One year ago, the idea of an AI agent autonomously spending cryptocurrency to buy access to another AI model, a dataset, or a code-analysis service would have sounded like science fiction. By the third quarter of 2026, it had happened 138 million times.
That figure, drawn from new research by blockchain analytics firm Chainalysis, captures the total transactions processed through x402 — the Coinbase-incubated protocol that lets software pay for online services automatically — on Coinbase’s Base layer-2 network. The number is striking not just for its size, but for how fast it got there: in mid-2025, x402 was handling roughly 1,500 payments total. In July 2026 alone, it processed approximately 75 million. The curve is nearly vertical.
What’s happening here is something the crypto industry has long theorized but rarely seen in the data: a genuinely new category of blockchain user, one that isn’t a person at all.
How x402 Turned AI Agents Into Crypto Spenders
To understand why the payment counts are exploding, it helps to understand what x402 actually does. The protocol is built on a simple premise: when a software agent needs to access a paid resource — a premium AI model, a live data feed, a smart contract audit — it can pay automatically using stablecoin micropayments on Base, without any human having to approve the transaction.
The name x402 is a nod to HTTP status code 402, “Payment Required,” a code that has sat largely unused since the early web. Coinbase revived the concept and built a protocol around it, and the timing proved fortuitous. As AI agents proliferated in 2025 and 2026, developers needed a way for those agents to pay for the compute, data, and services they consumed. Credit cards don’t work for autonomous software. Traditional bank wires take days. x402, running on a blockchain that settles in seconds and charges fractions of a cent, does.
The adoption has been rapid. Visa, Mastercard, Stripe, and Google have all backed or built on the protocol in various forms, with Google unveiling its own dedicated agent-payments protocol earlier this year. Coinbase has also rolled out tools letting businesses accept agent payments directly, further widening the network effect.
The Numbers That Tell the Story
The Chainalysis report, released on October 14, 2026 as part of the firm’s annual Geography of Crypto series, provides the most comprehensive picture yet of this emerging market. Several figures stand out:
- More than 87% of all x402 payments fell between a hundredth of a cent and 10 cents. These are not investments or trades. They are operational costs — agents paying for exactly what they use, one inference at a time.
- The time from wallet funding to first payment has collapsed. In late 2025, the average new agent wallet took 73 days from first funding to its first x402 transaction. By Q3 2026, that lag had shrunk to 18 minutes — a sign that onboarding friction is effectively gone.
- Wallet retention has improved sharply. Among wallets first funded in Q2 2026, 89.3% received additional funding within 30 days, compared to 70.6% for wallets first funded in Q3 2025. The agents using these wallets are coming back.
“AI agents are becoming economic actors where every agent is effectively a new ‘user’ of the blockchain,” Tejwani, whose full analysis is published in the Chainalysis North America Crypto Adoption 2026 report, noted in the firm’s research. The phrasing is careful, but the implication is significant: the next wave of crypto adoption may not require any human to open a wallet app.
Who’s Behind the Payments — and Who Isn’t
Not all AI agent spending looks alike, and the geographic breakdown in Chainalysis’s data reveals a sharp and somewhat unexpected divide.
When it comes to paying for AI model inference — essentially, buying access to run prompts through commercial AI models — wallets linked to U.S. users account for 88.1% of geographically identifiable payments. Wallets linked to China account for just 0.1%.
The gap likely reflects model access rather than general enthusiasm for crypto. Eric Jardine, Chainalysis’s head of research, offered this interpretation to CoinDesk: “One possible explanation is that Chinese developers are more inclined to use open-weight models such as DeepSeek and Qwen, reducing their need to purchase metered inference services.” Open-source models, by definition, don’t require a per-query payment to a commercial provider.
The pattern flips entirely for security and code-analysis services. There, Chinese-linked wallets represent roughly 90% of identified spending, most of it flowing to AgentLisa, an AI-powered smart contract auditing platform. Chinese developers building on-chain applications, it appears, are willing to pay in crypto for security tooling — just not for Western AI models.
What “Machine Commerce” Actually Looks Like Today
It is worth pausing on what these 138 million transactions actually represent in practice, because the picture is both more mundane and more consequential than the headline number suggests.
When an AI coding assistant needs to verify a smart contract, it may call AgentLisa and pay $0.003 in USDC. When an autonomous research agent needs to query a live data API, it pays a fraction of a cent for each call. When a software pipeline needs to run a prompt through a commercial language model, it pays per token. None of these transactions require a human to notice them, let alone approve them. They happen in the background, automatically, as a cost of running AI software.
Some skepticism is warranted here. TRM Labs, a rival blockchain analytics firm, raised questions earlier this year about whether a portion of apparent agent activity on x402 reflects genuine autonomous agents or simply scripts and self-dealing — developers cycling funds through their own wallets to simulate activity. Chainalysis acknowledged this uncertainty in its report, cautioning against reading too much into the gap between transaction counts and genuine commercial demand.
What the data can say with more confidence is that the rails are real and being used at scale. Whether every transaction represents a fully autonomous economic actor is a philosophical question that the market will sort out over time.
The Infrastructure Being Built for 100 Billion Agents
Even at 138 million payments, the x402 ecosystem is in its infancy if the projections of major industry figures hold up. Yat Siu, co-founder and chairman of Animoca Brands, has estimated that up to 100 billion AI agents could be transacting on the internet over the coming years — a figure that would dwarf any existing payment network’s transaction volume.
The infrastructure required to support that scale is being built now, and it extends well beyond x402. Decentralized GPU networks — including io.net, Render Network, Akash, and Gensyn — are positioning themselves as the compute layer for AI workloads that hyperscalers cannot or will not serve. Collectively, their protocol revenues reached an estimated $200 million annualized in early 2026, according to data from DeFiLlama and Dune Analytics.
The appeal is partly economic. AWS, Microsoft Azure, and Google Cloud control roughly 65% of available GPU capacity, according to SemiAnalysis research, and their on-demand pricing reflects that dominance. Decentralized alternatives are, in many cases, dramatically cheaper for workloads that don’t require enterprise-level uptime guarantees. McKinsey estimates the global AI infrastructure market will surpass $700 billion annually by 2030, and even a small share of that directed toward decentralized networks would represent an enormous expansion from today’s figures.
BlackRock’s Will Su, head of digital assets research, told Chainalysis that the firm is watching the space closely, noting that “AI agents potentially creating meaningful demand for frequent, low-value payments around the clock” is exactly the kind of use case that stablecoins and fast settlement blockchains are optimized for — far more so than traditional payment rails designed for humans making occasional large transactions.
What This Means for You
If you’re a developer, the practical implication is that the barrier to building agents that spend crypto has essentially disappeared. x402 integration is straightforward, stablecoin liquidity on Base is deep, and the ecosystem of payable services is growing every month. Building an agent that can autonomously pay for the resources it needs is now less a technical challenge than a product decision.
If you’re a crypto investor, the picture is more nuanced. The raw transaction counts are impressive, but the dollar values are tiny — 138 million payments averaging a fraction of a cent works out to a modest total sum. The story here is less about near-term revenue and more about the long-term plumbing being laid for a world where AI agents are economic participants. That plumbing favors blockchains with fast finality, low fees, and stable-value tokens — Base/USDC, Solana, and similar Layer-2 and high-throughput networks.
If you’re simply a curious observer of where technology is heading, the more remarkable data point may be the one about onboarding speed. In late 2025, it took an AI agent’s wallet an average of 73 days from first funding to its first real transaction. By this past summer, that had fallen to 18 minutes. The friction is gone. The question now is not whether AI agents will become a significant force in crypto markets — they already are — but how quickly 138 million payments becomes 138 billion.
Sources:
- CoinDesk — AI Agents Make 138 Million Crypto Payments as U.S. Dominates Spending on AI Models
- Decrypt — AI Agents Made 138 Million Crypto Payments in a Year. Most Were Tiny
- Chainalysis — 2026 Geography of Crypto: North America Report
- Yellow Research — AI Compute Demand Is Outpacing Supply, And Crypto Networks Are Stepping In
- Decrypt — What Is x402? How AI Agents Pay With Crypto
