1. What did Chainalysis find about AI agent payments?
On October 14, 2026, blockchain analytics firm Chainalysis reported that autonomous software agents have made more than 138 million transactions over the x402 payments protocol, which runs on Coinbase's Base network. The agents paid in crypto for data, computing power and access to artificial intelligence models, and wallets linked to the United States accounted for 88.1% of the AI model payments whose location could be identified.
The figures come from a Chainalysis report titled Inside North America's $3 trillion crypto economy: Agentic commerce accelerates, released on Wednesday. It is one of the first attempts to measure, with on-chain data, how software that acts on its own is beginning to spend money, and where that spending originates.
The report counts only transactions that can be identified as x402 activity. It does not attempt to size the full global market for payments related to artificial intelligence, a limit that matters when reading the headline numbers.
2. What is the x402 protocol?
x402 is a payments protocol incubated by Coinbase that lets an AI agent pay for a digital service automatically, in crypto, without a person approving each purchase. The name echoes the HTTP 402 status code, which was reserved in the early web for a "payment required" response but was rarely used in practice.
In a typical exchange, an agent requests a resource such as a dataset, a block of computing time or a query to an AI model. The service answers that payment is due, the agent settles the amount from its wallet, usually in a stablecoin, and access is granted. The whole sequence can happen in seconds and involves no card number, bank account or human signature.
Payment companies have taken notice. Visa, Mastercard, Stripe and other firms have backed the protocol, and Coinbase has begun allowing business customers to accept payments from AI agents. That institutional interest is part of why on-chain measurements of the protocol attract attention.
3. How dominant is the United States in AI model spending?
Of the AI model inference payments that Chainalysis could place geographically, 88.1% came from wallets linked to the United States. Wallets linked to China made up just 0.1%. Inference is the stage at which a trained model produces answers, and developers typically buy it by the request or by the token, which makes it a natural fit for small, frequent payments.
Eric Jardine, head of research at Chainalysis, offered one possible explanation for the gap. Chinese developers, he said, may be more inclined to use open-weight models such as DeepSeek and Qwen. Because those models can be downloaded and run on a developer's own hardware, they reduce the need to buy metered access from a hosted provider.
Jardine also cautioned against overreading the pattern. Chainalysis would not go so far as to say that open-source AI is starving Asia's on-chain inference market, he said. The result describes who pays through x402, not the total amount of AI usage in any country.
The picture reverses for security and code-analysis services. Chinese-linked wallets accounted for roughly 90% of the geographically identified payments in that category. Much of the money went to AgentLisa, an AI-powered platform that audits smart contracts.
The contrast suggests that agent payments are not uniform across regions. Spending on model access is concentrated in the United States, while a specialized auditing service draws most of its identified payments from Chinese-linked wallets. Both findings are limited to the transactions Chainalysis was able to attribute to a location.
4. Why are the payments so small?
More than 87% of the x402 payments analyzed were worth between one-hundredth of a cent and 10 cents. Amounts that low are not economical to process over traditional credit card networks, where fixed fees and minimum charges can exceed the value of the purchase itself.
This is the core argument for stablecoins in machine-to-machine commerce. Because they can settle fractions of a cent, agents can buy small units of data or compute as they need them. Stephanie Cohen, chief strategy officer at Cloudflare, has said stablecoins are particularly well suited to the kinds of transactions that AI agents are expected to make at first.
The same smallness is a caution. A very large count of tiny transactions does not necessarily amount to large economic value, and the report's own data shows how modest the totals still are.
5. How large is the market, and is the activity genuine?
In July, x402 processed roughly 75 million payments worth about $24 million over a 30-day period. Set against 138 million cumulative transactions, those numbers show that volume has grown quickly while dollar value remains small.
Earlier research has also raised questions about artificial activity on the network, meaning transactions that may be generated to inflate usage rather than to buy something of value. Transaction counts alone cannot show how much of the activity reflects real commercial demand, and the Chainalysis findings do not settle that question.
Yat Siu, co-founder and chairman of Animoca Brands, has estimated that as many as 100 billion AI agents could be transacting online within a few years, and using crypto to do so. Such forecasts remain projections, not measurements.
6. Are agent wallets becoming more active?
Chainalysis points to several signs of maturing behavior. Among wallets first funded in the second quarter of 2026, 89.3% received additional funding within 30 days. For wallets first funded in the third quarter of 2025, the figure was 70.6%. The analysis excluded wallets that made payments on only one day.
Wallets are also moving faster from funding to use. The average time between a wallet receiving funds and making its first x402 payment fell from 73 days in late 2025 to 18 minutes in the third quarter of 2026. That change suggests that the tooling for creating and operating agent wallets has become far easier to deploy.
A report author quoted in the research, identified only by the surname Tejwani, wrote that AI agents are becoming economic actors, with every agent effectively a new user of the blockchain.
7. Why does location attribution matter?
Chainalysis assigns activity to a region by linking wallets to services and infrastructure it can identify, which is why the report speaks of wallets linked to a country and not of people or companies located there. Many wallets cannot be placed at all, so every percentage in the study describes only the identified portion of payments.
That distinction affects how the 88.1% and 0.1% figures should be read. They show the distribution among attributable AI model payments on x402, not the share of worldwide AI spending. A developer in one country can use a wallet or service tied to another, and spending that runs through ordinary cloud billing or credit cards would not appear in the data at all.
For readers comparing the result with other measures of AI adoption, the practical lesson is that on-chain data captures one payment rail. It is a useful window into agent behavior on that rail, and a partial one for the AI economy as a whole.
8. Why do stablecoins suit machine-to-machine commerce?
A stablecoin is a token designed to hold a steady value, most often one US dollar, so a payment of a tenth of a cent is worth the same when it is received as when it was sent. That predictability is important for software that has to budget in advance and cannot tolerate price swings between the request and the settlement.
Card networks were built around people at checkout counters and online carts. They rely on identity checks, chargeback rules and per-transaction fees that make sense for purchases of several dollars but not for fractions of a cent. A blockchain such as Base, a network built to run on top of Ethereum, can finalize a transfer within moments and at a very low cost, which is why it is a natural home for protocols like x402.
There are trade-offs. Agents that hold funds need safeguards to limit what they can spend, and businesses accepting agent payments need ways to tell legitimate traffic from automated abuse. The report does not examine these controls, but they will likely shape how quickly the model spreads beyond early adopters.
9. What does the funding data suggest about agent behavior?
The jump in repeat funding is a more telling signal than raw transaction counts. A wallet that is topped up again within a month implies that whoever operates the agent expects to keep using it, which is different from a one-off experiment. The exclusion of wallets active on only a single day was meant to filter out that kind of short-lived testing.
The collapse in the delay before a first payment, from 73 days to 18 minutes, points the same way. Earlier wallets often sat idle after being funded, while recent ones appear to be created and put to work almost immediately. That pattern is consistent with agents being deployed by software and templates, not configured by hand.
Even so, these are behavioral indicators, not proof of profitability. They show that agents are being used, but not whether the services they buy generate enough value to justify the spending.
10. What happens next for agent payments?
The next test is whether the volume of agent payments rises in dollar terms, not only in transaction counts. Observers will be looking for evidence that more of the spending reflects sustained commercial demand and that the artificial activity flagged by earlier research is a shrinking share of the total.
Chainalysis concludes that transactions remain small but that autonomous software is beginning to develop spending patterns of its own, with the United States leading in paid AI services. Further reports using the same methodology would show whether the geographic split and the funding patterns persist.