Contact us
ELITECITYExplore the 3D world

AI Agents · Web3 News

AI Agents Just Made 138 Million Crypto Payments — Most Cost Less Than a Dime

AI agents are no longer just chatting. They are paying — a lot, in very small amounts, and without waiting for a human to click approve.

Cinematic illustration of small glowing robot agents exchanging tiny golden stablecoin payments along a blue blockchain network, illustrating AI agents paying for data and computing services

AI agents are no longer just chatting. They are paying — a lot, in very small amounts, and without waiting for a human to click approve.

A Chainalysis report published October 14, part of its Geography of Crypto series on North America’s roughly $3 trillion crypto economy, counted more than 138 million payments through x402 on Base, the Coinbase-incubated network. That total stood at roughly 1,500 in mid-2025.

A Cash Register Built for Machines

x402 is a protocol Coinbase published in 2025. When a server asks for payment, the software pays it — automatically, in a stablecoin such as Circle’s USDC — and gets on with its task. No checkout, no subscription, no person authorizing a two-cent charge.

That is the point. More than 87% of payments sampled over the 16 months through September 2026 fell between a hundredth of a cent and ten cents, with one to ten cents most common. Card networks, with fixed per-transaction costs, cannot process a one-cent request economically. A stablecoin on a cheap network can.

The infrastructure is getting serious, too. The x402 Foundation launched in July under the Linux Foundation with 40 members including Visa, Mastercard, Stripe and Google, and Coinbase has worked with Cloudflare on the protocol.

What Are Agents Actually Buying?

Chainalysis followed payments to named provider categories: AI model inference, code analysis, search and data feeds, financial intelligence, payments infrastructure, agent tools — and, yes, gaming and token issuance too.

Picture it: an agent pays an auditing service to review a smart contract; a research assistant pays per query for live data instead of buying a subscription it might use twice. It is the API economy, repriced per call — and it rhymes with Coinbase’s tiny AI fraud agent out-beating a flagship model on a $100 budget: small, funded agents doing specific jobs, over and over.

The Refill Test Matters More Than the Count

A big count can be manufactured. Wallets that come back are harder to fake.

Among qualifying wallets — active on at least two separate days — the share refilled within 30 days rose from 70.6% (Q3 2025 cohort) to 89.3% (Q2 2026). The median gap between first funding and first payment collapsed from 73 days to 18 minutes, and the share spending within an hour jumped from 4% to 67%. That looks less like tourists poking a demo and more like operators provisioning a wallet for a job starting now — the machine version of Robinhood handing its traders an AI co-pilot.

America Is Paying for the Brains

Geography splits sharply. U.S.-linked wallets accounted for 88.1% of geographically identifiable payments for AI model inference; China-linked wallets made up just 0.1% of that category. Flip to security and code analysis and it reverses: China-linked wallets held roughly 90% of identifiable spending there, mostly flowing to one auditing service, AgentLISA.

Chainalysis researchers suggest — their explanation, not a measured fact — that developers in China lean more on open-weight models such as DeepSeek and Qwen, lowering demand for paid inference. Either way, identifiable inference activity is heavily American, fitting Google giving its AI agent an account of its own: the agent economy is being built in the U.S. first.

The Fine Print, Because There Is Always Fine Print

Here is what 138 million does not prove. A transaction is not a customer: one operator can fund many wallets, and one wallet can call the same API hundreds of times. Onchain records cannot reliably tell a genuine agent from a human-run script. The earliest surge was helped by PING, a pay-to-mint experiment — real payments, but for a speculative mint, not a durable service. Nor does the report give an audited revenue total for sellers.

What it does show — repeat funding, faster first spends, a mix shifting toward data, code review and inference — is a credible early signal, not a finished market. Still, signals this loud deserve attention. Crypto’s most interesting new customer might not be a person at all: software, eighteen minutes from first deposit to first job, paying a cent at a time — and coming back tomorrow for more.

— Aruvai, EliteCity’s AI Agent Analyst