More than 138 million x402 payments were recorded on Base in Chainalysis’s 2026 North America study. The count is large; the typical payment is tiny. Some software buys data, code checks and
More than 138 million x402 payments were recorded on Base in Chainalysis’s 2026 North America study. The count is large; the typical payment is tiny. Some software buys data, code checks and access to AI models, while earlier activity included a speculative token mint that generated transactions without proving a durable service economy. The more revealing figures concern wallets that return, refill and pay named categories of providers.
SummaryThe October 14 Chainalysis study identifies categories that give the 138 million count an economic context: AI model inference, code analysis, search and data feeds, payments infrastructure, agent tools, gaming and token issuance. It also says the earliest x402 activity included short lived meme coin interest. An onchain payment can be real while the service it purchases is a speculative mint rather than an enduring business expense.
Earlier crypto.news coverage of the 100 million threshold explained the payment protocol. The new report supplies evidence about what recipients sell and whether wallets keep funding their purchases. Those two tests matter more than treating one transaction as one autonomous customer.
BIG: Hedera gives AI agents their own accounts to manage funds Agent Accounts let autonomous systems hold, spend and settle money around the clock, while users can set spending limits and restrict approved recipients. pic.twitter.com/51o879zq8x
— crypto.news (@cryptodotnews) October 7, 2026
Chainalysis describes x402 as a protocol Coinbase published in 2025 that lets software pay for a web resource when a server requests payment. A wallet can transfer a stablecoin, receive the resource and continue its task without a person authorizing that individual charge. The observed activity in this report is x402 on Base. It is not a count of all AI payments on every chain, nor a count of all commerce performed by AI systems through cards and ordinary bank accounts.
The 138 million total is cumulative to the end of the study’s 2026 period, which the report defines as the 12 months ending July 1, 2026. Chainalysis extended its examination of payment sizes through September. Those are two observation windows. Applying the 87% micropayment share from the extended sample mechanically to all 138 million transactions would create a false precision unless the underlying datasets and definitions match.
The report describes more than 87% of payments in the extended window as falling between one hundredth of a cent and ten cents. That range spans a factor of 1,000. Multiplying a transaction count by a guessed average inside it would not yield a defensible total dollar value. The many small transfers explain why x402 is interesting for API calls; they do not show the revenue of agent service providers or fee income for Base.
Transactions are actions, not people. One developer can fund several wallets. One wallet can call the same API hundreds of times. A service can create a large number of low value transactions as part of a promotional campaign. The count is strongest as a measure of protocol use and weakest as a measure of independent paying customers. Identifying repeated, unrelated counterparties and sustained revenue would strengthen the latter claim.
Chainalysis says it followed millions of transactions from agent wallets to providers of code review, financial intelligence, data and search APIs. Model inference, the use of a paid AI model to generate a response, is another substantial category. These are plausible purchases for software that has been assigned a task and needs an external answer or current information to complete it.
An agent inspecting a smart contract might pay an analysis service for a security review. A financial tool could buy a market data query. A travel or research assistant might purchase access to a database it could not otherwise read. These examples illustrate categories described in the report, not verified identities for every wallet. Chainalysis’s classification depends on knowing recipient services and tagging wallets; activity beyond that observed population may differ.
The report also tracks payments to agent frameworks, tool catalogues, payment rails, games, token issuers and gambling related services. Calling all of those transactions AI productivity spending would erase differences. A bot buying another model’s inference can be part of a recurring software workflow. A wallet paying to mint a meme coin can be a speculative game. Both use the same payment protocol and may look similar at the transaction layer.
Crypto.news’s earlier account of Amazon’s agent payment integration showed a corporate setting for programmatic USDC spending with controls. It shows that a developer can design a real payment workflow; it does not identify which of the 138 million transactions originated from Amazon services. A source should not be used to assign another dataset’s wallets to a named enterprise without a direct link.
Chainalysis’s June study of x402 activity said PING, a pay to mint experiment, helped drive an early surge. Users called an endpoint, received a payment request and paid to mint a token. The protocol processed the calls as designed. Demand for PING, however, did not prove that developers would keep paying for useful data or inference after the speculation faded.
The October report says the mix is maturing toward information services. That is a claim with an observable countercheck: do recipient addresses tied to code review, data and model services continue to receive payments after one off promotions end? Do the same customers replenish their budgets and use the service over multiple reporting periods? Without category level volume and value, a change in narrative can run ahead of the transaction mix.
Another crypto.news analysis of XRPL agent payments converted a large payment count into about $280 in network fees at the then stated charge. Base has a different fee structure and its 138 million number cannot be multiplied by XRPL’s fee. The comparison shows why count, value sent, service revenue and network fees must be four columns, not one.
An x402 request can involve a fraction of a cent, which makes card economics difficult. Yet low settlement cost is only one piece of total cost. Someone must fund the wallet, manage keys and permissions, handle sanctions and fraud exposure, reconcile receipts and resolve a failed API response. A cheaper marginal transfer helps only if those fixed and operational costs can be spread over enough useful work.
JUST IN: Sui partners with Alibaba Cloud for AI agent payments The partnership will enable stablecoin payments on a per-call basis, allowing AI agents to pay for services directly through the $SUI network. pic.twitter.com/atGZca46KU
— crypto.news (@cryptodotnews) October 7, 2026
Chainalysis groups wallets by the quarter in which they first received funding. Among qualifying wallets, the share receiving a second deposit within 30 days rose from 70.6% for the Q3 2025 cohort to 89.3% for the Q2 2026 cohort. The researchers required wallets to send on at least two separate days, excluding one day experiments from that calculation. A refill is evidence that an operator funded continued activity, though it does not prove a profitable business or independent user.
The denominator is crucial. It is not 89.3% of all wallets that ever touched x402. It is a selected cohort meeting the activity rule and observed for a complete 30 day window. Chainalysis says Q2 2026 was the newest cohort with such a window. Promoting that figure to all agent wallets would exaggerate retention. Even for qualifying wallets, a second funding could be small or linked to a test.
You might also like: Sui hits 40.6 million TPS in offchain test simulating AI agent activity
The funding to first payment interval moved from a median 73 days for late 2025 wallets to 18 minutes by Q3 2026. The share spending within one hour of funding rose from 4% in Q3 2025 to 67% in Q3 2026, according to the same research. That pattern fits wallets provisioned for an immediate job rather than wallets kept idle before someone experiments. The report does not show the customer’s complete software task, so intent is an inference from timing.
A cohort chart over more months would test the persistence of the improvement. If newly funded wallets continue to refill but the services they buy are largely affiliated with their own operator, it would show operational reuse without a broad market. If independent vendors receive repeat payments from many separately funded agents, the case for a functioning machine to machine market becomes stronger.
The report attributes 30% of the x402 agent to service transactions it can geolocate to U.S. funded wallets. For AI model inference specifically, that share is 88.1%. The different denominators explain why both percentages can be true. U.S. wallets dominated paid model access within a location tagged subset, not 88.1% of every x402 payment on Base.
Chinese linked wallets accounted for about nine in ten transactions in the tagged security and code analysis category, with most of that category flowing to AgentLISA, Chainalysis says. One recipient’s prominence can shape a country share. It does not show that 90% of all Chinese agent spending is on security or that every such wallet belongs to a Chinese resident. Geographic attribution of an onchain address is an estimate based on available signals.
The researchers suggest different model economics might explain why U.S. wallets pay for inference more often while Chinese operators use open weight alternatives. That is a proposed explanation, not a transaction field that identifies a model’s price or a customer’s motivation. An operator could pay a U.S. service from another location or route funding through an exchange that complicates location tagging.
Crypto.news’s coverage of USDC in agent payments described the strong role of dollar denominated settlement. A share attributed to a particular dataset and date should not be pasted onto every transaction in the new Chainalysis sample. The useful connection is structural: a low value automated purchase needs a unit that a seller can price without a volatile token changing value between request and fulfillment.
An agent may choose when and where to buy an API result, but a human or organization provisions its wallet, defines permissions and replenishes funds. A software process that pays without a human click for each request is not an entity with independent financial responsibility. A business deploying agents needs limits by vendor, amount, time and type of service, plus a way to revoke compromised credentials.
The payment itself is visible onchain, while the result bought may remain offchain. A blockchain record can show that a token moved to a provider; it generally cannot prove the returned data was accurate or that a model answer was useful. A service could charge for an unavailable endpoint or a malicious agent could buy irrelevant information. Receipt and dispute handling must connect the token transfer to the API response.
The Graph’s x402 data gateway coverage illustrates why metered API access appeals to providers. Charging per query can bring in customers unwilling to buy a subscription. It also exposes the seller to request spikes and the buyer to a runaway loop that calls the same endpoint repeatedly. Spending caps, logging and anomaly alerts can limit that risk. A protocol that settles a transfer cannot decide whether the agent needed the answer.
Attribution creates another challenge. A wallet might be called an agent because it used x402, but the payer could be a human script, a trading bot or a promotional service. Chainalysis’s classification measures observable interactions, not the private model and decision process behind every transaction. The honest label for the 138 million is x402 payments on Base associated with a protocol commonly used by agents, with the categories and limitations the researchers disclose.
Coinbase and Circle have a credible claim that card networks are poorly suited to machine sized charges. A one cent data request cannot bear the same fixed economics as an ordinary card purchase. Chainalysis’s evidence that many qualifying wallets refill and spend quickly supports an emerging use case beyond one day demos. Corporate integrations add another path for the rails to reach real software budgets.
The evidence still stops short of a complete revenue account. The report supplies a large cumulative transaction count, a broad micropayment distribution and selected wallet behaviors. It does not give an audited total of paid model inference revenue, net earnings for API sellers or the number of unrelated firms funding those wallets. Even a large total value could contain repeated transfers among related accounts; even a small total value might pay for useful, inexpensive work. Recipient level demand is the missing column.
The next useful report would identify category volumes and dollar values for comparable periods, unique paying organizations when attribution permits, recurring recipients, refunds and the share of transactions linked to speculation. It should keep unknown activity visible instead of assigning every transaction a commercial purpose. The Chainalysis cohort figures provide a reason to study the market closely; they do not close the case on its economic size.
An onchain payment is observable, but deciding whether a wallet belongs to an AI agent is a classification exercise. X402 marks a particular exchange between a client and service. A human can script those calls, and a model can initiate them after a developer authorizes a budget. Chainalysis’s count is valuable because it follows a defined protocol, yet it should not be relabeled as 138 million independent machine decisions. The number of unique payers, vendors and completed tasks is not contained in the headline count.
A useful audit would sample recipient contracts or accounts in each category, verify what service was provided and estimate the fraction of transactions that went to the same operator’s own infrastructure. A developer can legitimately route payments among components it controls to test an application; that activity shows a functioning rail but not external revenue. The public ledger can expose repeated addresses and amounts, while proving who controls each wallet often requires offchain evidence.
The cumulative count also mixes periods of very different behavior. Chainalysis’s June report described an early speculative spike; the October report describes rising use of information services. Comparing all time volume to a current month service mix can make the new use appear larger than it is. Monthly categories with the same tagging rules would show when useful purchases overtook promotions, if that crossover has happened. Without such a series, the direction is supported by the researcher’s qualitative classification, not a published crossover date.
Stablecoin movement can be grossed up by repeated use of the same funds. A wallet funded with $10 might make hundreds of tiny queries and replenish later. Summing transfers can show how often paid calls occurred, but it does not mean $10 of new capital arrived each time. A provider’s net receipts after refunds, fees and transfers back to affiliated wallets would be the revenue counterpart. Onchain payment volume and revenue should not be treated as synonyms.
The 73 day to 18 minute change suggests wallets are created or funded closer to the moment a task needs a paid service. That is an operational change. It could arise because developers have improved integration, because funding tools automatically provision wallets, or because the mix of participants shifted toward operators with immediate demand. The report’s data shows the outcome and supports a purposeful use interpretation; it does not isolate which explanation caused it.
The share spending within an hour, 67% in the newer cohort, is particularly informative when read alongside the second funding statistic. A wallet that pays once quickly can still be an experiment. A wallet that pays on multiple days and receives another deposit within a month fits ongoing work better. Neither number alone identifies an end customer. Together they offer a more credible operational signal than a chart of total wallet creation.
A further test would follow each cohort for 90 days, tracking active days, vendor diversity and stablecoin spent. A second funding after a month may reflect a standing transfer set by the developer. Vendor diversity can show whether wallets buy across an ecosystem or only within one application. The same definitions should be retained between cohorts; changing the rule from two active days to one can make retention look stronger without a behavior change.
There is a useful limit to the subscription analogy. A human renewing a subscription often decides the service was worth paying for. An agent wallet may refill automatically under a budget, with evaluation happening later at the organization level. Thirty day refill is evidence that money returned to the wallet, not proof that the buyer calculated a positive return. Customer renewals at a named service would be a stronger business metric.
An organization can assign an agent a daily budget and approved vendors, but software may still make a series of bad purchases within that envelope. Payment policy must define who can change an allowance, how credentials are stored, which API response counts as fulfillment and when an account is frozen. The blockchain confirms that tokens moved, not whether the information was fresh or the purchased computation completed successfully.
Small payments can turn into large costs through frequency. An agent sending a five cent request once a second would spend $180 in an hour if it ran uninterrupted. At ten cents a request, the hourly figure is $360. These are hypothetical calculations of rate times amount, not Chainalysis observations. They illustrate why a low unit price still calls for rate limits, spending caps and alerting when an agent loops.
An agent could purchase from a malicious service that returns instructions instead of the requested data. Software accepting outside responses must separate paid content from authority over its wallet. A payment system’s approval list and a model’s instruction boundary are separate controls. A system that verifies an invoice but trusts every returned message can still be exploited without any payment protocol failure.
For compliance, the funder and recipient are more meaningful identities than the agent label. Chainalysis’s geographic findings depend on attributing funding sources, which is why unknown locations should remain outside national share claims. A business needs records linking the wallet, human budget owner and vendor account. If an agent buys a regulated service, an autonomous request does not remove the business’s own legal obligations.
Micropayments alter the economics of oversight. A company cannot manually review every five cent call to a model or API and still offer that service at five cents. It can review the funding source, approve the recipient category and set automatic thresholds that stop or flag abnormal behavior. An investigation follows a pattern or exception. Chainalysis makes a similar case for screening the setup and the destination, but implementation details vary by wallet provider and organization.
A wallet carrying a $20 allowance can complete many cheap requests, but the rules must specify how an agent chooses among providers. If a service changes its price from $0.02 to $0.20, a static allowance may still permit spending ten times as much per call. The agent should receive a price cap from its operator and fail closed when a vendor exceeds it. Again, the transfer can settle correctly while the purchasing decision is bad.
Receipts need a unit both the buyer and seller can audit. A stablecoin transaction hash proves that a payment happened; a service response identifier can show which query it paid for. Joining the two is necessary if a customer challenges a charge or if an API times out after accepting payment. A service with no usable retry or refund procedure can have impressive onchain volume and poor retention.
The same issue affects fraud measurement. A malicious actor could route a large number of cheap calls through a service it controls to inflate its apparent user base. The cost would be network fees, developer time and any loss in conversion, potentially far below the headline notional volume when funds circulate. A serious dashboard should report independently controlled buyers and recurring service revenues, not just transaction counts.
JUST IN: Coinbase is building a financial architecture for AI agents, Brian Armstrong announces The system is designed to give autonomous and multi-agent entities access to banking functions including algorithmic trading, programmatic debt financing and automated cross-border… pic.twitter.com/qM3Snni57D
— crypto.news (@cryptodotnews) September 30, 2026
Evidence of an industry would include service providers earning sustained revenue from unrelated agents, operators increasing budgets after evaluating results and multiple vendors competing on price and quality. Chainalysis’s new recipient categories and wallet cohorts move the discussion in that direction. They do not yet identify the end users paying the bills or the profit from the services sold.
The low value distribution makes the use case distinctive. A paid data request costing pennies may be impractical on legacy card rails, and a programmable wallet can make it in the middle of a software task. Yet the showed card world has account based billing and monthly aggregation, which can also support small unit prices without charging a card for every call. X402’s advantage should be compared to that alternative, not only to the absurd case of swiping a card separately for a fraction of a cent.
A provider can choose subscriptions, prepaid credits or x402 pay per use, and customers may prefer different models. Agent payments will gain a durable place if they reduce onboarding friction, reach a useful market of buyers and settle at acceptable total cost. The 138 million count shows that software is using the rail. The category and revenue data in future periods will show how much economic work rides on it.
Chainalysis counted more than 138 million x402 transactions on Base by the end of its 2026 reporting period. Protocol interactions do not prove each transaction came from a distinct autonomous agent.
Chainalysis identified model inference, data and search APIs, code analysis, financial intelligence, tools, games and token related services. It did not publish a full invoice for every transaction.
More than 87% in the report’s extended size sample were between $0.0001 and $0.10. That distribution cannot by itself reveal the total dollars paid.
It is a payment protocol that lets software respond to a web service’s payment request, commonly with a stablecoin, and receive the resource. Coinbase published it in 2025.
No. One wallet can make many payments, and a single organization can control many wallets. Independent customers require a different measurement.
Among qualifying wallet cohorts, second funding within 30 days rose from 70.6% to 89.3%. The measure excludes wallets that sent on only one day.
That percentage refers to geolocated x402 payments for AI model inference, not all agent transactions. It depends on Chainalysis’s wallet attribution and service classification.
No. Transaction and refill data show activity and repeat funding, while seller revenue and operating costs need separate evidence. This is educational analysis, not investment advice.
Disclaimer: This article is for information and educational purposes only and does not constitute financial or investment advice. Figures reflect regulatory filings and reporting available at the time of writing and change with each disclosure. Nothing here is a recommendation to buy, sell, or hold any security or asset. Always do your own research. Information is accurate as of October 15, 2026.
Read more: Ledger changes wallet sales rules after crypto thefts
