Hugo Freire

When AI agents start funding other agents

AI agents are learning to buy the resources they need. The next step is financing the businesses they help build. Ethereum already has much of the infrastructure those businesses would need.

Read the essay
A line of light connects a team working beside servers to a second, smaller workshop at dusk.

Imagine a small business with five agents. One finds customers, one builds software, one buys data, one checks the work, and one manages the budget. The owner defines their authority and takes responsibility for what they do. After the business starts billing customers, it needs capital to buy more computing capacity, license better data, and reach more customers.

Some of that capital could come from a community that already uses the service. Some could come from agents acting on behalf of investors. The twist is that the same agent could buy a service today, commission a supplier tomorrow, and help fund the next business that supplies it.

My bull case is that Ethereum becomes a capital market for these businesses. Cheaper AI makes it easier to start them. Programmable payments make it easier to trade. Stablecoins, collateral, and confidential settlement give them reasons to keep capital onchain after the first payment clears.

What would investors fund?

The initial coin offering boom showed how quickly an open market can assemble capital. PwC and the Crypto Valley Association counted $7.0 billion across 552 ICOs and security token offerings in 2017, followed by $19.7 billion across 1,132 in 2018. Activity contracted in the second half of 2018. A second wave could fund operations that already have a record of completed work, costs, and repeat customers. PwC and Crypto Valley Association, March 2019

Suppose a group of small retailers finances a shared service that keeps catalogues up to date and compares suppliers. The service could sell prepaid usage, issue membership in a cooperative, or offer a contractual share of revenue. Each instrument gives its holder different rights. The token format alone tells you little about what the holder can enforce.

For a business already earning revenue, the most interesting decisions may look ordinary. Should it buy more computing time, secure a data licence, fund a supplier, or keep cash for a slow month? Agents can help make those decisions within a mandate set by the people who own the funds.

Why now

Stanford’s 2025 AI Index measured a fall from $20 to $0.07 per million tokens for inference at roughly GPT-3.5 level on one benchmark between November 2022 and October 2024. That is more than a 280-fold drop at a fixed capability level. The full cost of running a business still includes verification, supervision, and customer acquisition. Stanford AI Index

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini with open weights. Its checkpoint is about 20 GB, and the lab supports local inference on Apple silicon with at least 32 GB of memory. A small firm can run routine work on its own machine and pay for a larger model when a job warrants it. That makes it easier to start with a modest budget and expand as customers arrive. Inkling 2 Mini announcement · Model card

On the same day, METR estimated that the 50% success horizon doubled about every 3.6 months across two combined task suites. The horizon measures task difficulty in terms of how long humans take to do the work. The result is useful for thinking about verifiable tasks such as testing software, while judging an ambiguous result or negotiating with a supplier still calls for oversight. METR capability report

As more teams can build the same software, buyers can demand lower prices. The businesses worth financing may be the ones with reliable distribution, hard-to-replace data, or a way to prove what their agents delivered. Those advantages can outlast any one model release.

How agents pay

Pay-per-request already has measurable use. Visa and Artemis counted 109.6 million x402 transactions and about $15 million in adjusted volume from launch in May 2025 through 21 April 2026, after filtering identified wash and test activity. An October report from Coinbase and The Block counted 249 million x402 payments through September 2026 and said activity had settled near 20 million transactions a month after early peaks. The reports use different filters and coverage, and the later count includes payments beyond AI agents. Visa and Artemis · Coinbase and The Block, pp. 17–18

On 19 November 2026, AWS Marketplace added x402 payments for select products through AgentCore. An agent can buy one search or call a paid service within a budget and an approved list of suppliers. AWS puts the charges on its usual dollar invoice. This matters because one payment standard can reach buyers who use traditional billing and buyers who settle in stablecoins. AWS announcement

Wallets are moving in the same direction. MetaMask introduced scoped, time-limited permissions on 6 April 2026. Those permissions let an operator define what an application can spend and for how long. Once a customer can authorise a budget instead of signing every purchase, a small supplier can sell a single query, a data licence, or a few minutes of compute as the work calls for it. MetaMask Advanced Permissions

Ethereum’s starting position

An agent business needs more than a checkout. It needs somewhere to hold cash, pledge collateral, settle with suppliers, and keep track of who owns what. Ethereum already hosts a large share of the stablecoins and transferable real-world assets that could serve those needs.

Onchain assets, 21 November 2026 snapshot
AssetsEthereum mainnetAll tracked networks
Stablecoins$133.3B$273.2B
Distributed real-world assets$14.3B$40.9B

RWA.xyz tracks these categories separately. Issuers can restrict who may hold and transfer distributed real-world assets. Ethereum mainnet · All networks · Definitions

That puts mainnet at roughly 49% of tracked stablecoin value and 35% of distributed real-world asset value before counting its layer 2 networks. The Ethereum dashboard showed both balances down over the preceding month, while holder counts rose. The scale and range of assets are the advantage I see here. A new business can choose from markets that already exist as it moves from taking payments to managing reserves.

Two changes could make that base easier to use. Ethereum’s Fusaka upgrade went live on 3 December 2025 and introduced PeerDAS, which lets the network expand the data capacity available to rollups. On 4 November 2026, Gnosis launched Rollup X, a development network for the Ethereum Economic Zone, and set a spring 2027 target for its own transition. Its demo combined a rollup swap and mainnet liquidity in a single transaction. If that design reaches production across more zones, businesses could draw on shared markets while keeping less working capital idle across chains. Ethereum Fusaka upgrade · Gnosis announcement

Commercial privacy

A company can use public settlement and still need to protect salaries, supplier discounts, and treasury positions. Cheap agents make that more important. A competitor can monitor public wallets continuously and look for payment patterns that reveal a customer relationship or a cash shortage.

Two offices share the same illuminated network. Clear windows reveal the first office, while blue privacy glass obscures most of the second.
Businesses can share payment infrastructure while restricting access to commercially sensitive amounts.

Zama’s 21–24 January 2026 Ethereum auction kept bid prices public and encrypted the quantities people wanted. The auction calculated a clearing price from the encrypted data. Zama reported $118.5 million in commitments, including related sales, and $44 million paid by winning bids. On 12 March, GSR reported an over-the-counter Ethereum trade using the same protocol. Both parties completed identity checks, and the amount stayed encrypted onchain. Zama auction results · GSR trade announcement

By 31 October 2026, Zama reported over $122 million deposited in shielded positions, up from over $90 million in September. Confidential vaults accounted for more than $102 million of the October total. Its July engineering benchmark reached 1,040 confidential token transfers per second on eight H100 GPUs, 2.6 times its previous result on the same hardware. The test measured computation on one node. Production settlement also depends on the network. Zama October update · July benchmark

That makes confidential treasury operations more plausible. The next step is getting them into the wallets, custodians, and markets companies already use. Wallet addresses can still reveal relationships even when amounts are encrypted. Businesses still need rules for who can see records and how auditors verify revenue and solvency. Zerion and Zama on wallet privacy

The bull case for ETH

The biggest opportunity is capital that stays on Ethereum. One million agents spending $1 a day would generate $365 million in annual payments. At three days’ expenses per agent, that operation would need only $3 million in pooled working cash, which it could hold in stablecoins. The upside for ETH grows if the businesses behind those payments also borrow against ETH, hold it as collateral, or settle on rollups that pay Ethereum for data.

  • Settlement. Rollups that post data to Ethereum buy block space in ETH. Cheaper data can attract services with many small transactions. The impact on fee revenue then depends on how much demand grows as unit costs fall. Ethereum’s data scaling roadmap
  • Collateral. An onchain lender can accept ETH as collateral while funding a business in stablecoins. If more firms borrow this way and keep positions open, the capital committed matters more for the ETH thesis than the raw number of payments.
  • Security. Validators stake ETH to secure the network. As more financial activity settles there, the incentive to rely on Ethereum’s security grows. Issuance and fee burn still determine how that activity affects supply. Ethereum staking · EIP-1559

The early use cases may be practical forms of financing. A group could prepay for compute, fund a data supplier, or release money in stages when work passes independent checks. Venice’s DIEM offers an example of tying a token to a daily API credit. These markets can grow as agents get better at comparing recurring costs and checking whether a supplier delivered. Venice tokenomics

What I’m watching in 2027

Agents can amplify the same feedback loops that drove the ICO boom. If A funds B, B hires C, and C buys A’s token, the same money can create the appearance of demand. When many agents use similar models and mandates, they can also make the same trade at the same time.

I’ll look for repeat customers outside that circle, collateral that stays committed, and a clear record of what new financing buys. I’ll also watch whether confidential positions retain deposits after incentives taper and whether rollups can draw on shared liquidity in production.

I’m bullish because the essential pieces are arriving close together. A small team can already run a useful model locally, give it a budget, and buy services by the request. If those teams build businesses customers keep paying for, Ethereum can give them a common place to raise capital and put it to work.

The next market on Ethereum could finance work that agents already know how to do.

Sources