x402scan shows x402, the agent payment standard Coinbase started, moving about $1.9M a month. If AI agents are buying things, that's where it should show. I build dvmkit, tools for selling services to agents, and I want that number to be real. So I set out to find what those payments bought.
I took all 5.6 million payments x402scan counted in the 30 days to 16 November, on Base and Solana, and followed the money on-chain.
Two flows make up two thirds of the $1.9M. Both start and end at Binance, and neither bought anything I can see. The payments I can tie to services sold per request come to $6,905.

$679,000 that went from Binance back to Binance
The biggest flow is nine transfers, $679,000 in total, from one wallet to one Binance deposit address.
Since August that wallet has taken in $1.22M, all but a few cents of it from Binance, and sent $1.17M on to that one deposit address in 21 round sums. Twelve times, the same amount came in from Binance and went back out within a day. The day after my 30 days ended, $90,000 came in and the same $90,000 left 16 minutes later.

Why does x402scan count this as agent spending? It can't see what a transfer paid for, so it goes by who submitted it. This wallet never submits its own transactions. It signs each USDC transfer, and a relayer submits it and pays the gas, which is how x402 payments on Base work. The relayers here are ones x402scan attributes to FluxA, an x402 facilitator that also runs hosted wallets and payouts. So x402scan counts all nine as x402 payments, though nothing in them points to a purchase: no URL, no order, no reference.
It looks like someone using x402's gasless transfers to move money between Binance accounts, and a dashboard counting it as agent spending. I can't see who controls the wallet or why. Either way, a third of x402's headline volume shows no sign of touching an API. I tried to ask FluxA before publishing, but its X inbox is closed to messages.
$583,813 to an AI video app, mostly $200 at a time
The second flow went to an address I couldn't find in x402scan's seller registry or Coinbase's x402 catalogue. Then I found two posts from ArcNova, which sells AI video generation: demos of an agent paying under a dollar for a clip from its Nanara studio, each linking a transaction to this address. ArcNova promotes the total as traction. On 11 November it claimed "400,000+ USDC" settled, which matches the $421,122 x402scan had counted for the address by then.
Nanara charges by the second of video. The payer signs a maximum, like a card pre-authorisation, and is charged only for what's delivered, so you'd expect odd amounts, like $99.27. Instead 85% of the dollars came in exact multiples of $50, and 2,113 payments of exactly $200 brought in $422,600, nearly three quarters of the total. In a random sample of 18 payments, 9 had signed a $200 maximum and were charged all of it.
The payers behave like a fleet more than a customer base. Almost all of the 265 addresses were funded minutes before they started paying, and almost all paid on just one day. The typical one made 15 payments over 17 minutes. Nearly all their funding traces back to Binance, and 123 of them emptied their leftover USDC into one of just two Binance deposit addresses, most likely one or two accounts. Neither is the deposit address the video app's address cashes out to.
The payments also run on a timetable: all but two of the 3,846 landed between 02:00 and noon UTC.

Everything that arrived went on to Binance, almost all of it straight to one deposit address, on average within two hours. When I checked on 19 November, the address held nothing.
It looks like one operation, or very few, running funded wallets through Nanara on a timetable, with the money starting and ending at Binance. I have no evidence ArcNova controls the paying wallets, and I can't see what each payment produced. If these are genuine sales, ArcNova is by far the biggest seller on x402. If they aren't, x402's biggest seller is a fleet of wallets funded from Binance, paying round sums on a timetable. I asked ArcNova on 20 November what the $200 payments buy and who the main payers are. It hadn't replied by publication.
What the rest bought
Take out those two flows and $636,342 is left.
Nearly half, $288,865, went to addresses I couldn't tie to any seller. Another $252,311 bought prepaid value: cards and payouts from Laso, and service credits from dTelecom. That is real buying, but of stored value to spend later, not a service paid per call.
A further $63,832 matches farming patterns, like sellers paying their own listings: 57% of all payments, but 3% of the dollars.
That leaves $6,905 for 280 listings that sell an API call, a model response, or a search at a set price. Ten took 77% of it, and the median payment I could measure was half a cent. Some have payers who come back: at BlockRun, which gives agents one pay-per-call endpoint for AI models and tools, 221 of the 408 paying addresses paid on more than one day. I wrote up who these sellers are and what sells in a second piece.
So how big is it?
Take out the round trips and x402's per-request market is about $7,000 a month: mostly sub-cent calls to a few hundred APIs, with payers who come back.
Even if every untraced dollar were a per-request sale, it would be under $300,000. That's the number I'd start from until someone explains the big flows. MPP looks similar: $8,350 in total on MPPScan over the same 30 days, $3,254 of it to per-call services.
The transfers in both flows, and every per-request listing, are in the data, as CSV files. If you're named here and I've got something wrong, tell me and I'll correct it. I'll rerun this in December.
How I counted
- Window and chains. Every x402 payment x402scan indexed on Base and Solana from 17 October 23:50 to 16 November 23:50 UTC: 5,609,330 payments, $1,899,155.
- Labels. Exchange labels, such as "Binance deposit address", are Basescan's. Solana amounts carry none.
- Facilitators. Which facilitator a payment went through is x402scan's mapping of relayer addresses to facilitators.
- The order of the buckets. I set aside payments to addresses I couldn't tie to any seller first, then payments that match a farming pattern, then prepaid value. What was left split into listings of unclear purpose ($24,429) and per-request services ($6,905).
- Farming filters. I counted a listing's payments as farming or self-testing when it had just one paying address in the window, two addresses made more than 80% of its payments, 100 or more paying addresses averaged at most 1.2 payments each, at least 80% of 100 or more payers paid the same price about the same number of times, or its own page offers points or rewards for calling. Loosening or tightening the thresholds in the second, third and fourth of these filters keeps the money that matched no farming filter between $282,291 and $285,452.
- Prepaid value. $252,311 in all, including Laso's cards and payouts ($214,888) and dTelecom credits ($36,114). Feldspar reported that a single paying address made 98.6% of Laso's dollars in an earlier month, the 30 days to 28 October.
- For comparison. In December 2025, Bankless reported an Artemis estimate that put 47% of x402 transactions, and 14% of its dollars, down to leaderboard gaming.
- The ArcNova sample. The random sample of 18 payments used a fixed seed. I also decoded the first two payments of exactly $200 alongside it.
- Other rails. l402.space, a Lightning gateway that publishes its counters, shows $1.66 across 19 transactions. Stripe, Visa, Mastercard and PayPal publish no spending totals for agent checkout.