A headline is circulating that AI agents made more than 138 million crypto payments. That total is not what the Chainalysis posts that publish the size data actually measured. Over the 16 months ending September 2026, the October North America post found that more than 87% of x402 payments on Base fell between $0.0001 and $0.10. The most common size was one to ten cents.
That range belongs to that sample alone. It cannot be laid over a different cumulative count, and it cannot be multiplied by a guessed average. Chainalysis did not publish a dollar total for this activity. A guessed one would be fiction with a decimal point.
If you are budgeting AI automation, a count of times software touched a payment protocol will not help you. Ask who funded the wallet, what the call was for, and whether an outside vendor got paid again after the promotion ended.
A vending machine, not a workforce
x402 is a payment protocol Coinbase developed. Coinbase Ventures has said the company incubated it. Chainalysis’s June post notes that the name comes from the old HTTP status 402, “Payment Required.” A server asks to be paid, software pays (commonly in a stablecoin), and the task continues without a person approving every charge.
That last clause describes the protocol. Products built around it do not all work that way. crypto.news has reported that Base MCP, which connects into tools such as ChatGPT, Claude, Codex, and Cursor, still leaves final approval with the account holder. Integrated DeFi actions in that flow still ask for confirmation in a separate window. “The agent spent it” and “nobody was watching” are different sentences.
These studies measure x402 on Base. They do not measure AI commerce in general, and they do not measure card or bank spend. Chainalysis’s June post said the protocol also runs on chains such as Solana and Polygon, and that Base had the most active deployment. That is a comparison among the chains it looked at. It is not a census of every rail. Agent activity, in its method, means wallets talking directly to the protocol. A wallet can be a human script, a trading bot, a promotion, or a company paying its own infrastructure. Calling it an agent is a classification. It is closer to calling a spreadsheet a department than to meeting a new customer.
Three facts belong in the budget conversation before the charts do. The person or company that provisions the keys still owns the money. A five-cent request once a second is $180 an hour. At ten cents, it is $360. That is arithmetic. Chainalysis did not observe that pattern. And x402 is one way to settle a cheap call. Monthly or account billing can price the same API without a payment on every request. For AI for business, a named budget owner, a vendor allowlist, and a cap are what separate an API call from a bad hour. The label does not change the problem if the wallet is buying data for a sales team rather than inference for a product. AI for sales is not an exemption from a refill with a human name on it.
The mint that padded the early count
Early x402 was, in Chainalysis’s telling, mostly fads. The clearest was PING, a pay-to-mint experiment. A user queried a URL, got an HTTP 402 response, and paid 1 USDC to mint a token. The June post said transactions jumped more than 10, 000% in a single week, and that PING alone did more than 150, 000 transactions in its first month. Base’s near-zero gas fees made the loop cheap to repeat.
A one-dollar mint is not a micropayment. Set against a later pile of more than 100 million transactions, 150, 000 is small. It can still distort early dollar volume, and it proves a narrower point: an onchain payment can be perfectly real while the thing being sold is a speculative mint. Chainalysis treated PING as a high-concurrency load test. It did not treat PING as a durable information market. By the first quarter of 2026, it said speculative activity had cooled and growth had moderated. It never published a date when useful purchases overtook promotions.
The counters do not match, and they should not be blended. Chainalysis’s June post put cumulative x402 transactions on Base at well over 100 million through the first quarter of 2026. crypto.news has reported earlier figures cited by Base of roughly 165 million settled transactions and more than 480, 000 agents. Those Base figures are not in the Chainalysis posts. A total of more than 138 million has circulated in coverage of the October study. It is not stated in the posts that publish the size sample and the cohort charts. Different populations, different dates, different counters.
A protocol interaction is not a distinct autonomous agent. The 480, 000 figure has the same problem until someone shows how an “agent” was defined, and that the same wallet was not counted twice.
Most payments are tiny. Most of the money is not.
June and October are not the same measurement, and both can be true. In the June post, transfers of $1 or more were 49% of volume in early 2025 and 95% by early 2026. The band from 10 cents to $1 fell from 46% to 4%. crypto.news has read that 95% as value moved. Chainalysis’s own language was that economic weight shifted toward larger transfers, even while a large number of sub-cent transactions continued. The June post does not say the count shifted to $1 payments.
The early-2025 mix describes payments that existed. It does not describe a market already underway. Chainalysis separately said the count was near zero in mid-2025. Still tiny as a business. Once any volume was there, it was not too small to sort by size.
Hold that next to October. Over the 16 months ending September 2026, more than 87% of payments were still $0.0001 to $0.10, and the most common size was one to ten cents. Most checks are tiny. Most dollars, by early 2026, sat in fewer larger transfers. That pairing is why a transaction count is not seller revenue.
Net receipts after refunds, fees, and moves to affiliated wallets would be closer to revenue. Those nets are not in the published charts. One reason a summed transfer total can mislead is that the same funds can move more than once. That is a reason for caution. It is not a finding that it happened in this sample.
What the wallets called is more interesting than the raw count, and still incomplete. Chainalysis says it tracked millions of transactions from AI wallets into code review, financial intelligence, and data and search APIs. Model inference (paying a model to produce a response) is one of the largest categories. Other tagged uses include payment rails, agent frameworks, agentic tooling, gaming, and token issuance. Gambling-related services are on that list too. If a company wallet can reach them, that is a compliance exposure. It does not belong in a footnote.
The October post says the mix matured toward information services, with demand showing up in machine-readable data. That is a classification claim. There is no category-by-category volume and value series on a common timeline, so there is no dated crossover from fad to utility. The payment is onchain. The result usually is not. A transfer does not prove the data was accurate, the computation finished, or the answer was worth the dime.
One vendor can look like a country
Concentration matters more than passports. Among payments Chainalysis could place, U.S.-funded wallets were 88.1% of model-inference transactions and 30% of agent-to-service transactions. Those shares are not of all x402. The share that could not be geolocated was not published.
Chinese-linked wallets made nine in ten transactions in the tagged security and code-analysis category. Chainalysis says almost all of that category went to AgentLISA, a service that audits Solidity contracts with a language model. One recipient is not a national market.
“AI nationalism” is Chainalysis’s hypothesis. It is not a field in the transfer. Chinese operators may prefer Chinese open-weight models, including DeepSeek and Qwen, which Chainalysis describes as free to use, while U.S. wallets pay metered models such as ChatGPT. Plausible. The payments do not prove it. Location tags are estimates from onchain signals. They are not proof of residency.
Refilled wallets are not revenue
Refills, through Q2 2026 only
Among wallets that sent on at least two separate days, the share receiving a second deposit within 30 days rose from 70.6% for the third-quarter 2025 cohort to 89.3% for the second-quarter 2026 cohort. One-day experiments are out. This is not a share of every wallet that ever touched x402. Chainalysis calls the second quarter of 2026 the latest group whose 30-day window had fully closed.
Speed, a later chart through Q3 2026
On a later chart, the median gap from seed funding to first spend fell from 73 days for late-2025 wallets to 18 minutes by the third quarter of 2026. The share spending within an hour of funding rose from 4% in the third quarter of 2025 to 67% in the third quarter of 2026. The refill series stops so a 30-day window can close. The speed series runs through the next quarter. The two are related. They are not the same observation.
Chainalysis reads the refill the way a subscription renewal reads: somebody found enough utility to fund the wallet again, and newer wallets look built for a job already queued. That is a fair reading of the timing. It is not profit, audited seller revenue, or proof the recipient was an outside vendor rather than the operator’s own infrastructure. It also does not say whether the refill was an evaluated renewal or a standing transfer. What these charts do not show is persistence past 30 days (active days at 90, how many different vendors got paid, or how much stablecoin was actually spent).
June described a separate tester-to-payer conversion (a wallet that paid itself, then paid others in the same month) as improving fourfold in six months, and weekly retention as volatile but trending up. Neither rate was published. Leave both off the slide.
The budget still has a human name on it
Tejwani of Coinbase Ventures put the bull case in one sentence, quoted in Chainalysis’s October post:
AI agents are becoming economic actors where every agent is effectively a new “user” of the blockchain… We’re especially excited about projects building on the x402 protocol, which Coinbase incubated…
Coinbase Ventures is an interested party, and “user” is doing more work than the data can support. A new wallet is not a new economic actor. The budget owner is still the person or company that provisions the keys, sets permissions, and sends the next deposit. Autonomy of the click is not autonomy of the treasury.
Circle, as quoted by Chainalysis, draws a cleaner line: “card rails for consumer-prompted agent payments, stablecoin rails for machine-to-machine settlements that are too small for card economics.” Chainalysis argues that cards fit poorly under a dollar, and especially under a penny, while stablecoins that cost less than a cent to settle make API-sized charges viable. That settlement-cost claim is Chainalysis’s. It is not an independent measurement in this study.
Cheap settlement does not retire the rest of the stack: funding, key management, sanctions and fraud exposure, reconciliation, failed responses, and disputes. Those controls vary by product. These studies do not report the terms.
Announcements have widened around the rail. crypto.news has reported Stripe support for x402 on Base, AWS support through Amazon Bedrock AgentCore Payments, and other adopters. None of that reporting assigns any share of the measured transactions to Amazon or to any other named enterprise. Forecasts for the broader agentic-AI market are a different subject. They are not x402 volume. They are not what an API seller kept after refunds.
A machine-to-machine market a CFO should respect would show independent vendors receiving repeat payments from many separately funded agents, with category volume and value over comparable periods, and with concentration low enough that one mint or one audit service cannot carry the story. Until that shows up, this is activity worth watching. It is not a revenue line worth booking.
Questions worth asking before the next top-up
- Did AI agents make 138 million payments? No. That figure is not in the Chainalysis posts that publish the size sample, and no count there equals 138 million distinct agents. The on-record Base total is well over 100 million x402 transactions through the first quarter of 2026, which should not be averaged with the separate Base figures crypto.news has reported.
- What did the wallets buy? Tagged categories include model inference, code review, financial intelligence, data and search APIs, agent tools, payment rails, games, token issuance, and gambling-related services. There is no full invoice for every transaction, and a completed payment does not prove the result was useful.
- Were the payments economically meaningful? In the 16 months ending September 2026, more than 87% of payments were between $0.0001 and $0.10. By early 2026, Chainalysis put $1-plus transfers at 95% of volume, which crypto.news read as value moved. That is not a published revenue total.
- Does a second deposit mean customers came back? Among wallets that sent on at least two days, 30-day refill rates rose from 70.6% to 89.3%, and the median time from seed funding to first spend fell from 73 days to 18 minutes. That is repeat funding under a stated filter. It is not profit, and it is not proof the recipient was an outside vendor.
- Why do U.S. wallets look so dominant? They aren’t, except in one aisle. They were 88.1% of geolocated model-inference transactions and 30% of geolocated agent-to-service transactions. In security and code analysis, Chinese-linked wallets made nine in ten transactions, and almost all of that category went to AgentLISA.
- Should a company let an agent pay with no one in the loop? The x402 design allows unattended per-request settlement, but crypto.news has reported that at least one Coinbase agent surface still keeps final approval with the account holder. Unattended spend is a product choice. It is not a fact that arrives with the protocol.
- Before the next top-up, can you name the budget owner, the allowlist, the per-hour cap, and whether the recipient is an outside vendor? If you cannot, you are funding a demo. You are not buying a service. Those four answers are the difference between an API call and an hour that costs $180.
AI agents that can pay are a real shift in how software buys compute and data. The treasury that stands behind them is still a human decision, and it should be treated like one.