AI Agents Are Doing 29 Payments a Second — at 32 Cents Each

The fastest-growing payment network of 2026 moves less money in a month than a mid-tier regional bank moves before lunch. It processed 75 million transactions in 30 days — totaling just $24 million. Average ticket: 32 cents. And this week, Visa, Mastercard and Ripple all paid for seats at its table.

The network is x402, a protocol that lets AI agents pay each other in stablecoins, and it has become the clearest demonstration yet that agentic payments are not a keynote-slide fantasy. They are live traffic — roughly 29 transactions every second, flowing between some 94,000 buyers and 22,000 sellers, most of them software.

The dollar figures are trivial. That is exactly why the incumbents showed up.

A 1990s error code becomes a payment rail

The mechanism is almost comically simple. HTTP has carried a status code — 402, "Payment Required" — since the 1990s, reserved for a future nobody built. x402 finally activates it. A server that wants payment answers a request with a 402 and a price. The agent signs a small stablecoin transfer, usually USDC, resends the request with payment attached, and receives the data. Seconds, no account, no card, no prior relationship between the two sides.

In July the Linux Foundation announced the operational launch of the x402 Foundation, with Coinbase's contribution of the protocol complete and 40 members signed on. The premier tier reads like a payments-industry seating chart: Visa, Mastercard, American Express, Stripe, Adyen, Fiserv, Shopify, Google, Amazon Web Services, Cloudflare, Circle, plus the Solana and Stellar foundations. Ripple joined the same week, confirming the XRP Ledger already supports the standard.

When every company whose business model is payment intermediation joins a foundation for disintermediated payments, one of two things is true. Either they are there to smother it — the classic embrace-and-extinguish play — or they have concluded the traffic is coming regardless and would rather stand in the toll plaza than beside it. Their own research suggests the second.

Visa's report reads like a confession

On July 16, Visa and analytics firm Artemis published a joint report arguing that stablecoins will become the preferred payment rail for AI-agent micropayments. Think about who is saying this. The company that earns a percentage of every card swipe is forecasting a payments category in which cards are structurally uncompetitive.

They are right, and the arithmetic explains why. A 32-cent transaction dies inside interchange economics — the fixed fees exceed the payment. Card networks were built for humans making dozens of purchases a month, not agents making thousands of calls an hour. The only rails where a sub-dollar payment clears profitably are the ones x402 uses.

The volume backs the thesis. On Base, Coinbase's layer-2 network, agentic payments crossed 20 million transfers in a single 90-day window, with about 169 million total agentic transactions processed to date. Coinbase itself shipped tooling in June that lets agents manage trading and payments directly. Between early June and July, MetaMask, OKX and BNB Chain all launched agent-focused wallet and identity infrastructure. The picks-and-shovels layer is being poured in real time.

What agents actually buy for 32 cents

Skepticism about these numbers usually starts here: what commerce is really happening at 32 cents a pop? Mostly machine-to-machine mundanity — an agent buying a single API call, a data query, a snippet of compute, a scrape of a paywalled page. One inference chain might trigger a dozen such purchases. No human would tolerate a checkout flow for any of them; that is the entire reason the category could not exist before.

The honest caveat is that some of this volume is subsidized experimentation and developer noise, the way early Lightning Network traffic was. Fine. Every payment rail's first million transactions look like a toy. The Visa network's did too, in 1958, when it was a Bank of America mailer in Fresno. The relevant question is not whether today's $24 million matters — it doesn't — but whether the transaction shape it proves out becomes the default interface between AI services. Forty payment giants just bet that it will.

The regulation writers didn't see this customer coming

There is a regulatory wrinkle hiding inside the timing, and it deserves more attention than it is getting. The stablecoins these agents spend — overwhelmingly USDC — are the same instruments six U.S. federal agencies are racing to finalize rules for under the GENIUS Act this very week. Those rules were drafted with a mental model of human payment flows: remittances, merchant checkout, treasury management.

An economy of autonomous software making 29 purchases a second at 32 cents apiece fits none of those templates. Know-your-customer frameworks assume a customer who can be known. Transaction monitoring thresholds assume transactions large enough to monitor. An agent that splits its activity across ten thousand micro-payments per day is not evading those systems; it is simply invisible to them by design of scale.

Regulators will eventually notice that the fastest-growing user of regulated dollar tokens has no name, no address, and no jurisdiction beyond wherever its server happens to run. When they do, the fight over agent identity — who must attest to what an agent is and whose agent it is — will make the current stablecoin rulemaking look consensual by comparison. The x402 Foundation's membership list suggests the industry already knows this. You do not invite Cloudflare, Google and AWS into a payments body unless identity infrastructure is on the roadmap.

The optimistic read: machine payments could end up cleaner than human ones, since every agent transaction is programmatic, logged, and auditable from birth. Cash never left a trail. Agents can't help leaving one.

The permission problem nobody has solved

The hard problem is not throughput. It is authority. JPMorgan's Marianne Lake voiced the incumbents' core objection: consumers may not be ready for AI handling their money, with unresolved questions of trust, security and liability. She is asking the right question. When an agent misbuys — wrong dataset, wrong quantity, a hallucinated subscription — who eats the loss? The user who delegated? The model developer? The protocol? Card networks spent fifty years building chargeback machinery for disputes between humans. Nothing equivalent exists for disputes between software.

Spending limits and session keys are the current answers, and they are the payment-security equivalent of a padlock on a bank vault. Expect the first agent-payment scandal — an agent drained by a prompt-injected merchant, most likely — within the year. Expect it to be survivable, and to produce the insurance and attestation layer that card-dispute machinery took decades to grow. The scandal is a rite of passage, not a verdict.

There is a quieter consumer-side shift coming too. A person running three agents will generate thousands of micro-debits a month across chains — a transaction history no bank statement was designed to express. Portfolio and wallet trackers like The Crypto App already aggregate balances across chains for humans; the next version of that job is summarizing what your software spent on your behalf, and whether you'd have approved.

The stakes are the checkout page itself

Follow the logic one step further and the reason for the land rush becomes obvious. If agents do the buying, the checkout page — the web's most valuable real estate for thirty years — stops being seen by humans. Branding, impulse placement, loyalty programs, the entire psychology of conversion: all of it evaporates when the customer is a script optimizing for price and latency.

That is why Visa is in the foundation rather than fighting it, why Stripe and Shopify are members, why Google and AWS bought in. Nobody can afford to let a payment standard for the machine economy be written without them. The $24 million is not the prize. The prize is authorship of the rails that carry the next trillion, in payments too small for any human to notice and too numerous for any incumbent to ignore.

Twelve months from now, the interesting metric will not be x402's transaction count. It will be the first month an AI agent's spending decisions move a real company's revenue line — and the first earnings call where an executive has to explain that their newest major customer segment has no pulse.