200 million x402 transactions, $15 million moved. What the numbers actually say.

Two numbers are being used to tell opposite stories about agentic payments. Hundreds of millions of x402 transactions since launch. Around $15 million in actual adjusted volume. One camp reads that as a protocol going vertical; the other reads it as a rounding error with a press office. Both are misreading it. Sub-dollar is not a phase x402 will grow out of — it is the shape of the traffic, and it dictates what has to sit on top.

What actually happened

On 14 July the Linux Foundation announced the operational launch of the x402 Foundation, with Coinbase's contribution of the protocol completed and open governance in place. Forty members. Premier seats held by Visa, Mastercard, American Express and Stripe. A technical steering committee, and a search underway for an executive director.

Three weeks later Cloudflare shipped both sides of the market on those rails: the Monetization Gateway for sellers and Cloudflare Wallets for buyers. The wire format question is, for practical purposes, settled. What is not settled is whether anything meaningful is moving across it.

The numbers, without the spin

• x402: roughly $15.0 million in adjusted volume across 109.6 million transactions since launching in May 2025. That is an average payment of about 14 cents.
• MPP (the Machine Payments Protocol from Stripe and Tempo, with Visa contributing): about $25,000 across roughly 115,000 transactions in its first weeks after a mid-March 2026 launch. Around 22 cents a payment.
• Headline counts near 200 million circulate because raw transaction counts include a great deal of testing, faucet traffic, and self-dealing between related parties. Adjusted volume is the honest figure, and it is the smaller one.

Fifteen million dollars over fifteen months is, in payments terms, nothing. Visa clears that in well under a second. Anyone selling x402 as a rail that already moves serious money is selling something.

Why sub-dollar is the point, not a stage

Look at what agents actually buy, in the examples Visa itself uses in its July 2026 agentic payments research: a research agent querying five financial data providers at a few cents each and keeping the most complete answer. A marketing agent paying ten image models five cents apiece to render the same product shot and picking the best one. That is not a checkout. It is parallel speculative buying — the agent purchases several options and discards most of them, because at those prices comparison is cheaper than deliberation.

No human commerce pattern looks like this. It only exists because the marginal cost of an extra attempt collapsed. And it means the transaction count and the dollar volume will keep diverging, not converge. As agent deployments grow, expect counts in the billions while average payment size drifts down, not up.

What that does to the control problem

Every assumption baked into corporate spend controls comes from a world of few, large, deliberate payments. Invert the shape and each one breaks.

Per-transaction review is arithmetically dead

A fleet of agents can generate more payment events in an afternoon than a finance team reviews in a year. Approval cannot be per payment. It has to be per policy — rules written once, evaluated automatically on every quote, with human escalation reserved for the small tail that actually warrants it.

The risk is aggregate, not per-item

Nobody is defrauded out of 14 cents. The failure mode is a retry loop that pays the same endpoint four hundred thousand times overnight, or forty agents each quietly inside their own cap. A per-transaction maximum — the control both Cloudflare Wallets and most wallet providers ship with — does not see either. Velocity limits, rolling windows, and org-wide budgets do. We wrote about that gap in Cloudflare gave agents a wallet.

Reconciliation becomes a data problem

A million-line ledger of sub-cent entries is not something a CFO reconciles by hand, and it is not something an accounting system was built to ingest. It needs to roll up by agent, by project, by client, by merchant, by day — before it ever reaches the books. Attribution has to be captured at authorisation time, because there is no reconstructing it later from a chain of hashes.

Counting transactions is not measuring anything

The same trap the ecosystem fell into with the 200 million figure will hit teams internally. Transaction count is a vanity metric. The questions worth instrumenting: what did this agent's task cost end to end, what fraction of purchased calls were actually used, which merchants deliver value per dollar, and how much spend is retry waste.

What to take from the standards news

Vendor-neutral governance under the Linux Foundation, with the card networks and Stripe inside the tent rather than building a rival, is the strongest possible signal that x402 is the long-term wire format. Standards bodies attract that membership when the fight is over, not while it is on. Meanwhile MPP's existence is a reminder that more than one protocol will carry machine payments, and that any control layer worth building has to be protocol-agnostic rather than welded to a single spec.

So: build against 402 now, ignore transaction-count headlines, and assume the volume curve is real but slow. The infrastructure being laid this quarter is not for $15 million of traffic. It is for the point, probably a year or two out, where a mid-sized company's agents make a hundred thousand paid calls a day and someone has to answer for all of them.

Where Vayr Pay fits

We are building for the ratio, not the headline. Policy evaluated per quote in milliseconds so nothing waits on a human unless the rules say it should. Budgets and velocity limits that see across every agent and wallet, so aggregate risk is visible even when every individual payment is trivially small. Attribution captured at the moment of authorisation — agent, principal, project, policy version — so a million micro-payments roll up into something a finance team can actually read. Funds stay with the wallet provider. We take the decision and the record.

Fourteen cents a transaction sounds like nothing to govern. Multiply it by the counts this protocol is already producing, and it is the only thing worth governing.

• Cloudflare gave agents a wallet. A spending cap is not governance.
  Cloudflare Wallets and cloudflare.pay give AI agents an identity and a virtual wallet with allowances, allowlists, and per-transaction caps. That closes the buyer side of x402 — but per-wallet limits are not the same as company-wide spend policy, approvals, and auditable receipts.
• Cloudflare will let anyone charge agents. Someone still has to let agents pay.
  Cloudflare's Monetization Gateway turns any API, page, dataset, or MCP tool into a paid resource over x402. That solves the merchant side. The buyer side — how an AI agent actually decides to pay, under whose rules — is the other half of the problem.