Software paying other software: the phenomenon can now be measured. In the North American section of its annual report on crypto adoption, blockchain analytics firm Chainalysis states that artificial intelligence agents settled more than 138 million transactions through the x402 protocol on the Base blockchain in the third quarter of 2026. A year earlier, the counter stood at 58,800. In June, the same firm had recorded 100 million.

How x402 works, the crypto payment designed for machines

The x402 protocol was launched by Coinbase in May 2025. It revives a web code that had gone unused since the 1990s: HTTP 402, “Payment Required”. An AI agent calls an API; the server replies that payment is required first. The agent then sends a few cents in USDC, the stablecoin issued by Circle, on Base, and receives its response. No customer account, no subscription.

The amounts explain why this crypto payment has taken hold in this niche. According to Chainalysis, more than 87% of payments fall between one hundredth of a cent and ten cents, mostly between one and ten cents. “Sub-dollar transactions make no economic sense for a credit card network,” the firm writes.

What these agents buy is mainly the services of other agents: model inference, tool catalogues, data and search APIs, code auditing. “Robots are buying AI-generated knowledge on the blockchain,” the report sums up.

What the Chainalysis figures measure

Beyond volume, the report describes a change in the behaviour of agent wallets. At the end of 2025, the median time between the first deposit into an agent wallet and its first spend was 73 days. In the third quarter of 2026, that delay had fallen to 18 minutes. Two thirds of wallets now spend within the hour after being funded, compared with 4% a year earlier, and 89.3% are topped up within 30 days.

The geographical breakdown is more mixed. The United States accounts for 30% of geolocated agent-to-service transactions, the top position. US wallets pay for 88.1% of model inference, 41.8% of payments to agent frameworks and tool catalogues, and 23.2% of purchases of data and search APIs. China follows the opposite logic: nine out of ten transactions in the “security and code analysis” category come from wallets linked to that country, almost all of them going to AgentLISA, a service that audits Solidity smart contracts using a language model. These wallets also account for 46.7% of financial intelligence and 35% of payment rails, but 0.1% of inference.

Chainalysis puts forward an explanation it calls “AI nationalism”: DeepSeek and Qwen are available free of charge as open weights, whereas ChatGPT and Claude are billed per request. This is the firm’s hypothesis, not a result demonstrated by the data.

What they do not measure

Several caveats apply:

  • the country shares only cover transactions that can be geolocated, and Chainalysis does not specify what fraction of the total they represent;
  • the figures concern x402 on Base: they do not cover all payments made by AI agents;
  • a number of transactions is not revenue: with amounts mostly between one and ten cents, it says nothing about the total value exchanged.

The players cited also have a direct interest in these uses. Hoolie Tejwani, who heads Coinbase Ventures, speaks of a “maturing of the industry” in which “every agent becomes a new blockchain user”. Circle, for its part, describes a division of roles: cards for purchases triggered by a human, stablecoins for machine-to-machine settlements too small for card economics.

What machine-to-machine payments change for Swiss companies

The North American section of the report, on which this article is based, gives no figures for Switzerland. Its questions nonetheless concern any company that exposes an API or deploys agents, Swiss fintechs included.

The first is commercial. Billing a service per request, for a few cents, without creating a customer account, opens up a model that subscriptions and cards do not allow. The second concerns liability. Payments that are too fast, too small and too numerous cannot serve as a control point, Chainalysis writes; liability must therefore go back to whoever deployed and funded the agent, not to the agent itself. For a company, the framework is therefore set upstream: who funds the wallet, with what cap, for what spending.

Some players are already drawing the consequences. In August, Cloudflare opened wallets dedicated to agents, with a cap set by their owner. On the infrastructure side, Circle reported USD 73.3 billion of USDC in circulation at the end of September, up 19% year on year, a sign that stablecoin payment rails extend well beyond agents.

Agent adoption also remains uneven across regions, as the case of European SMEs facing AI agents shows. The Chainalysis figures describe a clear trend on a specific protocol; they do not yet make it possible to say what place machine-to-machine payments will take in the real economy.

Sources

  • Journal du Coin: “Agents IA : 138 millions de paiements x402 sur Base en un an selon Chainalysis”