What AI agents buy: Chainalysis' x402 data maps the demand — inference in the US, code checks in China, a cent a call, and wallets refilled like subscriptions

What Chainalysis published on October 14, 2026

On Wednesday, October 14, 2026, Chainalysis published “North America: Inside the Growth of Agentic Commerce”, the North America chapter of its 2026 Geography of Cryptocurrency report. The region's crypto economy came to $3.02 trillion in the 12 months to July 1, 2026, down 1.9%. The US shrank 2.2% to $2.78 trillion; Canada grew 2.3% to $247.7 billion.

The headline, though, was a payments protocol. To measure agentic commerce — AI agents and other software paying for services without a person approving each purchase — Chainalysis studied x402. It's the open standard that revives HTTP's 402 Payment Required code: a server quotes a price, and an agent pays it in a stablecoin inside the same request. Coinbase published it in May 2025. The x402 Foundation, under the Linux Foundation, now stewards it.

On Base, Coinbase's Ethereum layer-2 network, Chainalysis counted more than 138 million x402 transactions by the third quarter of 2026. Decrypt's summary put the mid-2025 count at roughly 1,500. Chainalysis says agentic commerce is still in its “experimentation phase,” as the experts it interviewed put it, but with signs of maturation. Circle and Coinbase were among those it spoke to.

We've covered the volume story before: the first x402 audit stripped bridges and loops out of August's headline, and the 2026 volume decline turned out to be composition, not collapse. This report adds what those didn't have: what the agents were buying, and where the buyers were.

What AI agents buy, category by category

Chainalysis sorted the services agents pay for into five groups:

• Model inference — paying per request for an AI model's output.
• Information services — code review, security and code analysis, market and financial intelligence, data and search APIs.
• Agent frameworks and tooling — the infrastructure agents use to run, coordinate and call other services.
• Payment rails — gaming, token-issuance and gambling services. Early on, memecoin minting drove much of x402's volume; that faded.
• Agents buying from agents — one agent paying another to do part of a task.

It then geolocated wallets where it could. Here are the shares it reported:

Shares of geographically identifiable payments in each category, as reported by Chainalysis. A dash means the report didn't give that figure.

Two things stand out if you sell software. First, the list is almost entirely business-to-business: models, data, checks and tools. Nobody is buying sneakers. Agent demand on x402 today looks like a developer's procurement list, not a shopping cart.

Second, the same category can be one country's biggest line and another's smallest. That's the most useful finding in the report, and it isn't really about geography.

Agents pay for what their owner can't get free

The inference split is stark: 88.1% of identifiable inference payments came from US wallets, 0.1% from China-linked ones. The report's explanation is what it calls AI nationalism: the Chinese market leans toward domestic models led by DeepSeek and Qwen. Then comes the economics: “China's open-weight models are free to use; US-developed, metered models, led by ChatGPT and Claude, are not.” US agents pay for every call because their owners' preferred models charge for every call.

China-linked agents spend their budgets elsewhere: about 90% of identifiable security and code-analysis payments and 46.7% of market and financial intelligence. Where inference is free, the money goes to specialized checks and data that a model can't produce by itself.

The lesson for sellers is general: an agent's real alternative to your API is whatever its owner can already get for free. If a buyer's own model can do what your endpoint does, your price is competing with zero. Ask three questions before you list an endpoint:

• Could the buyer's model answer without me? If yes, you're selling convenience, and convenience prices in fractions of a cent.
• Does my endpoint return something a model can't make up? Fresh data, a proprietary index, a verified check, a result someone stands behind. That's what every paid category above has in common.
• Where do my buyers' agents run, and what do their owners already pay for? A US agent may happily pay for model access. An agent running on a free local model may need your data but never your inference.

A cent a call is the going price

Over the 16 months to September 2026, more than 87% of the x402 payments Chainalysis analyzed fell between a hundredth of a cent ($0.0001) and ten cents. The most common size was one to ten cents.

Those prices don't exist on card rails. As Decrypt noted, at that size card fees can cost more than the service itself. Circle described the division of labor to Chainalysis: “card rails for consumer-prompted agent payments, stablecoin rails for machine-to-machine settlements that are too small for card economics.”

We worked through the arithmetic last week. The Coinbase Institute's October 7 paper puts the price floor for cards — the smallest payment that keeps fees at or below 5% — at about $1.19 on the cheapest small-ticket pricing. Chainalysis now shows where agent demand actually sits: roughly a hundred times below that floor.

So the unit you sell is a product decision. A single row worth $0.001 is uneconomic for everyone involved. A page of 50 rows at five cents isn't. Package your output so one call lands in the band where buyers already transact.

The real news is in the wallets: 18 minutes, then a refill

Most coverage led with 138 million. For a seller, the behavior figures matter more:

1. Time to first spend. The median wallet's age at its first x402 payment fell from 73 days in late 2025 to 18 minutes in Q3 2026.
2. Spending within an hour. In Q3 2025, 4% of wallets paid within an hour of being funded. In Q3 2026, 67% did.
3. Coming back. 70.6% of the Q3 2025 cohort got a second funding within 30 days. For the Q2 2026 cohort, it was 89.3%.

Together, these describe a different buyer from a year ago. In 2025, wallets were funded and then sat — someone experimenting, or minting a memecoin. In 2026, wallets are funded for a job, spend on it almost at once and get topped up again. In the report's words, “each successive group of wallets has arrived more committed than the last.” To us, that pattern looks less like experiments and more like a subscription paid in top-ups.

Hoolie Tejwani, head of Coinbase Ventures, told Chainalysis that “a lot of what we're seeing vs. 2025 comes down to maturation of the industry.” His other line is the one to pin above your desk: “AI agents are becoming economic actors where every agent is effectively a new ‘user’ of the blockchain.”

Three consequences follow for anyone selling to them:

• You get minutes, not a sales cycle. An agent funded 18 minutes ago doesn't read your pricing page or book a demo. If your endpoint can't be found, priced and paid in one round trip, that budget goes somewhere else.
• Revenue arrives as refills. A 90% re-funding rate means owners keep budgeting for their agents. You want to be the endpoint an agent returns to on its next top-up. That matches the September audit: sellers with ten or more payments kept 71% of their buyers.
• Count buyers, not calls. One agent can make thousands of one-cent calls. Your health metric is how many distinct wallets come back.

What the report doesn't tell you

Before you build a forecast on 138 million, read the gaps. Chainalysis flags most of them itself:

• No dollar total. The report counts transactions; it doesn't say how much money moved through x402 on Base. For a rough sense of scale, the x402.org dashboard's “last 30 days” panel showed about 75.4 million transactions and $24.2 million of volume across all networks when we checked on October 15, with 94,060 buyers and 22,000 sellers. That averages about 32 cents a payment and about $1,100 per seller — averages a few large sellers almost certainly dominate, given that the September audit found a median seller earning $2.50.
• One chain. The 138 million is Base only. x402 also runs on Solana and other networks; Solana carried 38% of a record week in August.
• Filtered wallets. The analysis left out wallets that paid on only one day.
• Agents or scripts. Decrypt pointed to earlier TRM Labs research suggesting much of the apparent agent activity may be scripts or self-dealing. On-chain, a scheduled script and an autonomous agent look the same.

None of that reverses the direction of travel. It changes how to use the report: as a map of what's being bought and how buyers behave, not as a market size to put in a pitch deck.

Compliance follows the funder — and the seller is a known counterparty

One section of the report deserves more attention than it got. Individual agent payments, Chainalysis writes, “are too fast, too small, and too frequent to serve as the compliance checkpoint.” The model offers other controls instead. The funding event identifies who sent the money. The budget is a human decision someone can review. The receiving service is a known counterparty. And the mechanical regularity of agent spending makes anomalies easier to spot than in human payment flows. In other words, accountability points back to whoever deployed and funded the agent, not to each payment.

Notice the third control: in this model, the seller is one of the checkpoints. A published endpoint, a real business behind it, a clear price and a wallet it controls are what make the receiving side legible.

The identity layer is moving too. On October 15, 2026, Proof launched x401, an open protocol that extends HTTP 401 so a website can ask an agent to prove who it represents and what that person authorized. Proof designed it to sit alongside x402: x401 for identity and authorization, x402 for payment. The card networks are building their own version, which we covered as Know Your Agent in September.

A seller's playbook from the data

1. Pick a category agents already buy. Inference, data and search, code and security checks, financial intelligence, agent tooling. If your API fits one, you're selling into demand that already exists.
2. Sell what a free model can't do. Fresh, proprietary or verified output. Price against the buyer's free alternative, not just against competitors.
3. Price in the one-to-ten-cent band. Bundle tiny units — rows, readings, pages — into one call that clears it.
4. Make the first paid call work in minutes. A 402 that states price, asset, network and recipient, and a response worth paying for twice. No signup and no API key for the buyer to request.
5. Design for the refill. Keep pricing and output consistent, so an agent on its next budget comes back. Track returning wallets.
6. Keep each payment in your own wallet. At a few cents a ticket, any processor that holds a balance, pools funds or adds a fixed fee per payment eats your margin and your control.

Where Payzum fits

Payzum sits in front of your existing API as x402 middleware — a proxy. It isn't the facilitator, and it never holds your money. In the dashboard you enter your endpoint, the API key or bearer token it already expects and a price per call. Payzum publishes an x402 URL for it and answers agents with the 402 Payment Required. The USDC payment on Base settles through an external facilitator — today Coinbase's — straight to your wallet. Payzum then forwards the paid request to your real endpoint with your key and returns the response.

There's no code to write and no protocol to implement, so an API provider can start serving agents the same day. Pricing is about 1,000 x402 transactions a month free, then roughly $0.001 per transaction plus gas. The same account gives you the REST API with API keys, signed webhooks and an integration playground. If you also sell to people, hosted checkout, payment links, invoices and subscriptions settle non-custodially to your wallet too.

Payzum doesn't implement x401 or any agent-identity protocol today, and we don't claim otherwise. More on the setup on the x402 page.

How it works, step by step

1. Add your endpoint: its URL, the API key or bearer token Payzum should forward and your receiving wallet on Base.
2. Set a price per call in USDC, in the band agents actually pay.
3. Publish the x402 URL Payzum gives you — in your docs, your agent-facing catalog or your MCP server.
4. Get paid per call: an agent calls the URL, receives the 402 and pays. The facilitator settles USDC to your wallet in about 2 seconds, and Payzum proxies the call to your API.

Selling to agents: API-key plans vs per-call x402

Fair objections

“138 million payments of a few cents — is there real money here?”

Not much yet, and the report doesn't pretend otherwise: it gives no dollar total and calls the market experimental. The case for acting now isn't size, it's cost. Becoming payable by agents takes a dashboard setup, and the buyers arriving now are the ones refilling budgets at a 90% rate.

“My API isn't an AI product. Do agents buy it?”

Most of what agents buy isn't AI. Data and search APIs, code checks and financial intelligence are all on the list. If your endpoint returns something fresh or proprietary that an agent needs halfway through a task, it fits — as market-data sellers found first.

“Should I wait for agent identity standards like x401?”

You don't need to. None of them is required for an agent to pay you over x402 today. Identity checks will matter more as buyers start presenting them; the payment side works now, and you can layer identity on when your market asks for it.

“Aren't many of these ‘agents’ just scripts?”

Some are, and that inflates the industry's counts. It doesn't inflate your revenue. Your own numbers — distinct paying wallets and how many come back — tell you what's real for your API.