AI agents made 138 million crypto payments. What did they buy?
Chainalysis counted more than 138 million x402 transactions on Base during its 2026 reporting period. They covered payments for AI inference, data services, code analysis and other tools. The volume points to heavy protocol activity, but it does not reveal how many independent agents or customers were involved, or whether the services made money.
- Reported uses include AI models, search and data APIs, code analysis, agent tools and token-related services.
- In a separate sample through September, more than 87% of payments ranged from $0.0001 to $0.10.
- Wallets received funds more often and spent them sooner, Chainalysis reports. Neither pattern proves customer satisfaction or profitability.
- Early activity included PING, a pay-to-mint experiment that helped drive a surge in transactions.
What x402 enables
Coinbase published x402 in 2025. The protocol uses the HTTP 402 “Payment Required” response to let a web service ask for payment before providing a resource. In a supported flow, software can make the payment, often in a stablecoin, a token designed to track a stable value, then resubmit its request with proof of payment.
That setup could work for a program paying for a model response, a search query or access to a data API. Rather than opening an account with every provider, software may be able to pay for each request. The 138 million figure covers Base, not every AI-agent payment made across crypto or traditional payment systems.
The appeal is clear. Billing systems designed for people and recurring plans can be awkward for tiny, automated charges. But x402’s ability to move a payment says nothing about whether a service returned useful results, or whether pay-per-request billing is better than subscriptions, prepaid credits or aggregated charges. That depends on costs, reliability and how customers use the services.
What the 138 million figure measures
In its October 14, 2026 study, Chainalysis reported more than 138 million x402 transactions on Base in the 12 months ending July 1, 2026. The firm’s earlier account, “The New Rails: How Digital Assets Are Reshaping the Foundations of Finance, ” described well over 100 million cumulative transactions through Q1 2026. The October figure is a later snapshot. It was not established by the earlier account.
Chainalysis found activity involving AI model inference, code review and analysis, financial intelligence, data and search APIs, agent frameworks, payment infrastructure, games, token issuance and gambling-related services. These categories give a sense of the services receiving payments. The figures discussed here do not give comparable dollar totals or provider earnings for each category.
A “transaction” is not the same as an “agent.” Chainalysis uses wallets interacting with x402 as an on-chain proxy for agent activity. A wallet might be controlled by a person, a company, a script or an autonomous system. One entity can also control multiple wallets. So the count does not establish 138 million distinct agents, buyers or completed tasks. Nor does an on-chain transfer prove that a request succeeded or produced anything useful.
The transaction count is not a measure of revenue, either. The figures discussed here do not provide a full account of independent buyers, provider costs, refunds, failed requests or net earnings. Repeated transfers can drive up activity without adding the same amount in new customer spending or provider income.
Small payments, separate sample
Chainalysis found that more than 87% of payments in an extended sample through September ranged from $0.0001 to $0.10. That sample extends beyond the July 1 end date of the 138 million transaction count. The figures cover different periods, so the payment-size percentage should not be applied to all 138 million transactions unless the underlying data and definitions are confirmed to match.
Even small charges add up. One five-cent request every second costs $180 an hour. At ten cents per request, that comes to $360. These are illustrative calculations, not observed spending totals, and they exclude fees and other operating costs. An agent making frequent calls needs budget controls and rate limits as much as it needs a payment protocol.
More refills, faster spending, but what does that mean?
Chainalysis says the share of qualifying wallet cohorts that received a second deposit within 30 days rose from 70.6% for the Q3 2025 cohort to 89.3% for Q2 2026. A second deposit means the wallet received funds again. The calculation excluded wallets that sent on only one day, and Q2 2026 was the newest cohort with a complete 30-day observation window.
The firm also found that the median time from funding to first payment fell from 73 days for late-2025 wallets to 18 minutes by Q3 2026. The share of wallets that spent within an hour of funding climbed from 4% in Q3 2025 to 67% in Q3 2026.
These patterns suggest more immediate and repeated use. But another deposit is not a customer review. It could reflect regular service use, automatic replenishment, testing or something else. The figures do not show whether the wallets belonged to unrelated buyers or whether providers earned more than they spent. Funding and retention patterns are useful signs of activity, not proof of a successful business. Questions about who is actually using agentic payment protocols remain part of the demand test.
Early activity included a speculative experiment
Chainalysis has named PING, a pay-to-mint experiment, as one driver of an early x402 surge. In its earlier account, a user paid 1 USDC after an endpoint returned a payment request. Chainalysis reported more than 150, 000 PING transactions in its first month and an activity increase of over 10, 000% in a single week.
That example shows how transaction counts can jump without proving lasting demand for information or computing power. A repeatable minting mechanic can generate plenty of transfers. The later report identifies a broader mix of service categories, but the figures discussed here do not show exactly when, or whether, information services overtook promotional activity.
Where was the activity attributed?
Chainalysis attributed 30% of the x402 agent-to-service transactions it could geolocate to U.S.-funded wallets. For AI model inference, it attributed 88.1% of geolocated payments to U.S.-linked wallets. The firm also found that Chinese-linked wallets accounted for about nine in ten transactions in its tagged security and code-analysis category. Most of those payments went to AgentLISA.
These percentages cover geolocated or categorized subsets, not all x402 activity or national adoption. An inferred wallet location does not necessarily establish where its owner lives. The figures discussed here also do not state what share of all activity could be geolocated. Read the percentages as estimates for the observed subsets, not as a map of every user. For broader regional context, Chainalysis’s 2026 North America crypto adoption report covers crypto activity across the region.
x402 and AI agent payments: key questions
-
Does 138 million mean 138 million AI agents?
No. Chainalysis counted x402 transactions on Base. Wallet activity does not prove that each transaction came from a different, autonomous agent.
-
What services received payments?
Chainalysis identified AI model inference, data and search APIs, code analysis, financial intelligence, agent tools, games and token-related services. The figures discussed here do not provide a full breakdown of spending by provider.
-
Do repeat deposits prove the market is profitable?
No. They show that wallets received funds again, not that customers were satisfied or providers made money from sustainable businesses.
-
Does the payment-size figure cover all 138 million transactions?
Not necessarily. The finding that more than 87% of payments ranged from $0.0001 to $0.10 comes from a separate sample through September. The transaction count covers the 12 months ending July 1, 2026.
-
What evidence would show that an agent-payment market is working?
Comparable spending by service category, independent repeat buyers, verified useful results and provider revenue measured against costs would offer a stronger test than transaction counts alone.
Promising payment rails still need proof of demand
x402 is a credible experiment in payments sized for software. Stablecoins give software a relatively predictable unit for pricing requests, while on-chain settlement can support automated purchases without a conventional account relationship. Whether that beats existing billing depends on reliability, costs and the safeguards buyers and providers need.
The infrastructure is only part of the work. Wallets still need funding, spending limits, vendor permissions and a way to handle failed requests. Wallet tools for AI agents address one part of that operational challenge. Chainalysis’s figures point to faster, more frequent x402 activity, but they do not show how much comes from independent buyers paying for reliable services at a profit. That is the tougher test, and transaction totals alone cannot answer it.
For a different kind of AI service, Google Flow offers a generative filmmaking tool. Its presence does not change what x402 transaction counts tell us: protocol activity, by itself, does not prove lasting demand.
