I’ve been watching the numbers roll in for months now, and still the scale catches me off guard. More than 138 million crypto payments moved through a single protocol on Base, almost all of them triggered by software rather than people clicking buttons. The typical transfer sits somewhere between a fraction of a cent and a few cents. That kind of volume forces a simple question: what exactly were these agents buying, and does any of it point to lasting demand?
Inside The 138 Million Count
The figure comes from a detailed review that closed its main window at the end of a twelve-month stretch ending in early July. Researchers later stretched the payment-size look through September. Two different observation periods sit side by side, so treating every statistic as if it covers the exact same set of transfers would overstate precision.
The protocol itself lets a piece of software answer a web server’s payment request, send a stablecoin, and keep working. No human has to approve each tiny charge. Almost everything counted here happened on Base. It is not a global total of every AI-related payment across every chain, and it certainly does not capture ordinary card or bank transfers made by the same systems.
More than 87 percent of the payments examined in the extended sample fell between one-hundredth of a cent and ten cents. That range covers three orders of magnitude. Anyone multiplying the full 138 million by a guessed average inside that band ends up with a number that looks precise yet rests on shaky ground. The sheer quantity of micro-transfers explains why the protocol feels useful for API-style calls. It does not by itself reveal how much revenue the service providers actually kept.
Transactions are actions, not customers. One developer can spin up several wallets. One wallet can hammer the same endpoint hundreds of times. A promotional campaign can flood the ledger with low-value calls that never turn into ongoing business. The count works best as a measure of protocol usage. It works least well as a headcount of independent paying agents.
What The Agents Actually Purchased
The researchers followed millions of transfers from agent wallets to named categories of providers. Model inference sits near the top. An agent that needs a quick answer from a paid AI model pays for the generation, receives the output, and continues its task. Code analysis tools, financial intelligence feeds, general search and data APIs all appear repeatedly. These purchases make practical sense for software handed a concrete job that requires fresh external information.
Picture an agent reviewing a smart contract. It might send a few cents to a security-analysis service, get the report, and decide whether the code is safe enough to interact with. A research-style agent could buy a short window into a proprietary database. A market-watching tool might pull a priced data point it cannot obtain for free. None of these examples claim to identify every single wallet in the set. Classification depends on knowing the recipient services and tagging the sending wallets. Activity outside that observed group may look different.
Payments also reached agent frameworks, tool catalogues, payment infrastructure, games, token issuers and even gambling-related endpoints. Lumping every one of those transfers under “AI productivity spending” blurs important distinctions. An agent buying another model’s inference can form part of a steady software workflow. A wallet paying to mint a short-lived meme token is closer to a speculative game. Both use the same payment rail and can look identical at the raw transaction layer.
An onchain payment can be real while the service it purchases is a speculative mint rather than an enduring business expense.
Early activity included a noticeable burst of pay-to-mint interest. Users called an endpoint, received a payment request, and paid to create a token. The protocol handled those calls exactly as designed. Demand for that particular mint, however, did not prove that developers would keep paying for useful data or inference once the speculative wave passed. Later reporting suggests the mix has tilted toward information services. The real test is whether addresses tied to code review, data and model endpoints continue receiving payments after one-off promotions fade, and whether the same customers keep topping up their budgets across multiple periods.
Wallet Refills Offer A Stronger Signal
Raw transaction counts can mislead. Refill behavior is harder to dismiss. Researchers grouped wallets by the quarter they first received funding. Among those that met a simple activity filter—sending on at least two separate days—the share that received a second deposit within thirty days climbed from roughly 70 percent in an earlier cohort to nearly 90 percent in a later one. A refill shows that someone decided the wallet still needed money. It does not prove the activity generated profit or that the operator is an independent third party.
The denominator matters. That high percentage applies only to the filtered cohort observed for a full thirty-day window. It is not 90 percent of every wallet that ever touched the protocol. Even inside the qualifying group, a second funding could be tiny or tied to further testing. Still, the upward trend is hard to ignore.
The time from funding to first payment shrank dramatically. Median lag dropped from more than two months for late-period wallets in one year to roughly eighteen minutes by the following mid-year. The share of wallets that spent within one hour of receiving funds rose from a low single-digit percentage to around two-thirds. That pattern fits wallets provisioned for an immediate job far better than wallets left idle for weeks while someone experiments.
I’ve found that timing data often reveals more about intent than pure volume. A wallet funded and spent almost immediately suggests the operator already knew what the agent needed. A wallet that sits for weeks before making a single call feels more like curiosity. Neither pattern alone identifies an end customer, yet together they paint a clearer operational picture.
Geography And Category Differences
About thirty percent of the geolocated agent-to-service transactions were linked to U.S.-funded wallets. For AI model inference specifically, the share climbed above 88 percent inside the location-tagged subset. Different denominators explain why both figures can be true at once. U.S. wallets dominated paid model access within the tagged sample; they did not account for 88 percent of every payment on the chain.
In the security and code-analysis category the picture flipped. Wallets linked to Chinese funding sources made up the large majority of tagged activity, with one prominent recipient absorbing most of that traffic. A single service’s popularity can shape an entire country share. It does not prove that nine out of ten Chinese agent payments overall go to security tools, nor that every such wallet belongs to a resident of that country. Geographic attribution of an onchain address remains an estimate built on available signals.
One possible explanation offered by the researchers is that model pricing and availability differ by region. Operators in some places may lean toward open-weight alternatives while others pay for hosted inference. That remains a hypothesis rather than a field visible in the transaction data. An operator could also fund a wallet from one jurisdiction and route traffic through exchanges that obscure the original location.
Autonomy Still Has A Budget Owner
An agent may decide the precise moment and target of a purchase, yet a human or organization still provisions the wallet, sets the permissions and decides when to add more funds. Software that pays without a human click for each request is not an independent financial actor. Any serious deployment needs hard limits by vendor, amount, time window and service type, plus a reliable way to revoke compromised credentials.
The payment itself is visible on the ledger. The result purchased often remains off-chain. A blockchain record can confirm that tokens moved to a provider. It generally cannot prove the returned data was accurate or that a model answer proved useful. A service could charge for an unavailable endpoint. A poorly designed agent could buy irrelevant information in a loop. Receipt handling and dispute resolution have to connect the token transfer back to the actual API response.
Perhaps the most interesting operational risk sits in frequency rather than unit price. An agent sending a five-cent request once per second would burn through hundreds of dollars in an hour if left unchecked. At ten cents the hourly figure climbs higher still. These are simple rate-times-amount illustrations, not observed totals. They show why even a low unit price demands rate limits, spending caps and anomaly alerts.
Software that accepts outside responses must also keep paid content separate from authority over its own wallet. A malicious service could return instructions instead of the requested data. Approval lists for payments and instruction boundaries for the model are separate controls. A system that verifies an invoice yet trusts every returned message can still be compromised without any failure in the payment protocol itself.
Why Micropayments Change Oversight Economics
When each call costs only a few cents, manual review of every transaction becomes impractical. Companies instead review the funding source, approve recipient categories and set automatic thresholds that stop or flag abnormal patterns. Investigation follows exceptions rather than individual calls. Screening the setup and the destination makes sense; the exact implementation varies by wallet provider and organization.
A wallet carrying a modest allowance can complete many cheap requests, yet the rules must still specify how the agent chooses among providers. If a service suddenly raises its price tenfold, a static daily budget may still permit the higher spend. The agent needs a price-cap instruction from its operator and should fail closed when a vendor exceeds it. The transfer can settle correctly while the purchasing decision remains poor.
Receipts need a unit both buyer and seller can audit. A stablecoin transaction hash proves the payment happened. A service response identifier can show which query it covered. Joining the two becomes essential when a customer challenges a charge or an API times out after accepting payment. A service lacking usable retry or refund procedures can post impressive onchain volume yet suffer weak retention.
Fraud measurement faces the same issue. A malicious actor could route a large number of cheap calls through a service it controls simply to inflate apparent usage. The cost would be network fees, developer time and any conversion loss—often far below the headline notional volume when funds circulate among related addresses. Serious dashboards should track independently controlled buyers and recurring service revenues rather than raw transaction counts alone.
Separating Industry From Demonstration
Evidence of a real 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. The newer recipient categories and wallet-cohort data move the conversation in that direction. They do not yet identify the end users writing the checks or the net profit earned by the services sold.
The low-value distribution is what makes the use case distinctive. A paid data request costing pennies sits awkwardly on traditional card rails with their fixed costs. A programmable wallet can complete the same purchase in the middle of a software task. Yet the existing account-based world already supports small unit prices through monthly aggregation and prepaid credits. The advantage of the new protocol should be measured against those alternatives, not only against the absurd case of swiping a physical card for every fraction of a cent.
Providers can still choose subscriptions, prepaid credits or pure pay-per-use. Customers may prefer different models depending on predictability and cash flow. Agent payments will earn a durable place if they reduce onboarding friction, reach a useful market of buyers, and settle at acceptable total cost once operational overhead is included. The 138 million figure shows that software is already using the rail. Category-level volume and revenue data over comparable future periods will show how much genuine economic work rides on it.
Practical Signals Worth Tracking
Several concrete indicators will clarify whether the activity is maturing or simply cycling through new forms of experimentation.
- Comparable-period transaction counts and dollar values using the same classification rules
- Share of payments and value flowing to inference, data, code review versus speculative token activity
- Repeat-funding rates measured with a clearly stated activity filter and completed observation window
- Concentration—whether a handful of services or wallets dominate each category
- Public product terms describing agent budgets, vendor permissions and handling of failed paid requests
A useful next step would also sample recipient contracts in each category, verify what service was actually delivered, and estimate how much traffic stayed inside the same operator’s own infrastructure. Routing payments among internal components is legitimate testing. It demonstrates a working rail without proving external revenue. The public ledger can expose repeated addresses and amounts; proving who controls each wallet usually requires off-chain evidence.
Cumulative counts mix periods of very different behavior. An early speculative spike and a later rise in information services can sit inside the same all-time total. Monthly category breakdowns with consistent tagging would reveal when useful purchases overtook promotions—if that crossover has already occurred. Without such a series, the direction of travel rests on qualitative classification rather than a published turning point.
The Missing Money Column
The strongest remaining gap is the absence of audited totals for paid model-inference revenue, net earnings for API sellers, or the number of unrelated firms funding those wallets. Even a large total value transferred could contain repeated movements among related accounts. Even a small total value might still pay for genuinely useful, inexpensive work. Recipient-level demand remains the missing column in the ledger.
Stablecoin movement can also be grossed up by repeated use of the same funds. A wallet funded with a modest amount can complete hundreds of tiny queries before needing a top-up. Summing the transfers shows how often paid calls occurred; it does not mean fresh capital arrived with every call. A provider’s net receipts after refunds, fees and transfers back to affiliated wallets would be the true revenue counterpart. Onchain payment volume and actual revenue should never be treated as synonyms.
In my experience the most convincing markets eventually produce named services with visible customer lists and recurring revenue that survive the loss of any single promotional campaign. The current data set has not yet reached that clarity. It has, however, moved well beyond pure demonstration. The combination of rising refill rates, shrinking time-to-first-spend, and a visible shift toward information services gives the market a reason for continued close attention.
What The Protocol Does And Does Not Prove
The protocol marks a particular style of exchange between client and service. A human can script those calls. A model can initiate them after a developer authorizes a budget. The large cumulative count is valuable precisely because it follows a defined technical standard. It should not be relabeled as 138 million independent machine decisions. The number of unique payers, vendors and completed useful tasks is not contained in the headline figure.
Attribution remains a classification exercise. A wallet may be labeled an agent because it used the protocol, yet the payer could be a human script, a trading bot or a promotional service. Observable interactions are measured; the private model and decision process behind every transaction stay invisible. The honest description of the 138 million is a set of protocol payments on Base associated with a mechanism commonly used by agents, complete with the categories and limitations the researchers themselves disclosed.
Card networks struggle with machine-sized charges. A one-cent data request cannot carry the same fixed economics as an ordinary consumer purchase. Evidence that many qualifying wallets refill and spend quickly supports an emerging use case that goes beyond one-day demos. Corporate integrations that embed programmatic stablecoin spending with explicit controls add another path for the rails to reach real software budgets.
The evidence still stops short of a complete revenue account. Large cumulative counts, broad micropayment distributions and selected wallet behaviors are all present. Audited totals for paid inference revenue, net seller earnings and the count of unrelated funding organizations are not. That gap does not erase the signal; it simply defines the next set of questions the market needs answered.
Looking Ahead Without Overclaiming
The next useful reports will keep unknown activity visible rather than forcing every transaction into a commercial narrative. They will publish category volumes and dollar values for matching periods, identify unique paying organizations when attribution allows, track recurring recipients, and separate refunds from gross volume. They will also retain the same activity filters across cohorts so that retention improvements can be trusted.
A further test would follow each funding cohort for ninety days, tracking active days, vendor diversity and total stablecoin spent. A second funding after a month may simply reflect a standing transfer set by the developer. Vendor diversity can show whether wallets buy across an ecosystem or stay inside one application. Consistent definitions between cohorts remain essential; changing the activity rule from two active days to one can make retention appear stronger without any real behavior change.
There is a useful limit to the subscription analogy. A human renewing a service often decides the product was worth paying for. An agent wallet may refill automatically under a pre-set 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 measured at the named service level would be a stronger business metric.
For compliance purposes the funder and the recipient remain more meaningful identities than the agent label. Geographic findings depend on attributing funding sources, which is why unknown locations should stay outside national share claims. A business needs records linking the wallet, the human budget owner and the vendor account. If an agent purchases a regulated service, the autonomous request does not remove the business’s own legal obligations.
The low settlement cost of a micropayment is only one piece of total cost. Someone still has to fund the wallet, manage keys and permissions, handle sanctions and fraud exposure, reconcile receipts and resolve a failed API response. Cheaper marginal transfers help only when those fixed and operational costs can be spread across enough useful work.
I’ve come to see the 138 million figure less as a finished score and more as an invitation. It proves the rail is being used at scale. The categories and retention patterns suggest the usage is evolving. Whether that evolution produces a durable machine-to-machine market or remains a mixture of genuine workflows and promotional noise will be settled by the next few reporting cycles. The data already on the table is interesting enough to keep watching closely.
What stands out most after working through the details is how ordinary the underlying need feels. Software needs information. Sometimes that information carries a price. When the price is measured in fractions of a cent and the buyer is itself code, traditional payment systems look clumsy. The protocol under discussion removes some of that clumsiness. The harder questions—who ultimately pays, what value is delivered, and whether the economics hold once operational overhead is counted—remain open. Those questions are worth answering carefully rather than rushing to declare a new economy fully formed.
The honest reading of the current evidence is measured optimism. Repeat funding is rising. Time from funding to first spend is collapsing. The mix of purchases appears to be shifting toward services that can support ongoing workflows. At the same time, concentration around a few recipients, the lingering presence of speculative activity, and the absence of clear revenue figures all counsel against treating the headline number as proof of a mature market. The next reports that publish matched category volumes and independent buyer counts will tell us far more than another cumulative total ever could.
Until those numbers arrive, the practical takeaway for anyone building or funding agents is straightforward. Treat the wallet as a controlled budget with explicit limits. Prefer services that provide clear response identifiers tied to each payment. Monitor frequency as carefully as unit price. And keep the distinction clear between internal testing traffic and genuine external revenue. The rail works. Whether the economy that rides on it becomes large and durable is still being written in real time.
