider and a customer about the minimum level of service they will provide. For example, a cloud company might promise 99.9% uptime, or support replies within two hours. If the provider fails to meet that promise, the customer may be entitled to a refund or a service credit.

If the agreement already specifies which logs, monitoring services, and exclusions count as evidence, an automated adjudicator has a much narrower thing to deal with.

Research is a good example of where GenLayer works well up to a point. Some parts of research are easy to check. Did the report answer all the questions? Are the sources real? Validators can inspect those things.

But something like “is this insight original enough?” is much harder. There may be no clear rule or evidence that settles it.

The work cannot be judged by a simple yes/no script, but there is still enough evidence and structure for different validators to make a reasonable decision. If the task is almost entirely about taste or vague quality, the system has much less to work with.

If you think using several AI judges automatically makes the decision reliable, no.

If all five validators use similar models, they might all make the same mistake. They could misunderstand the same instruction, trust the same bad source, or all be fooled by manipulated content. GenLayer tries to reduce this by having validators check decisions independently and by bringing in new validators if someone appeals.

The evidence can also be unreliable. A webpage might change, an API might give different answers at different times, or a document might contain instructions designed to confuse the AI reading it. So the system only works well when validators can properly inspect the same evidence. Sometimes there may simply be no clear answer. GenLayer can return an “Undetermined” result instead of forcing a yes or no.

GenLayer’s design leans on Condorcet’s Jury Theorem, the idea that a group of independent reasoners beats any single one. That only holds if the reasoners are truly independent and AI models learn from overlapping data. Kleros researchers tested something close to this on 99 real customer disputes from a Latin American crypto exchange. ChatGPT 5.5 sided with the customer in 3 cases and Claude Opus 4.7 in 14. When they reran the cases on the next Claude version, the platform’s win rate rose from 86 to 95%. Mixing models helps when their leanings are random. If a whole generation of models drifts the same way, a mixed panel drifts with it.

How big is the market now? The market is hard to size because only a fraction of agent transactions will ever become disputes. But existing systems show the demand for adjudication is already great.

Merchants handled about 261 million chargebacks in 2025, worth roughly $33 billion in dispute processing at $128 each. At the other end, GenLayer currently shows around 25,800 testnet decisions a day. Even at $1 per decision, that is only about $9.4 million a year. For comparison, eBay at its peak handled roughly 60 million disputes a year, or about 164,000 a day. GenLayer only becomes a large business if machine commerce eventually produces dispute volumes at that kind of scale.

For this to work, the contracts need clear rules, and the evidence needs to be reliable. The validators need to be independent enough that their agreement actually means something. Appeals also need to stay cheap.

Human commerce already has courts, arbitration and dispute teams for when two sides disagree about the same deal. If agents start doing more business with each other, they may need a faster and cheaper version of the same thing, because machines always need a cheap, fast version of everything.

Maybe it is a good thing we teach machines how to decide what is fair before we teach them to do everything else.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

For the first time in modern history, central banks around the world hold more gold than US Treasuries. Long-dated Treasuries have lost over 40% of their value since 2020, and the countries that used to absorb new issuance ( China, Japan, Europe) are now net sellers. The US government needs someone to keep buying its paper, and over the past couple of years it has gone from trying to shut down stablecoin issuers to writing laws that turn them into Treasury buyers.

A direct line runs from America’s debt problem to stablecoin regulation and the growth of perps, prediction markets, and on-chain stocks. I want to trace that line today, because at the end of it, the US government is funding its borrowing through a global crypto roadhouse.

Let’s dig in!

TOKEN2049: The Largest Crypto State

Over 25,000 attendees, 300 speakers, 500 exhibitors and 1,000 side events will come together at TOKEN2049 to rewrite the future of finance. And you are invited!

TOKEN2049 Singapore is just over three weeks away. The global crypto industry meets at Marina Bay Sands, 7-8 October, for the world’s largest crypto event.

Book by Wednesday, September 23, to lock in the current rate of $539 ($599) and get the exclusive 10% discount before the price goes up Today!

Book before the price goes up!

Patient Money Out, Hot Money In

If you hold USDC right now, your dollar is sitting in a Treasury bill. Under the GENIUS Act, every stablecoin issuer in the US has to back its tokens with short-term government debt. Circle and Tether together hold about $149 billion of Treasuries, and we holders don’t earn anything on that because GENIUS specifically bans issuers from paying interest on stablecoins. So you have a $149 billion pool of money that federal law requires to sit in T-bills and has no legal way to chase a better rate.

The $149 billion might feel like pennies compared to the entire $32 trillion debt market, but that’s not what matters here. For example, Money market funds already hold trillions of dollars in T-bills. And the difference is they can sell those bills whenever they want. If rates shift or spreads open up elsewhere, the fund can move its money because it is a buyer by choice, and it stays a buyer only as long as bills are the best option available. But a GENIUS-regulated stablecoin issuer cannot do that. It must hold reserves against every token in circulation, no matter the yield, for as long as anyone holds the token. And that’s why it gets locked in by the federal statute.

Washington used to have locked-in buyers on the long end of the curve too. Countries that sold goods to the US got paid in dollars, and those dollars needed somewhere to go. The only market deep and liquid enough to absorb that volume was US Treasuries, so Japan, China, and every other export economy ended up buying American bonds as a side effect of selling Americans cars, electronics, and clothing.

But not anymore, because foreign official holdings have dropped to 12.8% of outstanding Treasuries, the lowest since 1993.

What took their place is mostly hedge funds running leveraged bets out of the Cayman Islands. They buy a Treasury bond, sell the matching futures contract, and capture the tiny spread between the two prices. The spread is usually only a few basis points, so they lever it 50 to 100 times to make real money.

Close to $2 trillion of this trade is sitting in the market right now. And we already know how these buyers behave under stress: in March 2020, due to the COVID crash, every fund running this trade tried to exit at the same time, which led to the Treasury market freezing, and the Fed had to spend $1.6 trillion in three weeks buying Treasuries because nobody else would. Sellers couldn’t find buyers at any price, so the central bank became the buyer of last resort for its own government’s debt.

So Treasury Secretary Scott Bessent has a problem. The buyers who held American debt for decades are gone, and the buyers who replaced them are the kind that panic and run all at once. So he changed the strategy to shift more of the government’s borrowing away from long-term bonds, where nobody reliable wants to lend anymore, and toward short-term bills.

He doubled the Treasury’s buyback program on the long end to $4 billion per operation, and is paying for those buybacks by issuing new short-term bills. In 2024, before he got the job, Bessent co-wrote a paper accusing Janet Yellen of doing this exact same thing and called it stealth monetary easing that hid how bad the borrowing situation really was. Now he is running the same play at twice the scale, which tells you the situation got worse since he wrote that paper.

But the strategy only works if someone keeps buying those bills reliably. And that is where GENIUS comes in, because the stablecoin issuers regulated under that law are now required to buy exactly those bills that Bessent is flooding the market with.

And quite frankly, stablecoin holders are a much better buyer for Washington than the central banks ever were. Because China or any other country can call one meeting and decide to dump its Treasury holdings. But millions of stablecoin holders scattered across different blockchains have no way to coordinate a sale, even if every single one of them wanted out at the same time. That makes stablecoin holders the new captive buyer, and a more locked-in one than any central bank ever.

Another thing is, not all demand is equal for T-bills, since it splits into domestic and foreign money. When an American moves money from a bank account into a stablecoin, it doesn’t really create new demand for T-bills, because the bank was probably already holding bills with that money anyway. The demand just shifts from one holder to another.

But when a trader in Istanbul swaps lira into USDT to trade perps on Hyperliquid, that is net new money flowing into American government debt from outside the country, and no existing T-bill demand is being displaced by that transaction. About 60% of stablecoin demand comes from abroad, which is why the Treasury’s own borrowing committee projects the stablecoin market at $2 trillion by 2028 and counts on roughly $900 billion of new bill demand from it.

And we can already see the effect in the data. Every time $3.5 billion of new money flows into stablecoins, it immediately pushes 3-month T-bill yields down about 0.71 basis points, and that drops further to around 4 basis points over the next 10 days. The more stablecoins in circulation, the bigger that effect gets.

But then the catch is that GENIUS bans yield on stablecoins. Every other way of holding dollars pays you something for it. A savings account pays interest. A money market fund pays interest. The T-bill sitting underneath your stablecoin pays interest too, just not to you, because the issuer keeps that yield. So why would anyone still choose to hold a dollar token?

Because the token gets you into the trading venues. You hold USDC because Hyperliquid lets you trade 100x-leveraged crude oil perps with it at 2 AM on a Sunday. The trade-off for the yield is access to the venue. But all that T-bill demand exists only because people want to trade. If the trading dries up, people redeem their stablecoins, and the issuers have to sell the T-bills backing them. We already know that is exactly what happens, because when crypto trading collapsed after the October 2025 crash, USDC supply dropped by nearly $6 billion.

Crypto is Washington’s Las Vegas

Two years ago, the US government was actively trying to destroy the stablecoin industry. Under the Biden administration, federal regulators ran what the crypto industry now calls Operation Choke Point 2.0, where the FDIC and the OCC pressured banks to close accounts belonging to crypto companies without passing any new law or holding any public hearing; they just called the banks and told them crypto clients were too risky. Signature Bank, Silvergate, and Silicon Valley Bank all failed in 2023, and each had been a major on-ramp for stablecoin issuers. The federal government spent two years trying to cut stablecoins off from the banking system entirely.

Then Trump won, hired David Sacks as his crypto czar, and the entire posture reversed in weeks. Sacks went on TV and said the GENIUS Act “could create trillions of dollars of demand for our Treasuries practically overnight.” You don’t go from prosecuting an industry to writing it into your borrowing strategy unless the borrowing situation got bad enough to force your hand. And the fact that Sacks could say that on camera without a single person in Washington pushing back means the entire political establishment had already agreed that stablecoins were going to be the next source of Treasury demand.

And then the CFTC pushed it further. In May 2026, Chairman Michael Selig approved the first regulated perpetual futures contracts in US history, saying that “perpetual trading activity has predictably occurred offshore” and that American firms were “competitively disadvantaged.” The regulator, whose entire job is to protect markets from excessive speculation, approved 100x-leveraged offshore perps and decided the problem was that American firms were losing fees and deposits to foreign venues.

As funny as it may sound, it proves that the US derivatives regulator is now actively trying to grow speculative trading volume because growing that volume is how you grow the government’s newest source of debt financing.

And the speculation has already outgrown crypto. trade.xyz runs on Hyperliquid and was responsible for 55% of all Hyperliquid volume in August 2026. Its top markets were crude oil, silver, Brent crude, a Nasdaq-100 proxy, the S&P 500, and AI chip stocks like SK Hynix and Micron.

Every dollar deposited to trade crude oil or SK Hynix perps on Hyperliquid is a dollar that Circle has to back with a Treasury bill, which means the crypto derivatives market has turned into a funnel that takes money from people in Istanbul, Lagos or São Paulo who want to bet on oil and semiconductors and pushes it directly into American government debt.

But someone has to lose for this machine to keep running. When a trader loses a leveraged position, stablecoins don’t leave the system. They move from the loser to the winner and stay on-chain. T-bill demand only falls when someone actually redeems their stablecoins and takes dollars out. So the system does not actually need trading volume in the abstract; it needs a constant flow of new depositors willing to put fresh money in and lose it.

That’s very similar to how a casino works, because the house does not care about total bets placed; it cares about new players walking through the door.

Out of about $35 trillion in total stablecoin transaction volume in 2025, only about $390 billion were actual payments for goods and services. That’s roughly 1%. Circle’s own real-world payments network did $5.7 billion in Q4 2025, against trillions in on-chain USDC volume. So when someone tells you stablecoins are a payments innovation, they are technically correct in the same way that Las Vegas is technically a city with restaurants, but the restaurants are not why anyone goes to Las Vegas.

So the US government is funding its debt by keeping a roadhouse open. And the weirdest part is that both parties agree. Republicans want it because they see deregulation and dollar dominance. Democrats went along because 15 of them voted for cloture on GENIUS, and nobody wants to be the one who voted against something that kept Treasury yields down.

The crypto lobby spent over $130 million on the 2024 election cycle, and that money bought bipartisan agreement to keep the stablecoin-to-T-bill pipeline running, because both parties need the debt to keep getting funded and neither wants to explain to voters why yields spiked.

Maybe this works for decades. Maybe stablecoin supply hits $2 trillion by 2028, as the Treasury’s own committee projects, and $900 billion of new bill demand shows up right when Washington needs it most. Or maybe crypto has another October 2025 where the trading collapses, USDC sheds another $6 billion, and Washington discovers that its strategy is nothing but an empty suit.

Until Next time!
Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

I usually think twice before reading AI books now, and even more before recommending one, because they can become dated very fast. A book written five years ago can feel like it is describing a completely different industry.

Still, this one had me hooked for a week because it gave me a way to think about AI without trying to predict every version of what comes next, despite the name.

This one was published before ChatGPT, before Claude, before people were simply asking models to write code, operate browsers and call other software on their behalf. The AI examples in the book include fraud detection, recommendation systems, self-driving cars, translation, and inventory management. Fair enough, all of them are still relevant. But then, Alexa is still impressive enough to appear near the beginning of the book.

These days I have been looking at a rather different version of AI. There are agents with wallets, agents buying API calls, agents asking for permission to access Google Drive, and entire companies being built around controlling what these systems are allowed to do. Every week, another product claims to have solved some part of the journey from giving an AI a task to letting it complete that task with no humans.

That made me curious about what an economics book written before all of this would make of the current moment.

Ajay Agrawal, Joshua Gans and Avi Goldfarb are economists rather than computer scientists. Probably the reason why they don’t do much of trying to define intelligence or asking when machines will become conscious, no scary stuff.

The authors argue that the internet made search, communication and distribution much cheaper. Computers made arithmetic cheap. Artificial light became so cheap that we stopped thinking about whether turning on a lamp was worth the money. When the price of something useful falls far enough, we start using much more of it, including in places where we had never previously thought of using it. Their argument is that machine learning does the same thing to prediction. It is a wonderfully casual way to explain AI, and I mean that as a compliment.

Sharpen Your Investment Research with Amsflow

Wall Street and crypto used to run on separate screens, logins, and research stacks. That gap is closing everywhere else. Why not yours too?

Enter Amsflow - an all-in-one intelligence platform for investment research, covering every major global equity market, commodities, and deep crypto analysis, all in one place. Sentiment reading from Wall Street to the EU, India, Japan and China - down to gold and crypto.

Amsflow lets you run a Screener across 70,000+ securities with 550+ metrics. Track your Portfolio with real factor attribution, not just a P&L line. Watch entire industries rotate over a decade of history with Rotation, scrubbing through time instead of squinting at a static chart. Built sleek and clear, for anyone curious enough to look past the ticker.

You’ve been among Token Dispatch’s most loyal readers, so here’s an offer just for you - 20% off your Amsflow subscription, every month, for six months.

Use code ‘TOKENDISPATCH20’ at checkout for 20% off.

Sign Up with 20% Off

For the last few years, it has become difficult to talk about artificial intelligence without eventually arriving at civilisation, or human replacement. Prediction Machines drag the conversation back towards prices.

What do we consume more of because it became cheaper? What becomes more valuable around it?

With that, a surprising amount of the AI industry becomes easier to look at.

Prediction is the process of taking information you have and using it to generate information you do not have yet. Information you do not have yet does not even have to be about the future. Detecting whether a transaction happening right now is fraudulent counts as prediction. Translating an English sentence into Japanese counts too, because the machine is using one set of information to produce another.

Considering the way things are going these days, I am not sure that prediction can comfortably hold everything we now put inside the word AI. A language model does predict the next token. An image model also generates an image by estimating what visual information should come next or fit together. So if you stretch the word “prediction” far enough, you can still put most modern AI inside that definition.

The problem is that this starts becoming a little too broad to be useful. Today, an AI system might research something, make a plan, choose a tool, write code, spend money, check the result and try again. All of those actions may ultimately be powered by prediction, but calling the whole thing “prediction” leaves out a lot of it. So the point here is that prediction may still explain the mechanism underneath AI, but it may no longer fully explain the economic or practical thing we are interacting with.

Knowing what is likely to happen and telling you what you should do about it are two different things. The authors talk about rain and umbrella to explain that. A weather forecast gives you the probability that it will rain. You still have to decide whether to carry an umbrella. That depends on how much you dislike getting wet, how much you dislike carrying an umbrella and what you are doing that day.

The prediction can improve enormously while the preference remains yours. The book calls that second part judgment. At this point, I feel like I am teaching you “how to be a human.” Anyway, the book says that as prediction becomes cheaper and more widely available, judgment can become more valuable because more decisions can now be made using good predictions.

An agent can find flights, compare hotels and work out an itinerary, but once it has to actually book something, another set of questions appears. Now somebody has to decide whether the agent can spend ₹5,000 or ₹50,000, and whether it can buy a non-refundable ticket.

We are getting better at solving the “can the agent figure this out?” problem. Now we have to decide what it is allowed to do once it does.

I wrote about a similar argument recently through Aident, a company building a permission layer for AI agents. The product works between the agent and the software it wants to use, handling things like credentials, approvals, spending limits and an independent record of what the agent actually did. In my test, I let an agent use outside tools to research products, generate images and audio, publish posts and deploy a small website. It searched for tools as it went, checked their cost, switched when one failed, and kept moving.

Spending approval is one thing. Action approval is another. An agent might be allowed to spend a small amount on research, for example, while still needing permission before sending an email or changing a customer record.

That made me think differently about what “judgment” looks like once AI starts acting on its own. In the book, judgment is about assigning value to possible outcomes and deciding what matters. In an agent system, some of that judgment then has to be translated into rules the agent can actually follow, like how much it can spend, which accounts or tools it can access, and so on.

This is also why I think the permission layer around agents is more interesting than it appears now. It is really where companies start turning their own judgment into something software can follow.

In the Aident piece, I described that layer as an enforceable job description for an agent, defining what it can see, spend, alter, publish and escalate.

Read: Who Gets to Say Yes?

Crypto already has systems where software can hold value, move money and execute rules automatically. AI brings in systems that can decide when to use them. But have we built enough judgment around AI now before things turn into an irreversible action…Especially now, when some of the people building the most powerful AI systems are themselves asking whether things are moving too fast.

A smart contract can execute automatically, but first it needs some way to know what is happening outside the blockchain. An oracle brings in information such as an asset price or the result of an event. The contract then has rules for what should happen with that information. If collateral falls below a certain level, liquidate it. If a condition is met, release the payment. The oracle supplies the information, while the judgment about what that information should trigger is written into the system beforehand.

AI agents make that arrangement more complicated. The software is no longer only waiting for a number and following a rule that somebody already wrote. It can interpret information, choose between possible actions and decide which tool to use next. Crypto has been figuring out how to give software reliable information from the outside world. AI is pushing us towards the next problem of what happens when the software can also decide what to do with that information?

When something important becomes cheaper, the things needed around it can become more valuable. The authors call these complements. In the case of AI, those complements include data, judgment and action.

A bad transaction, liquidation or payment is harder to undo. So as AI gets better at producing decisions, more value may move towards the systems that check the information, define the rules and control what happens next.

The book also argues that asking whether AI will replace a job can be the wrong unit of analysis. Jobs contain many tasks, and technology may remove some of them while making other parts more important.

People who had previously spent large amounts of time manually calculating figures did not necessarily become useless once spreadsheets arrived. In many cases, their knowledge became more useful because they could spend less time performing arithmetic and more time asking better questions of the numbers.

The authors expect something similar with AI. Workflows get broken into pieces, machines take over some tasks, and the job is then assembled again around whatever remains.

Blockchains made it easier to create and move assets, but that made things like custody, liquidity, identity and compliance more important. Stablecoins made moving dollars easier, so more businesses appeared around issuing them and controlling how they’re used.

Read: Dollars Are a Technology

I am not even sure its central definition of AI would survive unchanged if the authors were writing the book today.

I still came away from it with a better way of looking at the industry.

The book wants you to stop staring at the technology itself and look at the economic system around it. That is one of the most useful habits I might have got from this book, and it felt like a useful antidote to how AI is usually discussed.

When I look at a new AI product now, I find myself trying to identify the thing whose price is actually falling. Then I look around it and see who is building to fix the gap that is exposed. If software can make thousands of decisions without waiting for us, somebody still has to decide what those machines are allowed to optimise for.

Crypto makes those decisions financially real because an agent can move money, interact with smart contracts and trigger transactions directly. That makes trusted data, permissions and spending controls much more important once AI starts acting on its own. Another reason why these two systems fit together almost too neatly. They are lobsters!

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

Before the 1929 crash, companies could sell shares with limited disclosure, traders could organise pools to manipulate prices, and stock exchanges were largely expected to police themselves. The crash made the cost of this arrangement impossible to ignore. Congress first passed the Securities Act of 1933, which required companies selling securities to give investors meaningful information.

The Securities Exchange Act followed in 1934 and created the SEC to oversee exchanges, brokers and the secondary market. The government was establishing a bargain that the claim had to be real, and the market moving that claim between people had to follow rules.

Nothing fundamental changed when certificates went digital, and floors became server rooms. Issuers still made the shares, transfer agents recorded who held them, exchanges matched trades, and market makers still supplied the other side so prices kept moving.

A blockchain can bring these jobs much closer together. The share, the ownership record, the payment and the trading pool can all operate through connected smart contracts.

The SEC’s obsession for much of the past decade has been asking when a crypto token should be treated as a security. The Clarity Act did not make it through Congress —RIP — but the SEC has now created a smaller, temporary route through its Innovation Exemption. This one gives companies five years to find out. A qualifying Tokenised Securities Venue (TSV) can facilitate onchain trading of tokenised US stocks through permissioned automated market makers and liquidity pools.

So if a token is already a genuine stock, can it trade through a crypto-style pool that only approved investors are allowed to use?

• The SEC’s answer is yes, but the old protections must remain.

• The token must carry normal shareholder rights.

• The company behind the stock must be informed and can object.

• Only approved wallets can trade.

• If the stock stops trading on Nasdaq or the NYSE, its tokenised version must also stop.

• The smart contracts must be public so regulators and investors can inspect what happened.

Creating a real stock and approving the pool only gets the market started. Someone must still provide the shares and cash, keep the price accurate and carry the risk. That is where we are going today.

TOKEN2049: The Largest Crypto Stage

Over 25,000 attendees, 300 speakers, 500 exhibitors and 1,000 side events will come together at TOKEN2049 to rewrite the future of finance. And you are invited!

TOKEN2049 Singapore is just over three weeks away. The global crypto industry meets at Marina Bay Sands, 7-8 October, for the world’s largest crypto event.

Book by Wednesday, September 23, to lock in the current rate of $539 ($599) and get the exclusive 10% discount before the price goes up next week.

Book before the price goes up!

A normal stock exchange tries to match someone who wants to sell with someone willing to buy. An automated market maker, or AMM, gives both of them a pool to trade against.

The seller needs the pool to contain enough cash. The buyer needs enough shares. Whoever supplied those assets earns part of the trading fees.

This sounds fairly easy, but suppose Apple trades at $200 on Nasdaq while its tokenised share costs $198 in the pool. A trader can buy the cheaper token, hedge it against the traditional market and keep the difference. This pulls the prices closer, but the liquidity provider’s pool sold the stock too cheaply.

Crypto calls one part of this impermanent loss. When a stock falls sharply, the pool can end up holding more of the falling stock and less cash. When it rises quickly, the pool may sell the stock too early.

Traditional market makers handle this by widening the difference between their buying and selling prices, trading less, hedging elsewhere or withdrawing. A simple AMM follows its programmed rules. This is uncomfortable with listed stocks, which can jump after earnings, a lawsuit or a post from a CEO who should perhaps have gone to sleep. A 24-hour pool remains awake.

The Innovation Exemption creates room for these pools to operate, and it offers conditional relief to certain firms supplying their own capital. It cannot guarantee that the trading fees will be greater than the losses from stale prices, adverse trades and badly balanced inventory.

Once we see the market this way, tokenised-stock liquidity separates into three problems.

1. The venue must check who is allowed to trade. A wallet address alone cannot show who owns it, whether that person has passed identity checks or whether they are legally allowed to hold the stock.

2. The pool needs enough shares and cash. If a pool holds only a few shares, even a small trade can empty it and push its price away from Nasdaq.

3. The liquidity provider needs enough capital to keep the pool running. It must fund the shares and cash, protect itself when prices move and earn enough from fees or spreads to make the risk worthwhile.

I tracked who gets to fix these layers.

Uniswap can decide who enters the pool

Uniswap v4’s Permissioned Pools deal directly with the first problem.

Uniswap v4 is normally permissionless. Anyone can create a pool, swap through it or provide liquidity. A regulated security cannot behave that way because its token contract may be required to reject transfers to wallets that have not passed the issuer’s checks.

Uniswap v4 keeps balances inside one PoolManager contract and updates virtual balances as people trade. If a restricted stock only checks wallets when a real token transfer occurs, it cannot see everything happening inside the pool. An unapproved person could receive exposure without triggering the stock token’s ordinary transfer check.

Permissioned Pools solve this by checking the wallet before the trade itself is allowed.

The real stock is held inside a separate smart contract called the Permissions Adapter, which the issuer has approved. Uniswap then uses a virtual version of the stock to track trades within the pool. Before anyone can buy, sell or add shares and cash, the adapter asks the issuer’s compliance system whether that wallet has permission.

It does not matter whether the issuer uses Securitize’s DS Protocol, Tokeny’s ERC-3643 or its own investor registry. Uniswap only needs a yes-or-no answer from that system.

The permission to trade is also different from the permission to provide liquidity. Someone may be allowed to buy a tokenised stock without being allowed to supply shares and cash as a market maker. Uniswap checks the trader before every swap and checks the liquidity provider before accepting their assets. Permission can also be withdrawn later. If a liquidity provider is no longer eligible, the issuer can close the position. The LP receives the unrestricted asset, such as USDC, while the tokenised stock is moved to an address chosen by the issuer. The stock does not forget that it is regulated merely because it entered DeFi.

The system is already live and Uniswap says Superstate, Securitize and Dowgo are building with it. None has thereby been approved as a TSV. The SEC published no list of winners. Uniswap has built technology resembling the permissioned AMM the order now makes possible.

The stock still has to come from somewhere

A token can carry Apple’s name and follow Apple’s price without making its holder an Apple shareholder.

Robinhood’s Stock Tokens, for example, are debt securities issued in Jersey. They provide economic exposure to the underlying stocks, but Robinhood states that holders receive no legal or beneficial ownership of those shares. The SEC’s new exemption requires the token to carry the same rights as the traditional stock. For the moment, we will step around Robinhood’s awkwardness and see who owns that layer currently.

Superstate’s Opening Bell helps public companies make their actual shares available on Ethereum and Solana. Galaxy Digital became its first major example in September 2025, when it allowed shareholders to move its Class A common stock onto Solana. These tokens were Galaxy’s SEC-registered shares, carrying the same rights as the traditional stock. When tokens moved between verified investors, Superstate updated Galaxy’s official shareholder list in real time.

As an SEC-registered transfer agent, Superstate maintains the shareholder registry across traditional book-entry shares, tokenised holdings and supported DeFi applications. When a company already has a transfer agent, Superstate works alongside it and builds a bridge between the existing record and the blockchain. Superstate also offers a Direct Issuance Program. This allows a listed company to sell newly issued shares directly to eligible investors using real-time market prices. The investor pays in stablecoins and receives the actual shares as tokens in an approved wallet.

Securitize covers similar ground. It issues and administers regulated digital securities, maintains investor records and enforces the rules governing who can hold or transfer them. Uniswap’s pools can query Securitize’s DS Protocol when deciding whether a wallet may trade or provide liquidity.

It is also the tokenisation and transfer-agent platform behind BlackRock’s BUIDL fund. BUIDL became available for onchain trading through UniswapX in February 2026, although this was not an AMM pool. It used a request-for-quote system in which approved investors could receive prices from whitelisted firms, including Flowdesk, Tokka Labs and Wintermute. Securitize Markets facilitated the transactions, and every participant had to be approved by Securitize. But look at how the two layers could work together.

Securitize controlled who could legally hold and trade the security. UniswapX supplied the technology for finding a quote and settling the transaction onchain. These companies are not necessarily Uniswap competitors. A transfer agent records who owns the share and whether a transfer is valid. Uniswap handles the exchange between approved participants.

SEC requires a Tokenised Securities Venue to verify that each token carries the same rights as the equivalent traditional share. If an unaffiliated third party tokenises a company’s stock, the venue must notify the original issuer and give it an opportunity to object. If the stock stops trading on its primary exchange, its tokenised version must stop too.

The SEC order also limits the number of symbols and trading volume available to each venue during the exemption. Somebody must still supply enough shares and cash to make those permitted markets useful.

How do you keep the price from becoming ridiculous?

A passive AMM, such as a standard Uniswap pool, is always open for trading. Its price depends on how many shares and how much cash are sitting in the pool.

Concentrated liquidity allows the liquidity provider to place most of that money close to the stock’s current price. This helps the pool offer better prices without requiring as much capital.

The bigger problem is the clock. A simple AMM sets its price from its own pool, while arbitrage traders keep it close to the outside market by buying wherever the asset is cheaper and selling wherever it is more expensive. Let’s say Nvidia finishes Friday at $200. On Saturday, the company announces something terrible. People now expect the stock to fall, but there is no fresh Nasdaq price to tell the pool how far. If the pool is still buying near Friday’s price, someone who sees the news can sell into it before the price adjusts. The people who put money into the pool are left holding shares that may be worth less.

RedStone’s COO, Marcin Kaźmierczak, pointed out that regular NYSE and Nasdaq trading takes up only 32.5 hours of a normal 168-hour week. Trading before and after those sessions gives us more prices to work with, but weekends still leave a gap.

Oracles, which bring outside data into blockchain applications, help narrow that issue. Chainlink’s 24/5 US equity streams cover regular, pre-market, post-market and overnight sessions. They report buying and selling prices, market status and information about how old the price is. When the traditional market closes for weekends or holidays, however, those feeds carry stale values. An application can use prices from tokenised markets instead, but those are no longer a live reference to the underlying stock market.

This leaves another decision for the trading system. Knowing the latest available price is useful, but how much inventory should we offer at that price, and how cautious should we be when the information gets old?

Byreal, a Solana exchange incubated by Bybit, approaches this by offering several possible routes for each trade. It compares concentrated liquidity pools with quotes from professional market makers through a request for quote, or RFQ. Or, let’s just say it asks a trading firm what price it will offer for the requested amount.

A professional firm may offer a better quote after considering its own inventory and prices elsewhere. Byreal can draw on its own pools, external AMMs including Raydium, Orca and Meteora, and market-maker quotes. The useful innovation is the ability to compare these sources rather than expect one pool to handle every order equally well.

BisonFi, from Jump Crypto, is a proprietary AMM, or PropAMM, with an offchain pricing engine that watches other markets and sends updated prices to an onchain trading program. The assets and settlement remain onchain, but the price can change before someone trades against an outdated quote.

Its program also responds to inventory and the freshness of its information. If it holds too much of one asset, it can adjust prices to encourage trades that restore the balance. If its price update is old, it can widen the gap between its buying and selling prices or offer less inventory. These are ways of continuing to make a market while limiting how much risk the firm accepts.

UniswapX addresses the same broader pricing problem through competition. Firms compete to fill a trader’s order using different liquidity sources. Separate from Uniswap’s permissioned pools, you can see how execution can improve when a trader is not confined to a single pool’s quote.

What all of these offer is a better way to trade under that uncertainty. For builders working on genuine tokenised stocks, the opportunity is to combine reliable market data, competitive quotes and controls that respond when information becomes scarce. Existing crypto-market capabilities provide a useful foundation, although applying them under the SEC exemption remains a separate task.

Suppose everything else is ready. The token represents a real share, the pool checks who can trade, and a market maker knows how to keep its prices competitive. That firm still needs enough shares to sell and enough cash to buy from everyone else.

A firm making markets in ten tokenised stocks may need shares and stablecoins across ten pools. Without financing, every dollar supporting those quotes comes from its own balance sheet. Traditional finance already has businesses called prime brokers that lend to trading firms. When deciding how much security to require for a loan, they can consider several investments together. For example, a firm might own shares while also holding a separate trade designed to gain value if those shares fall. Looking at both gives the lender a better picture of the risk than examining either alone.

Arkis brings this very portfolio margin into digital markets. Lenders supply money, while Arkis assesses a borrowing firm’s supported investments together and monitors the assets backing its debt. If those assets lose too much value, the system can sell positions to help repay the loan.

Arkis is adjacent to tokenised-stock liquidity, not a stock AMM. But it could help finance the firms putting shares and cash into trading pools. A market maker that can borrow against its combined investments may be able to support more trading with the money it already has.

There is still a large jump between that possibility and the exemption. Arkis has not established that these US stocks will be accepted as collateral, that TSV positions can enter its margin system or that its lending structure meets the relevant securities requirements.

The SEC has also offered a temporary conditional dealer exemption to liquidity providers that use proprietary capital in TSV pools, quote prices or commit capital. That removes one regulatory obstacle.

A single stock trade can support several businesses. One keeps the ownership records, another runs the trading pool, and others supply the shares, set prices or lend the money.

They can earn administration fees, trading income or interest, so they are not all competing for the same payment. Who earns the most will depend partly on where people choose to trade. Buyers and sellers want good prices and enough shares and cash available to complete their orders. But the firms supplying those assets need enough customers to make the effort worthwhile. Each side is waiting for the other.

When prices move sharply, and someone wants to sell a large holding, can the pool complete that sale at a reasonable price, while leaving the firms supplying its shares and cash willing to keep doing so? That is when we will know whether tokenisation has created a market people can rely on.

We never know. Robinhood is probably naming a feature as we speak.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

We subscribe to a product for a month when we may only need one small thing from it. Monid was interesting because it breaks some of those products back into the individual things they can do. You keep one balance, call whichever tool you need and pay for that use.

That was: Software by the Sip

I wanted to see how much the difference counts when the software is expensive enough that you would normally think carefully before subscribing. So I went back to video generation, which is very good at making credits disappear.

Let’s see how it turned out in this piece.

TOKEN2049: The Largest Crypto Stage

Over 25,000 attendees, 300 speakers, 500 exhibitors and 1,000 side events will come together at TOKEN2049 to rewrite the future of finance. And you are invited!

TOKEN2049 Singapore is just over three weeks away. The global crypto industry meets at Marina Bay Sands, 7-8 October, for the world’s largest crypto event.

Book by Wednesday, September 23, to lock in the current rate of $539 ($599) and get the exclusive 10% discount before the price goes up next week.

Book before the price goes up!

I asked Monid to make me a 20-second video of a yellow duck wearing headphones and walking through a crowded subway. The style was supposed to look like an oil-pastel animation drawn frame by frame on paper. Monid found ByteDance’s Seedance 2.0. Seedance itself has a 15-second maximum, so that is what it made.

The finished video was 15 seconds, 720p, with audio. The exact Monid receipt was $2.2743 for 324,900 tokens.

The duck was pretty good, in fact (although I have no idea what the people behind are up to, they seem pretty aggressive. It stayed yellow, kept the headphones and moved through the subway as requested. The animation was smoother than the hand-drawn movement I had asked for, not that happy about it for an unexplainable reason, and a “No Smoking” sign sneaked into a prompt that explicitly said no text, so we are not talking about cinema history here.

What would the same habit cost if I made ten of these videos every month through Higgsfield?

Ten Monid generations at the price I paid would cost $22.74. There is no monthly plan involved, so $22.74 of usage means $22.74 is deducted from the balance.

At the time of my test, the Higgsfield plan that could cover the same ten 15-second Seedance generations cost $47 per month. Those ten videos consumed roughly 660 to 675 credits, worth around $26 at that plan’s credit rate. The plan came with 1,200 credits, leaving roughly 540 credits after the ten videos. Higgsfield’s subscription credits expire at the end of the credit cycle rather than rolling over. Higgsfield’s current help centre still says subscription credits reset each billing cycle and unused credits expire.

These were Higgsfield’s prices when I ran the experiment. The exact numbers may change later.

If I made the same ten videos through Higgsfield, they would consume about $26 worth of credits, but I would need a larger subscription to cover them. That can be completely reasonable if I also use its other video models, image tools, camera controls, identity features and editing workflows. In this case, its subscription pays for a larger creative workspace rather than Seedance alone.

Seedance sits inside a larger creative suite with features such as Soul ID, Cinema Studio and Marketing Studio. If all I want is ten Seedance clips, however, much of that product disappears from my calculation.

I only want the capability.

Like we discussed last week, Monid describes itself as the “OpenRouter for agent tools.”

There are now more than 1,700 tools from over 55 providers listed on the site, across 71 categories. Its dedicated OpenRouter page says a request can return several possible sources with their live prices, and the system can route towards the cheapest one capable of doing the job.

Monid Founder Shengkun Ye said in September that the platform had crossed four million transactions and raised a $2.1 million pre-seed round. A few months before that, Monid was being described as a wallet with just over 200 endpoints. The catalogue is already several times bigger now.

OpenRouter provides the obvious blueprint because AI models have become unusually easy to compare. You can send similar work to several models, record price and latency, watch for failures and route the next request accordingly. The customer usually cares more about the output, price and speed than about who actually ran the model.

Tools are harder because a TikTok scraper and a video generator do not have the same unit of work. One provider charges by result, another by call, Seedance charges according to generated video tokens, and browser automation may charge according to time. This makes it difficult for Monid to compare tools on price alone or decide which one offers the best value. Outputs are different too, as a data tool might return a table of prices, while a video model gives you an MP4 file.

Monid is trying to put all of these very different tools behind the same buying experience. It wants more software companies to sell through this system.

An API developer can list a tool on Monid instead of finding every customer themselves. Monid brings the users, keeps track of how many times the tool is used, collects the payment and gives the provider its share.

That is when it could become a software marketplace instead of a collection of APIs. Which brings us to wholesale software.

Wholesale is a word we usually associate with physical goods.

When a factory makes thousands of products, and a distributor buys them in bulk, and those products eventually reach customers through shops and marketplaces. Wholesale markets have traditionally existed because producers and buyers do not always want to deal with each other one-on-one.

A wholesaler would buy goods from many producers, sort and standardise them, hold inventory, then break those large quantities into smaller amounts that retailers could use. The product itself usually did not change very much. How it was packaged, priced and distributed changed over the years. And it makes some software unusually suited to the same model.

An API already turns software into a defined unit of work. When you send a request, it returns something. The amount consumed can usually be measured. Monid is effectively trying to do the wholesaler’s job around those units. It gathers capabilities from many providers, gives them a common place to be discovered and paid for, and lets the buyer purchase only the amount they need.

A good wholesale software capability probably needs a few things.

• The job should be narrow enough that another system can describe it clearly.

• The result should come back in a format another program can use.

• Software should have a clear thing you can count and charge for.

• The customer should be able to switch suppliers without rebuilding their entire business.

That explains why search, enrichment, scraping, market data, media generation and verification appear so heavily in Monid’s catalogue. But where does the idea start to weaken?

Take Figma. Figma has APIs and an MCP server. In it, you can read files, work with variables and comments, and connect parts of Figma to other systems. But the reason companies buy Figma is much larger. Several people can view and edit a file simultaneously. They can iterate through designs and basically maintain a shared working state over months. Figma itself describes multiplayer files, cursor chat, comments and the shared canvas as central parts of the product.

So if you reduce Figma to one API call that says make this design, you are only capturing a small part of what people actually use Figma for.

Same with Salesforce. Salesforce is where a company keeps track of its customers and sales work. A team might use it to store contact details, past conversations, deals in progress, follow-ups and internal notes.

Getting one small piece of that information, such as a customer’s email or the status of a deal, is easy to imagine as a single paid call. But the full product is much harder to sell that way. The value comes from keeping the entire customer record and workflow together continuously. That ongoing workspace does not break neatly into tiny one-off purchases.

Wholesale software could therefore take individual capabilities out of a product and take some of its customers with them. How much of the original business survives depends on the value of everything surrounding those capabilities.

Monid is threatened when software companies start to charge for individual tasks within their products. Salesforce itself sells Agentforce through Flex Credits, with individual actions consuming credits. Its published rate is $500 for 100,000 credits, with a standard Agentforce action consuming 20 credits. That means a task such as updating a customer’s details or answering a question costs about ten cents, on top of any other charges. Companies still use Salesforce to store customer information and manage their work, but now pay separately for specific tasks the AI performs.

Salesforce is also investing in the machinery needed to support this model. It completed its acquisition of usage-billing company m3ter on July 1, 2026. The stated purpose was to bring metering and rating into Agentforce Revenue Management so businesses could track usage and bill for consumption or outcomes. That is a concrete commitment from a company whose business was built around subscriptions.

Look at Intercom, the customer-support software to get an even closer perspective of a capability becoming separable from its original workspace. A business can use its AI agent, Fin, to answer customer questions inside the support software it already uses, including Salesforce. It does not have to subscribe to Intercom’s helpdesk. Fin’s pricing starts at $0.99 per outcome, with a minimum monthly commitment. So the business can get the AI’s help without switching the software its support team uses.

Established companies can separate their capabilities, change their pricing, and maintain the customer relationship. Monid’s opportunity depends on making it sufficiently easier to compare and buy across those companies that customers prefer going through it.

Establishing a broader shift would require evidence of repeat spending, customers replacing existing subscriptions and providers winning business they would otherwise never have reached.

When would software brands disappear?

Today, buying software usually begins with a brand.

If you decide to use Canva, you open Canva and choose a Canva plan. You learn where Canva puts everything. A company might spend months choosing Salesforce and train employees on Salesforce. They might spend several more years complaining about Salesforce while continuing to renew it.

Brand choice happens before the work now, and that is the model Monid can reverse if it wants to.

Suppose I ask for the current price and reviews of a product. There may be several providers capable of retrieving them. Monid’s discovery layer can show those providers with prices and choose according to the job.

If that becomes normal, I may stop starting by picking which software company I should use. And instead, you just ask for the result.

The provider moves one level down. This could make these big brands uncomfortable since a surprising amount of software economics is built around being remembered. A 2025 study by 6sense of nearly 4,000 business buyers found that 95% bought from a vendor already on their shortlist when the buying process began.

An analogy I can’t resist is, if I want a mocha in Bangalore, I first decide which café to visit, then whether to book a Rapido or an Uber. Once I get there, I look at the menu and find out what’s actually available. Would it make me happier to start with exactly what I want and let someone figure out the rest? Probably. zomato’s of the world tries to fix this to some extent. But a cab company, where I can get in and click on what I feel like eating, gets me to a place I just pay once for all, easier! But no fun, so that’s okay.

The point is, companies spend heavily on websites, content, sales teams, demos, free trials, onboarding and brand. Once a customer is inside, they bundle more features into the account because those extra features make leaving feel more uncertain.

If I’ve already paid for Higgsfield this month, I’ll try to use the tools included in my plan before spending money somewhere else. With Monid, I can buy each task separately, so I have less reason to keep choosing the same provider. I can pick whichever works best for that job.

Monid could learn which tools work well by tracking what happens when people use them. If one keeps failing and another usually works, it could recommend the reliable one next time. But four million transactions alone don’t prove it can do this well. It needs enough results for each type of tool to make a fair comparison. Monid currently ranks tools mainly by how closely their descriptions match the request. It also shows speed and stability information, with performance playing a smaller role in its documented ranking. Learning which tools consistently do the best job remains an opportunity.

Let’s go back to the “OpenRouter for tools” claim. The ranking system is harder to copy than the catalogue.

For an amazon seller, being available on Amazon is useful. Appearing near the top when somebody searches is considerably more useful. If Monid chooses a tool behind the scenes, the user may never notice which company supplied it. Providers could then spend more effort getting Monid and other platforms to recommend them.

Smaller companies might welcome the customers this brings. Bigger brands with unique data or popular tools may prefer selling directly, where they control pricing and keep the customer relationship. They could offer only selected features through marketplaces while keeping their best tools inside their own apps.

For instance, Customers pay for Bloomberg’s combination of financial data, analytics, news and working tools. Selling individual pieces through a marketplace could bring in more buyers, but Bloomberg would have to weigh that against giving another company more influence over pricing and access to its customers.

How far this goes depends on whether providers think the extra business is worth giving up some control.

The fun consequence is that we may get software companies that barely look like software companies. Someone could build an API that takes a product link and returns accurate specifications and the latest price, then sell that service without building an entire app around it.

Today, the standard startup checklist would eventually arrive. Building the tool is only the beginning. Then you need a website, accounts, payments, pricing plans and a way to convince people to sign up. Before long, you’re writing welcome emails and posting on LinkedIn. All because the company needs to acquire, bill and retain human customers.

A wholesale marketplace can take some of it away.

Monid is already inviting API owners to bring the capability, while Monid handles discovery and metering.

There would still be work. It needs reliable infrastructure, clear schemas, security, compliance where relevant and enough support to stop the endpoint breaking. But for plenty of smaller jobs, the company may not need a consumer application at all.

This could create a long tail of tiny software suppliers that could not support the normal cost of running a SaaS business. You can’t afford a whole sales team when each request earns you a fraction of a cent. In such a case, it could work if the marketplace brings enough business without you having to find each customer yourself.

It also creates dependence. As loyalty is close to zero, a supplier could lose customers when the platform starts recommending another tool.

Does Monid need to become Amazon?

Monid has a reason to keep adding tools. More choice could attract more customers, which could encourage more providers to join. But I am not convinced the biggest catalogue automatically wins.

RapidAPI already taught us that an enormous API marketplace can exist without permanently owning the relationship between buyer and provider. Once they know which provider they want, they can go directly to it. Monid therefore needs to keep making the purchase easier. That means showing accurate prices, keeping instructions up to date, handling billing problems and helping agents find tools that work.

The mistakes can also spread quickly. An agent could keep using a tool that returns wrong information, repeating the error across hundreds of tasks sometimes.

In 2012, A faulty automated trading system placed millions of unintended orders in about 45 minutes and lost roughly $440 million. This is probably old and not an AI example, but it shows how quickly an unchecked automated loop can do damage.

In 2025, Anthropic had Claude run a small office shop. It told customers to pay into a Venmo account that didn’t exist. It also kept handing out discounts after being told to stop. Even a year after that, we are no strangers to slops.

So, to survive that, Monid’s next challenge is helping agents choose well. As we saw already, the recommendations currently rely mainly on matching tool descriptions. Reliable comparisons of actual results would make those recommendations more useful.

If it becomes very good at choosing among suppliers, it can gain plenty of marketplace power without making humans browse an Amazon-sized software mall.

The machines are the ones doing the shopping anyway.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

Stocks have been part of internet culture for longer than crypto would want to admit. The GameStop saga proved that a subreddit could move a stock by 2,400% on pure collective delusion, and finwit has since turned every ticker into a tribal identity. The cultural energy around equities is massive, and has been, for years.

But it has always been weirdly disconnected from the capital itself. The memes were on Twitter (X), the trades settled on NASDAQ, and the culture around a stock could make or break a company’s public perception without ever touching its cap table. Stocks have been internet-native content for a long time, but they never became internet-native as assets.

Then recently, Robinhood built its own blockchain and launched tokenised equities on it. And people did something really unexpectedly interesting- they started launching memecoins paired against those stock tokens. A coin called $BONER, paired against Hims, cornered 81% of all tokenised HIMS supply in its pools. And somehow, the stock tokens actually got liquid. I think Robinhood has accidentally built a wealth-effect machine. The chain is pulling two-thirds of all Ethereum L2 revenue right now and has brought in more liquidity to tokenised stocks than all institutions.

Velo: Supercharge Your Analysis

There’s never been a better time to be an analyst. Everyone has access to powerful AI models, and even non-coders can build the ideal tools they need. The difference between a mediocre analyst and a great one, however, is the same as it’s always been - access to the best information. Most are limited by poor-quality or insufficient data.

With the Velo API, we put our extensive data in your hands, allowing you and your AI to see the whole picture. Spot, futures, options, price action and order flow data - all at your fingertips. Power your AI in a single command with our MCP. Build your perfect suite of tools with Velo’s SDKs.

If you’re a corporate client and want to enhance the analysis you offer to your users, leverage Velo by plugging in our data, embedding our charts, or working with us on a bespoke package.

Choose Velo

The Speculators always get there first

Tokenised equities are not new. People have been trying to put stocks on the blockchain since at least 2018, when security token offerings were supposed to be the next ICO. tZEO had raised $134 million back then for this. Securitise got backing from every blue-chip name in finance. Binance, FTX - they all gave it a fair shot, but Mirror Protocol on Terra was one of the most complete attempts. They built fully synthetic tokens for Apple, Tesla, Google, and a bunch of other stocks, all tradable 24/7 by anyone in the world with no broker and no KYC.

And it worked really well for about a year before the SEC sued Terraform Labs partly over it, and then the whole chain collapsed. But the interesting thing about Mirror is that even when it was live and fully functional, its liquidity was terrible. You could buy synthetic Apple, but almost nobody bothered, because why would you when the real Apple stock sat right there on NASDAQ with better spreads, actual ownership rights, and a brokerage app that let you buy fractional shares for free?

The assumption was always that if you build the right infrastructure and make the assets accessible, traders will show up. But accessibility has never actually been the thing that brings people into a new market.

Joseph de la Vega wrote the first known book about a stock market back in 1688, about the Amsterdam exchange. And one of his most surprising observations was that the exchange drew in investors and gamblers in roughly equal measure. The Dutch East India Company ran real-time ship activity and trade routes, but the people who actually made its shares liquid weren’t the ones studying shipping manifests.

They were the ones placing bets on rumours about what cargo might be on a vessel that hadn’t docked yet, trading on secondhand gossip from sailors at the harbour, buying and selling based on how they felt about the weather.

The Chicago Board of Trade observed this again in 1865 when it built grain futures for farmers and found that the market only worked once speculators showed up to take the other side. This pattern goes back to financial markets themselves.

The reason is that if every participant in a market is rational and well-informed, they mostly agree on what something is worth, and when everyone agrees on price, there’s little reason for anyone to trade. You need people trading on sentiment, narrative, vibes, and the sheer excitement of watching a number go up, because they are willing to take one side of the trade. Take out the noise, and you end up with a dead market. The people who look like they are just gambling are actually performing a very important function in a new market; they are showing up before there is any rational reason to.

Crypto, of all, understands this better than anyone. Uniswap got its first real liquidity from yield-farming degeneracy in DeFi summer, and not from institutional allocators doing due diligence on its AMM design. Pump.fun tokens mostly die within a day, but the protocol still pulled in over $600 million in revenue last year because all that frantic activity generates infrastructure underneath it.

Hyperliquid launched as a perps exchange and let people trade memecoins on it, and now tokenised stocks and RWA make up 67% of its HIP 3 Volume. Every time something in crypto has tried to skip the degen phase and go straight to serious institutional adoption, it has ended up with a compliant and well-built but completely empty order book.

How a memecoin ate the float

So Robinhood did the one thing every other chain got backwards. Every blockchain ecosystem you know has followed roughly the same growth arc: launch DeFi protocols, attract builders with grants, wait for memes to show up organically, and then eventually try to bring real-world assets onchain at the very end.

But Robinhood launched with tokenised stocks as the foundation from day one and was fairly quiet for a few weeks; then a launchpad called long.xyz gained traction and started letting people create memecoins paired against the stock tokens instead of ETH or stablecoins.

That’s where it gets interesting, because the pairing changes where the liquidity actually resides. On Pump.fun, when you buy a memecoin, the other side of your trade is SOL, and the memecoin and the idea it references have no economic relationship beyond sharing a cultural joke.

But on Long, when you buy a memecoin paired against tokenised Nvidia, the router has to acquire NVDA first to fill your order. At the protocol level, every memecoin purchase becomes a buy order for the stock token on the other side. The equity exposure gets locked inside the liquidity pool as the reserve asset that prices the joke. If you have ever looked at how Olympus DAO’s bonding programme worked in 2021, this should feel familiar.

OHM sold tokens at a discount in exchange for LP tokens, and the high-quality asset on the other side went into the protocol’s own treasury. The protocol owned its liquidity, and OHM’s multi-billion-dollar market cap was denominated against a reserve it had itself absorbed.

You can see something similar with $BONER, paired against tokenised Hims & Hers. The total tokenised HIMS float on Robinhood Chain was around 73,685 tokens, and $BONER’s pools were holding 39,136 of them, roughly 53% of the entire on-chain supply. When Wall Street closed for the weekend and nobody could mint new stock tokens, the memecoin’s demand kept running while supply was frozen, and on-chain HIMS started trading at a 37% premium to its NYSE close.

After the squeeze, Robinhood’s Jersey entity expanded the HIMS float by 4.8x to relieve the pressure. But 53.1% of the new supply ended up right back inside memecoin pools. The supply response did not reduce the memecoin’s grip on the float. It just gave it more liquidity to absorb. For 11 out of all 17 tickers on the Robinhood chain, more than 40% of the wrapper’s total trading volume is a side effect of someone trading a joke.

A meme is basically a carryless cultural asset: you hold it, and your outcome depends entirely on the meme’s momentum. But pair it against a real underlying, and the cultural asset starts accumulating something outside of itself. The meme acquires carry. Finance has recycled the same primitives for centuries, but what we are starting to attach them to has never looked like this.

The accidental flywheel

Once carry exists, things start compounding in directions that are hard to predict. For example, with Artificial Inu, which is paired against tokenised Nvidia. When you buy $AI, 80% of the buy fee goes into a permanent NVDA vault that can only grow and can never be withdrawn from. When you sell, the fee splits between burning $AI tokens and locking even more NVDA permanently.

Every single trade, buy or sell, ratchets up the stock exposure in one direction. Roughly $3 million in NVDA has been permanently removed from circulation through this mechanism alone, and no human consciously decided to accumulate Nvidia stock. The memecoin’s own trading churn did it automatically.

And then Long took it a step further with LongX, which takes a 3x leveraged perpetual position on NVDA, wraps it into an ERC-20 token, and then uses that token as the quote asset for another memecoin. You now have a share being tracked by a debt security issued out of Jersey, wrapped into a leveraged perp on a DEX, tokenised into an ERC-20, and sitting as collateral inside a joke token’s pool.

That’s...Four layers of financial wrapping on top of a single Nvidia share. At one point, this whole thing held 16% of all NVDA open interest on its exchange. That’s what makes it a wealth-effect machine. A BNB Chain study found that users who entered through speculative memecoin trading converted to holding tokenised equities at 8.6%, compared to 0.6% for users offered equities directly.

Speculation was 14 times more effective at turning people into equity holders than simply showing them the equity.

Robinhood Chain is doing $4 million in daily revenue two months after launch, roughly double Hyperliquid and about two-thirds of all Ethereum L2 revenue combined. Pons alone is the second-highest fee earner across all of DeFi right now. A memecoin launchpad, outearning almost every protocol in the ecosystem.

But there are also things worth watching. Robinhood is running a 90-day gas subsidy for its own wallet users that expires around late September, and nobody knows what happens to transaction volumes when those users start paying gas for the first time. The permanent vaults also haven’t been tested at the scale where things tend to break. And the weekend premium phenomenon, where on-chain prices decouple from the NYSE because issuers cannot mint new tokens over the weekend, might be a feature right now, but it could start looking like a problem to a regulator who has never heard of a bonding curve.

Regardless, it still does matter beyond just Robinhood. Every memecoin that pairs against a stock token locks up tokenised equity inside its pool. Before this, the money always flowed away from the meme. You made the joke on Twitter, you traded the ticker on Robinhood, and the value ended up with the platform or the exchange. The meme itself never held onto anything.

What we are watching right now is a joke token accumulating a balance sheet of tokenised equities through its own trading activity. Blockchain gives it a common programmable environment, where some arbitrary things can start having economic relationships with other arbitrary things.

A meme can be paired against a stock. A leveraged derivative can become the collateral for a cultural bet. A weekend supply freeze can create an arbitrage that funds a sequencer. And none of these was deliberately designed by a product team. They emerged because the primitives were composable and early adopters found them first.

That’s it for today!

Vaidik

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 165,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.

Dario Amodei, who runs Anthropic, said that the companies building the most powerful AI models should deliberately slow down. After a few hours, Sam Altman, Elon Musk and Demis Hassabis had broadly lined up behind the idea. This bothered me only because Elon Musk agreeing with this particular gang is an event on its own.

Essentially, the people pushing AI forward the fastest were now suggesting that perhaps everyone should slow down. Including, presumably, themselves?

Donald Trump was not convinced. Speaking in Ireland, he downplayed the warnings and returned to the geopolitical argument that has followed every attempt to regulate AI in America.

China is not going to stop, so why should the United States? His administration’s position remains that losing the AI race would be worse than moving too quickly.

Amodei is worried about models becoming capable faster than humans can understand or control them. OpenAI has already temporarily slowed parts of its own training while strengthening safeguards around models it now considers capable of serious cyber work. The company’s chief scientist, Jakub Pachocki, wrote last week that he expects AI systems to increasingly contribute to their own development and called the situation one requiring “extreme caution.”

The internet and I naturally arrived at the more entertaining possibility that all these people had seen something terrible and were politely trying not to tell us.

Maybe. But can AI end the human race? No one has a non-diplomatic answer, and we have no precedent to support either view.

As of now, we know AI agents are getting access to things.

Gmail, GitHub, Slack, calendars, company databases, cloud servers, social accounts, payment methods, CRMs, the ability to generate media, buy software, and publish content that other people can see.

So now, when you ask what this model is allowed to do, that becomes one of the larger infrastructure markets created by agents. The industry has been improving intelligence. Now companies have to build the layer around intelligence that decides where it is allowed to go.

So, this is the best monday to talk about Aident, one of the companies making that bet.

Not the possible apocalypse, we probably can’t stop that. But I have been testing a small version of the problem with Aident Loadout, which gives an AI agent access to outside tools while putting credentials, approvals, spending controls and an audit trail around what it does. Let’s see what it tells us about the broader AI permission problem and how the Aident Loadout experiment puts it to the test.

TOKEN2049: One Crypto Stage, Myriad Perspectives

The future of finance is being written, and you are invited!

This year, TOKEN2049 will celebrate the largest crypto and TradFi gathering. People from the largest banks, fund houses, and VCs will share the stage with leaders from top prediction markets, digital asset exchanges, crypto advisories, and blockchains.

TOKEN2049 is your front-row seat to the next financial revolution, shaped by deals, partnerships, and product announcements.

Dare to miss this lineup at your own risk!

Let us spoil you a bit with an exclusive 10% discount coupon!

TOKEN2049, Here I Come

In 2006, Twitter had a similar headache.

Third-party developers wanted their apps to use Twitter accounts, but the obvious way to make that work was terrible. A user would give another application their Twitter username and password, and that application would hold the credentials indefinitely.

Blaine Cook, then Twitter’s chief architect, started discussing a better system with people working on OpenID. What emerged in 2007 was OAuth, an open standard for delegated access. Instead of handing a photo-printing website a user’s entire Flickr password, for example, they could give it permission to access the photos it needs. The service got a limited token, and the user kept the master key.

That decision became part of the plumbing of the modern internet. Whenever Google asks whether you want to allow some app to access your calendar, or GitHub asks whether an application may read your repositories, you are living inside the solution to that old problem.

AI agents make this awkward again. OAuth is very good at answering one thing. Does an app have permission to access Gmail? It was designed for a time when apps did one clear, predictable thing. A calendar app had calendar features. A photo printer printed photos. We knew what we wanted out of the software. An agent is different because its useful feature is precisely that you have not programmed every action in advance.

For example, when you tell it to organise a launch, it decides this requires researching competitors, generating images, writing posts, opening X, publishing something, and perhaps updating a spreadsheet afterward. Tomorrow, you ask for something different, and the same agent discovers another set of tools.

So “has access to X” is no longer a sufficient security model.

Should it publish this particular post?

Can it delete one?

Can it message customers?

Can it spend $2 on research without asking? Can it use the same Gmail connection when you switch from Claude to another agent next month?

This is now permission because software is no longer that predictable. Agents are moving into companies faster than the systems controlling them.

Every time a company wants one agent to work across software owned by several different vendors, somebody needs to handle authentication, tools, policy and records of what happened. The more models become interchangeable, the stranger it becomes to rebuild those connections for every new agent.

Okta found that 81% of CISOs and security executives were concerned about AI agents having excessive access, while fewer than half were confident they could identify every agent in their organisation or centrally control what each one could do. Microsoft has responded by extending Entra identity infrastructure to agents, arguing that every agent should have its own identity, scoped permissions and audit trail.

That makes identity and access management a more useful neighbouring market than the entire AI industry. Traditional IAM already handles who can enter which system. Agents need another layer that decides what they can do once inside, like read or write, spend or not spend, delete or ask first.

Aident calls its product Loadout. The easiest way I’ve found to think about it is as a capability and control layer placed between an agent and the software it wants to use.

Aident says Loadout already connects to more than 1,000 apps and gives agents access to 27,000 different actions.

Instead of giving your passwords or API keys directly to the AI agent, you can store them inside Aident Vault. Connect an app once, and you can then use the same connection across tools like Claude Code, Codex, Hermes, ChatGPT, and Cursor.

Aident also keeps a record of what the agent actually did. You can see which action it ran, whether it succeeded, and how much it cost.

But first, let’s clarify what Aident means by “safety.”

Aident does not make an AI model aligned. It cannot guarantee that Claude, Hermes or some future agent will never make a stupid decision. What Aident tries to do is limit what happens after that mistake.

Your credentials can stay hidden from the agent. You can restrict what an app connection is allowed to do. You can also check an action before the agent runs it.

Aident splits agent control into two checks. Spending approval asks only whether the agent may spend money. You can let it auto-spend small amounts, like $10 on research, without asking you each time.

Action approval asks whether it may perform a specific act. That check applies even when the act costs nothing. Sending an email, deleting a file, publishing a post, or changing a customer record can still cause real harm. The two risks are different. Money risk is about the budget. Behaviour risk is about irreversible or high-impact actions. So an agent might spend $5 on its own, yet still need your OK before it emails anyone or deletes anything.

Finally, Aident’s audit trail keeps its own record of what happened. So instead of asking the agent if it has send that email, you can check an independent log to see whether the action actually ran.

So I gave it something ridiculous to do

I connected Aident to Hermes and asked it to create an online pet company.

The eventual brand was Blorb & Co. I wanted ten digital pets, all with specific personality, rendered in 3D. The agent had to research similar products and conversations on X, choose an audience, develop the positioning and

tone, create a logo, posters and campaign copy, animate a couple of characters, give all ten voices, build a quiz and website, and then actually publish an organic campaign.

Through Aident, Hermes used research tools such as Exa and SerpAPI and generated the characters and creative assets through Fal.

Then it produced voice lines through ElevenLabs, made two animated clips, created trading cards, and built a small meme system that was so stupid, but maybe it can be fixed with a bit more cleaning. I eventually connected my X account, and the agent published six posts.

Later I connected to Vercel and it deployed the interactive pet site.

The whole experiment cost roughly $3 in Aident usage.

The agent was not working through a workflow I had designed beforehand.

It searched the available capabilities, inspected what a tool expected, checked prices and then used the tool. When RedFoxHub’s X search broke, it moved to SerpAPI instead. The route changed while the work was happening.

A few things broke during the experiment. The text-to-speech tool expected a different input than the agent was sending and later ran into rate limits.

Vercel was the biggest headache. File uploads did not work properly, and some deployments looked successful even though the files were missing. I eventually got around this by putting the entire website into a single HTML file.

@adoptablorb.vercel.app

OAuth also required me to manually approve account access. That is actually a good security feature, although it was not always obvious what I needed to approve or where.

A permission layer can make outside software available to an agent, but it still inherits the unreliability, schemas, rate limits and strange decisions of all that outside software.

Aident’s product looked to me as an attempt to put one controlled surface over a very messy internet.

Btw, figure out which pet you are ( I am Nugget) -

Which Pet Are You?

The next question is how Aident makes money if much of what it connects to belongs to somebody else. Its current model gives us a rough answer. Loadout supports pay-as-you-go usage, paid Pro and Team plans, and additional credit purchases. For some built-in services, Aident supplies the underlying provider access and deducts the cost from a user’s Aident balance.

For other applications, you connect your own provider account or API key, in which case that provider can bill you directly.

That creates two businesses sitting on top of each other. The first resembles a software marketplace. Aident aggregates demand, buys or routes access to outside capabilities and gets paid as those capabilities are consumed. More agents doing more work should mean more transactions. This is the obvious revenue line, and it is the one I experienced while watching a few cents disappear every time Hermes generated something or ran a search.

But I would not want that to be the whole company. If Aident’s value were simply adding a margin to somebody else’s API, a customer making ten million calls would eventually work out that it can negotiate directly with the provider. Aident itself already supports BYOK, where users bring their own provider credentials and Aident’s deduction for the external action can fall to zero.

That business is the control plane. A company may happily bring its own OpenAI account, Salesforce contract, GitHub organisation and cloud bill while still paying Aident to decide which agents can reach them, store the credentials, apply approval rules, enforce spending policy, maintain the integrations and keep an audit record. At enterprise scale, governance and security can become the thing being sold rather than the underlying API call. Aident’s public pricing already points in this direction with paid plans and an enterprise tier around security, governance and customised deployment.

That is probably the healthier economics if Aident succeeds. Marketplace revenue grows with the amount of work agents perform. Subscription and enterprise revenue grow as Aident becomes more important to the organisation. One monetises traffic. The other monetises trust.

Aident is early to an opportunity that other companies have clearly noticed too. Composio helps AI agents connect to apps like Gmail, Slack and GitHub without developers having to build every connection from scratch. It supports more than 1,500 apps and is especially useful for companies building AI products for lots of users because each user can keep their own accounts and permissions separate.

Arcade is more focused on big companies. It lets businesses control exactly what an AI agent is allowed to do, keeps login details away from the agent and records every action. Those records can also plug into the company’s existing security systems.

Zapier is the giant in this market. Its MCP product lets AI assistants use more than 9,000 apps and 40,000 actions. Zapier says users have already created more than 450,000 MCP servers and made 18.5 million tool calls through them.

Zapier already owns a gigantic map of how business software talks to other business software.

The distinction is becoming smaller as Zapier moves deeper into agents, so saying that Zapier only runs fixed workflows would now be unfair. But its historical centre of gravity is still automation. Someone knows the workflow, connects step A to B to C, and wants it repeated reliably.

Aident starts from a slightly different assumption. The agent may not know which tool it needs until it starts the job. It discovers capabilities while working, inspects the current schema, preflights an exact action, checks the cost and then executes it. The wedge is making tool access, spending, and permissions feel native to an agent that decides its route as it goes. Even Aident’s own comparison concedes that Zapier remains difficult to beat for repeatable business workflows.

That also tells you who Aident is for right now.

If you want Gmail to copy every attachment into Drive, this is probably unnecessary. Zapier is excellent at those things, which is an underrated quality in infrastructure.

Aident becomes more interesting for developers building agents, teams using several different agent clients, and companies where agents are expected to move between research, communication, coding, media and outside APIs without a human rebuilding the integration every time.

I would say that the AI model itself may turn out to be the most replaceable part of an agent stack. A company might use Claude for one team today, move a workflow to ChatGPT next month, give developers Cursor or Codex, and switch again when a better model shows up a few months later.

Since the models are improving so quickly, committing too deeply to one of them may not make much sense.

Once agents start doing real work, companies still have to decide where the line is. Maybe an agent can draft a customer email but not send it. And if something goes wrong, someone needs to know exactly what it did.

Those decisions are just company rules, like the ones normal employees follow, but now software has to understand and follow them. Access and spending rules are currently fragmented across disconnected consoles, tokens, cards, and manual offboarding checklists.

As the market matures, the permission layer becomes an enforceable job description defining what an agent can see, spend, alter, publish, and escalate.

That is why I think Aident is sitting in an unusually good pocket of opportunity, even if the company itself is still early and parts of the infrastructure were rough in our test. Governments will probably spend years arguing over how powerful the models themselves should be allowed to become.

If agents become normal parts of how companies operate, that rulebook has to live somewhere outside the model itself. Which makes the opportunity slightly different from the race to build the smartest agent. The valuable position may be owning the layer a company trusts enough to let all of those agents through.

—Thejaswini

Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000+ subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌

📩 Fill out this form to submit your details and book a meeting with us directly.

Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.