138 million AI agent payments: what the x402 count measures

Chainalysis counts more than 138 million payments by AI agents through x402, a protocol incubated by Coinbase that lets software pay for online services in crypto without a human pressing a button. Decrypt and CoinDesk both covered the research on Wednesday. This article looks at what each number in the report actually counts, how the wallet tests are built, and why a big transaction tally can say less than it seems.

What the research found

The study is part of the analytics firm's annual Geography of Crypto series, in a chapter on North America's $3 trillion crypto economy. Its subject is agentic commerce: software agents that hold a budget and spend it on data, computing and tools.

According to Decrypt, the running total of x402 payments on Base, Coinbase's layer-2 network, was near 1,500 in mid-2025 and passed 138 million in the third quarter of 2026. Chainalysis stretched its review window to the 16 months ending in September 2026 because the technology is so new.

The payments are tiny. Chainalysis found that over 87% fell in a band from one hundredth of a cent up to ten cents, with one to ten cents the most common size. Card networks cannot handle amounts that small economically. CoinDesk added a sense of scale in dollars: across 30 days in July, the rails carried roughly 75 million payments with only $24 million in value.

The metric and the method

A transaction count is the simplest on-chain metric there is. Every payment is a public record on the ledger, so an analyst can add them up. But a count measures activity, not people and not value. One busy program can produce millions of entries.

The report goes beyond the raw count with two behaviour tests. The first is a cohort test. A cohort is a group of wallets sorted by when they were first funded. Of wallets first funded in Q3 2025, 70.6% received a second top-up within 30 days. For wallets first funded in Q2 2026, the share was 89.3%. The firm reads refilling an emptied wallet as something like a customer renewing a subscription. It also left out wallets active on only one day, keeping those that paid on two or more separate days.

The second test is speed. The median wallet age at first x402 spend fell from 73 days in late 2025 to 18 minutes by Q3 2026. The share of wallets paying within an hour of funding went from 4% to 67% over a year. Read together, these figures suggest wallets are increasingly set up to spend quickly. Early activity, Chainalysis noted, was largely fads, with memecoins such as PING drawing wallets that later faded.

Geography is the third layer. A country share only exists for wallets that could be linked to a place. Among identified payments for AI model inference, CoinDesk reported that U.S.-linked wallets held 88.1% and Chinese-linked wallets 0.1%. In security and code-analysis tools the picture flipped: about 90% came from Chinese-linked wallets, much of it paid to AgentLisa, an AI smart contract auditor.

Limits: where the numbers can mislead

Every one of these measures rests on choices that a reader should keep in mind.

• Counts are not demand. Decrypt pointed to earlier work by rival tracing firm TRM suggesting that a lot of the apparent agent traffic could be scripts or self-dealing rather than true autonomous agents. CoinDesk made the same point: transaction counts alone do not show genuine commercial demand.
• Coverage is partial. The figures include only identifiable x402 activity, not the whole market for AI payments, as the firm itself cautioned.
• Geolocation is an estimate. A country share describes the wallets that could be tagged, not every wallet. Eric Jardine, head of research at Chainalysis, said Chinese developers may lean on open-weight models like DeepSeek and Qwen, but warned against reading too much into the gap.
• Filters shape results. Dropping one-day wallets makes cohorts look more committed by design. That is a reasonable choice, but it means the retention figures describe repeat users only.

Jardine put the caution plainly: "That's one possible interpretation, but we wouldn't go so far as to say open-source AI is starving Asia's on-chain inference market".

Why the pattern also matters for security

Chainalysis argues that compliance work changes shape here. Screening each tiny payment makes little sense. Instead, the checks move to the funding setup, the destination service and the spending pattern. For on-chain readers, that is a useful habit too. When a wallet is funded and starts spending within minutes, the funding source and the counterparty tell you more than the payment itself. The same habit helps spot fake volume, because a script that pays itself would show up as the same few wallets sending value back and forth.

What to watch next

Three things would make the picture clearer. First, value alongside counts: dollar totals, like the July figure, show whether activity is growing in substance or just in volume. Second, independent checks from other tracing firms on how much traffic is self-dealing. Third, the share of activity that can be geolocated at all, since country splits only cover that slice. None of these metrics is a signal about any token's price. They describe how a new payment rail is being used.

Sources

• Decrypt, 14 Oct 2026
• CoinDesk, 14 Oct 2026
• Chainalysis, 14 Oct 2026

Drafted by AI from the research and on-chain sources listed, then published automatically once checks confirmed the figures against those sources and found no copied text. Not reviewed by a person. Metrics are explained, never turned into price calls.