The x402 protocol moved $2.6 billion last month. Almost none of it was AI agents.

01 · The agent that never stops buying

One wallet, one seller, 52 days without a break

On Base, a blockchain built by Coinbase, there is a wallet that has bought a thing every third of a second since the middle of July. It pays about seven tenths of a cent each time. It has done this at four in the morning and at four in the afternoon, on weekends, without a pause longer than a few minutes, for 52 days. It has now made more buys than there are seconds in 5 months.

Nobody approves those buys. That is the point of them.

What the wallet is using has a name. It is called x402, a way for software to pay for something over the internet without a card, an invoice or a person. A program asks a server for a thing. The server says the thing costs a fraction of a cent. The program signs a payment, the server checks it, and the answer comes back. The whole exchange takes a second and settles in stablecoins on a public blockchain, which means anyone can count it.

The software doing the buying is called an agent, and x402 was built for the AI kind. Whether an AI sits behind any one wallet is not something a ledger records. What the chain shows is software paying without a person approving it, and that is what we counted.

So we counted it. Bitquery indexes every transaction on Base, Polygon, Ethereum, Arbitrum, Optimism and Solana, and on the five EVM chains this payment type leaves an exact mark in the data. We took every one made in 2026, through the first week of September, and asked what a payment count on its own cannot answer: which chains carry this, what is being bought, who earns from it, and whether the money goes anywhere near the agents. Five chains turn out to be running five different businesses on one signature, and the smallest of them is the one everybody writes about.

One limit belongs up front. Our Base records reach back 58 days, so this cannot speak to Base before July, and no growth rate here includes it.

02 · One signature, several businesses

Why the same payment type means five different things

The method underneath x402 is older than x402 and open to anyone. It lets a person sign away a fixed amount of stablecoin and have somebody else post it to the chain and pay the network fee. That somebody is called a facilitator. That is useful to an agent buying an API call. It is just as useful to a shop taking a card-style payment, to a trading venue taking a deposit, and to a bridge moving money to another chain. All of them land in the data looking the same.

Two things tell them apart. The first is price, because a person moves dollars and an agent moves a fraction of a cent. The second is where the money lands, because a bridge or an exchange collects into one address while a market of sellers does not. Read those together and the chains sort themselves.

03 · Five chains, five businesses

What each one is actually being used for

Put the chains side by side and none is doing the same job as another. The median payment, the one sitting in the middle when you line up a day's traffic, runs from a tenth of a cent to $72.

04 · Most of the money is one bridge

Nine tenths of the money never goes near an agent

Arbitrum carries under 2% of the payments on this rail and 90% of the dollars. Follow that money and it goes almost nowhere: a single verified contract called CctpExtension took 97.5% of the chain's whole volume in August, from 34,577 different payers, at a median of $749 and one payment of $10 million.

CCTP is Circle's cross-chain transfer protocol, the plumbing that moves USDC from one blockchain to another. What the data shows is people signing gasless authorisations to push money into a bridge. It is a real and large business. It has nothing to do with agents, and it is the reason any headline dollar figure for x402 looks enormous.

Take that one contract out and the rail moved $317 million in August rather than $2.59 billion.

05 · Where the payments are

Base and Polygon carry 98% of the traffic, and two loops carry them

The payment count lives somewhere else entirely. Base and Polygon between them settled 17.8 million of August's 18.3 million payments, and both are dominated by a single repeating relationship.

On Base one wallet has sent 13.2 million payments to one address over 52 days, and that address has almost never spent a cent: 6 payments out in the whole window. On Polygon 5.7 million payments were all for exactly the same amount, one cent, cycled between a few hundred paying addresses and a few dozen receiving ones. All that Polygon traffic moved about $57,000 in total, which is roughly what a used car costs.

Those two together are 83% of every agent payment we found. What remains is about 2.2 million payments a month across thousands of buyers and sellers on four chains, and that is the honest size of agent commerce today.

06 · Solana, where it looks different

The one chain that behaves like a market

Solana carries a lot of this traffic and cannot join the table above, because it has no marker to key on. There is no authorisation event: a sponsored payment there is an ordinary token transfer where somebody other than the owner pays the network fee, which is also what every router and exchange sweep looks like. The closest honest proxy is sponsored payments of a cent or less, sampled a day at a time.

On that measure Solana looks nothing like Base or Polygon. A single day in August carried more than 800,000 of these payments, paid for by over 15,000 different sponsors. No single seller took more than 5.4% of the day. The same measure is 71% on Base and 99% on Polygon.

So the chain with the least protocol machinery has the most convincing spread of buyers and sellers. It is also the chain where we are least sure what we are counting, because sponsored-and-tiny describes a shape rather than a standard, and it catches dust and spam alongside real trade.

07 · What is actually growing

Settled value by month, and it is not the agents

Read the two series against each other and they disagree. Payments on the chains we can watch all year grew a little over three times since January. Value on those same chains barely moved, because the money that arrived in 2026 arrived on Arbitrum, inside a bridge, and never went near a paid API.

08 · Who earns from the fees

One processor charges. Everybody else works for nothing.

An x402 payment carries no fee of its own. The full signed amount goes to the seller, and whoever posts it to the chain pays the network cost out of their own pocket. The question worth asking is who has built something on top that does take a cut.

On Base, exactly one has. A contract from Coinbase's Commerce Payments Protocol handles merchant checkout, holds the money in escrow, and writes the fee terms into the payment itself: a minimum rate, a maximum rate, and the address that collects. That makes the whole fee layer readable straight from the chain, which is unusual. It exists on no other chain we checked.

That escrow also records whether each trade finished, which makes this the one corner of the rail where completion can be measured at all.

A refund rate that low is not what a market full of disputes looks like. Whatever is being sold through this contract, buyers are keeping it.

The processor charging that 1% is itself a Coinbase Smart Wallet, and the address collecting the money is an ordinary wallet with no name attached. Two things stand out. The fee-paying side is tiny, fewer than 75 merchants, and it is the only part of this rail where anybody is visibly earning from the payments rather than from whatever the payments buy. Everyone else runs the pipes for free and hopes the volume turns into money later.

09 · What it costs to run the pipes

Gas takes a fifth of a typical agent payment

The people posting these payments carry a real cost. In August the largest facilitator on Base burned about 4.5 ether in network fees to deliver its payments, one payment per transaction, with no batching anywhere. Set that against the median payment on that chain and roughly a fifth of the money being moved is spent moving it.

Measured against the total dollars instead, the same fees come to under 2%, because a handful of large payments carry most of the value. The gap between those two figures is the finding. A typical agent payment barely covers the cost of its own delivery, and the rail stays solvent because bigger transfers ride on it.

That facilitator is worth a look on its own. Twenty wallets relayed nine of every ten payments on Base, serving 74,942 sellers, and every one of the twenty was funded by the same wallet in the same instalments to within a rounding error. It is one system wearing 20 faces. We could not work out whose, and we are not going to imply an answer. The one address the public labels as Coinbase's x402 facilitator settled nothing at all on Base in our window, so searching by label finds none of the real traffic.

10 · Does anybody make a living from this?

The median seller earned $2.50

Across 54,383 addresses that took a payment on Base in August, the middle one collected $2.50 for the month. Two in five earned less than a dollar. Retention is worse. Of the 80,642 sellers that took a payment in July, fewer than 6 in every 100 took anything at all in August, and nine in ten of August's sellers were new.

That average hides the one piece of good news. Split sellers by how busy they were and the casual ones melt away while the serious ones stay, almost to a wallet.

About 1,300 sellers took 10 payments or more in July and roughly three quarters were still trading a month later. That is the honest size of the recurring seller base, against a headline of nearly 75,000. Buyers hold up better: a casual buyer is nearly five times more likely to return than a casual seller. Standing up a thing to sell is a weekend experiment that usually gets abandoned, while paying for what you need is a habit.

11 · The recordAddresses behind the story

Bitquery's public explorer does not yet carry Base, Polygon or Arbitrum routes, so these are listed rather than linked. Figures were measured on 5 September 2026 and move by about a hundredth of a percent between runs.

12 · How we countedMethod, and what it cannot tell you

Every figure came from Bitquery's own index of these chains, reading the decoded contents of each payment instruction rather than the transfer records, which mis-report this payment type. We kept only payments that went through, and removed the duplicate that appears because the stablecoin contract forwards each call to a second contract behind it. Counting without that step doubles everything. Each chain's USDC contract was checked by symbol and decimals before use. Solana is read from a separate index and sampled a day at a time, which is why it sits outside the main comparison.

This payment type is used by things other than x402, which is why the price split and the destination both matter, and why a bridge and a shop sit in the same data. Reading intent from behaviour is inference, and it is flagged as inference each time.

You can run all of this yourself. The same index sits behind the Base, Polygon and Arbitrum APIs, the address APIs for tracing a wallet's counterparties, and Coinpath for following money between them. For the same method applied to one chain's whole economy, see our count of

, and the arbitrage piece for the other end of the same economics question.