A TikTok told me x402 was an infinite money glitch. So I measured all 15,133 endpoints.

I got here the embarrassing way. A video told me there was an infinite money glitch in agent payments, that the window was closing, and that I should move fast. I am a support engineer, not a crypto person, and I did not buy anything. What I did instead was spend an evening building the thing to see whether it was real.

It is real in the sense that it works. x402 is Coinbase's HTTP 402 payment standard, and it lets AI agents pay for API calls with stablecoins. An agent hits my endpoint, gets a 402 back with payment requirements, pays in USDC on Base, and gets its response. The whole loop settles in a couple of seconds and costs a fraction of a cent. Shipping a paid endpoint end to end took one evening including the deploy.

So the protocol delivers. The next question is the one the video was actually making a claim about: what do sellers on this protocol earn?

I could not find that number anywhere. So I measured it.

The short answer is that the median x402 endpoint receives one call per 30 days, twenty-eight endpoints out of fifteen thousand clear a thousand calls, and total revenue across every seller on the protocol is somewhere around ten thousand dollars a month. Here is how I got there, and the script so you can check it yourself.

The published numbers disagree by two orders of magnitude

Before measuring anything, I tried to just look it up. That turned out to be the interesting part.

x402.org reports roughly $24 million in volume over 30 days, across about 75 million transactions. Artemis, analyzing the same period, puts it under $2 million and estimates real commercial throughput at around $28,000 per day, with roughly 95% of the headline figure being protocol signaling rather than buyer-to-seller exchange. One third-party dashboard claims $148 million in ecosystem volume, a figure that exceeds every other source by one to two orders of magnitude, reports two different services with byte-identical revenue and payment counts on its own category page, and simultaneously states it has tracked under $1 million in revenue.

Chainalysis, which is more optimistic than most, still declines to say the volume is predominantly organic. Its own analysis notes that a pay-to-mint memecoin drove over 150,000 transactions in a single month in late 2025, and that weekly wallet retention peaked at 87% during that period and then collapsed to 5%.

None of this is necessarily anyone lying. It is a measurement problem with a structural cause: nearly everyone publishing x402 numbers has a position in x402. The dashboard above monetizes itself through x402 and charges $1 to $25 per query. The most-cited block explorer for the protocol, x402scan, is operated by Merit Systems, whose own service StableEnrich ranks first on it. Quantum Economics calls payment-to-delivery attribution the field's blind spot, and an academic measurement study covering 119 million transactions across fifteen facilitators focuses entirely on facilitator security and publishes no seller revenue distribution at all.

That is the gap. There is no disinterested measurement of what x402 sellers earn. There does not need to be one, because the data has been public the whole time.

The discovery API tells you everything

Coinbase's Bazaar is x402's discovery layer, the catalog agents query to find services they can pay for. Its endpoint is public and unauthenticated:

GET https://api.cdp.coinbase.com/platform/v2/x402/discovery/resources

Every resource in it carries a quality object:

{
  "l30DaysTotalCalls": 874,
  "l30DaysUniquePayers": 870,
  "lastCalledAt": "2026-08-17T21:51:00.522Z"
}

That is per-endpoint usage data, published by the facilitator that settles the payments, for free, on every listing. Paginate the catalog and you have the demand side of the entire protocol.

One thing to watch, because it cost me a wrong answer. The price field on accepts[0] is amount, not maxAmountRequired as it appears in the x402 payment requirements spec. Reading the wrong key returns undefined for every record, and if you coerce that to zero you will conclude that 96% of the catalog is free. It is not. Check your field names against a single record before you aggregate fifteen thousand of them.

Here is the whole thing:

const BASE = 'https://api.cdp.coinbase.com/platform/v2/x402/discovery/resources';

const all = [];
for (let offset = 0; offset < 25000; offset += 1000) {
  const res = await fetch(`${BASE}?limit=1000&offset=${offset}`);
  if (!res.ok) break;
  const { items = [] } = await res.json();
  if (!items.length) break;
  all.push(...items);
}

// Pages can overlap between requests. Deduplicate on resource URL.
const seen = new Set();
const unique = all.filter(r => !seen.has(r.resource) && seen.add(r.resource));

const calls = unique
  .map(r => Number(r.quality?.l30DaysTotalCalls ?? 0))
  .sort((a, b) => a - b);

const at = p => calls[Math.floor((calls.length - 1) * p)];

console.log({
  resources: unique.length,
  median: at(0.50),
  p90: at(0.90),
  p99: at(0.99),
  max: calls.at(-1),
  totalCalls: calls.reduce((a, b) => a + b, 0),
  over1000: calls.filter(c => c > 1000).length,
});

Node 18 or later, no dependencies, runs in about ten seconds.

What it says

Pulled on 18 August 2026: 15,133 unique resources across 1,559 distinct domains.

Measure	Calls in 30 days
Median endpoint	1
75th percentile	3
90th percentile	7
99th percentile	207
Busiest single endpoint	60,220

Half of every endpoint listed in the catalog was called once or less in a month. Not once a day. Once.

Twenty-eight resources out of 15,133 cleared a thousand calls. Five cleared ten thousand. The top ten hold 52% of all 326,665 calls in the catalog, and 8,425 endpoints, 56% of the total, had not been called at all in the previous seven days.

Pricing is uniformly low: 81% of listings are priced under ten cents, and the largest single band is one to ten cents.

Multiply call count by listed price and you get an estimate of revenue. The top domain, StableEnrich, comes out at $1,147 for the month. Nothing else clears a thousand dollars. Summed across the entire catalog, every x402 seller combined earns $10,180 a month.

That estimate is generous, and deliberately so. It assumes every call settled, at list price, with no failures and no free tiers. It also only covers CDP-indexed resources, so sellers on other facilitators are invisible to it. The true number is lower than what I am reporting, not higher.

For scale: Nansen, a funded crypto analytics company with a real brand, appears in a separate x402scan snapshot at $152 for a month.

The platform already ate the obvious business

There is a second finding in that quality object, and it matters more than the revenue numbers if you are thinking about building here.

A reasonable-sounding x402 business is provider verification: check whether an endpoint is reachable, whether it implements the protocol correctly, and whether it has a track record. Sell that to agents as a preflight check before they spend money with a stranger. I built one. There are at least a dozen others in the catalog right now offering variations on it.

The first two checks are commodities. Reachability is a GET request. Protocol compliance is reading a spec against a response. Anyone can write both in an afternoon, which is exactly why a dozen people have.

The third check, reliability history, is the only part with a durable advantage, because it requires observations accumulated over time that a competitor cannot simply recompute. And CDP now ships it free, on every listing, computed from settlement data that no third party can see. l30DaysTotalCalls and l30DaysUniquePayers are reliability history. The facilitator sits on the payment path, so its version is more complete than anything an outside observer can assemble by polling.

This is the ordinary pattern of a platform absorbing a complement, and it is worth naming plainly because people are actively building into it. If your product is a metric the platform can compute better than you can from data only it holds, you are not building a business, you are building a feature it has not shipped yet. In this case it already has.

The hype recruits sellers, not buyers

Here is the thing I did not expect to find, and it is the reason I told you how I got here.

Look at the shape of the market again. Fifteen thousand endpoints. Three hundred and twenty-seven thousand calls a month, half of which go to ten of them. That is not a market with weak demand and matching weak supply. It is enormous supply pressed against almost no demand at all, and lopsided distributions like that do not happen by accident.

They happen because the marketing is aimed at one side. Nobody makes a video telling AI agents to go spend money. The videos tell people there is money to be made, and the people who act on them do what I did: they build an endpoint and list it. Every one of those fifteen thousand listings is somebody's evening. The demand side has no equivalent recruiting campaign, because the demand side would have to be autonomous agents with funded wallets and a reason to buy, and those are considerably harder to manufacture than enthusiasm.

The volume data supports this directly. Independent analysis attributes roughly 47% of x402 transactions to non-organic activity, mostly self-payments and leaderboard farming, and one December snapshot put gamed activity at 81% of transaction volume. Chainalysis tracks wallets that transfer to themselves before ever making a real payment. Those aren't buyers either. They are more sellers, paying themselves in a circle, hoping a token airdrop eventually justifies it.

I am a data point in my own dataset. The video worked on me, I added an endpoint, and the endpoint sat there. That is the whole mechanism, and it is visible in the percentiles.

What I would tell someone considering building here

Not that x402 is fake. The protocol works, the settlement is fast, and the developer experience is decent.

The problem is the market, and it is worth being unromantic about the size. The categories that convert at all are data aggregation and inference routing. Everything else, on the evidence, is a ghost town.

If your plan involves per-call revenue from agent traffic, run the numbers against the actual distribution before you write code. The entire seller economy on this protocol is currently worth about one mid-market SaaS seat.

If you are building because the rails are interesting and you want the experience, that is a completely different and much better reason, and the protocol is genuinely worth an evening. Mine cost me nothing but the evening, and I got a dataset out of it.

The next time something tells you the window is closing, the window is the product. Take the week. The numbers are public.

The script is above. The numbers move. The method holds.