AI agents need payment rails before they need shopping demos

Circle is the most important company in AI infrastructure that almost nobody is covering. Not because it sells the best AI model or runs the most popular chatbot, but because 98.6 percent of all money flowing between AI agents settles through a single digital currency it issues: USDC, a stablecoin designed to hold a steady dollar value on public blockchains. If that sounds like a concentration problem, it is.

The agent payments market exists in a gap between what crypto advocates promised and what traditional finance ever addressed. When researchers count it, they get numbers so far apart they tell you more about methodology than about commerce: $73 million in headline volume over twelve months, $1.6 million after filtering out wallets that trade with themselves, $24 million by a different methodology. What all three estimates agree on is that the real market is tiny, growing from a low base, and built on infrastructure nobody has stress-tested at scale. The money is real. The infrastructure is not yet.

USDC dominates because the alternative is nothing. Traditional payment rails charge a per-transaction fee that makes sub-cent payments — the kind an AI agent might make to buy computing time, access a dataset, or call a software tool — economically irrational. Blockchain-based systems don't charge that fee. AI agents also cannot open bank accounts because they cannot satisfy Know Your Customer requirements: no passport, no selfie, no human behind the transaction. Stablecoins are the only option that works today, and USDC has the widest adoption across the payment protocols that matter for machine-to-machine commerce. According to Keyrock, which tracks on-chain agent payment activity, 98.6 percent of machine payments settle in USDC.

That single-issuer concentration has drawn more than $8 billion in acquisitions from incumbents over the past twelve months, Keyrock found. Stripe, Visa, and Mastercard are all building competing infrastructure, betting that if agents become normal commercial actors, the payment layer beneath them is worth owning. Stripe designed its next-generation payment standard with OpenAI and Anthropic in the room, according to Emerging Fintech. Visa has called itself the "hyperscaler of payments" for AI agents. The technical argument is coherent: 76 percent of agent payment activity falls below the $0.30 card-fee floor, making traditional rails impractical for the sub-cent transactions that dominate machine-to-machine commerce, Keyrock found. x402, a Coinbase-backed protocol for machine payments, processed more than 100 million transactions in its first six months.

The market-size problem is not just counting methodology. It is that the volume comes largely from developers testing API integrations, not from agents making commercial purchases at scale. Coinbase CEO Brian Armstrong told Emerging Fintech that AI agents "cannot meet Know Your Customer requirements and therefore cannot use traditional banking infrastructure" — which is accurate, but also means the current ecosystem runs on a workaround for a legal category that does not yet formally exist. Whether that workaround becomes permanent infrastructure or gets displaced by something regulated is the open question.

Three major regulatory frameworks hit enforcement in mid-2026: the EU's MiCA framework, the US GENIUS Act, and the EU AI Act, according to Keyrock. None directly address autonomous machine-to-machine transactions — payment laws written for human-initiated commerce do not automatically cover software acting on its own behalf. Whether an agent can enter a binding payment contract, who is liable when an agent pays the wrong party, and how consumer protections apply to machine decisions are questions the current rules do not clearly answer.

Over 104,000 agents are registered across various directories and registries, Keyrock found. Most are not yet moving meaningful commercial volume. The companies building for that volume — Stripe, Visa, Circle, Coinbase — are not waiting for regulatory clarity. They are building on the assumption that when the volume arrives, the infrastructure they are laying now will be the layer that matters. The bet is reasonable. It is also not yet proven, and it is entirely concentrated on one company's coin, one regulatory framework, and a set of technical standards that have never processed large-scale commercial traffic.

What to watch: whether actual commercial micropayments begin displacing test traffic in the next two quarters, and whether any regulator treats autonomous machine transactions as a distinct legal category requiring its own rules. If that happens, the infrastructure that exists now will be load-bearing. If it doesn't, the $8 billion in acquisitions will have bought an option on a market that has not arrived yet.