AI agents settled $73 million across 176 million blockchain transactions in the twelve months ending May 2026, according to a Keyrock report published May 21. The figure is small relative to global payments volume, but the infrastructure race it has triggered is not: Amazon Web Services, Coinbase...
"76% of agent transactions fall below the 30-cent fixed-fee floor common in card payments… traditional rails are structurally incapable of serving the agentic economy." — Johan Östman, Head of Research, Keyrock
AI agents settled $73 million across 176 million blockchain transactions in the twelve months ending May 2026, according to a Keyrock report published May 21. The figure is small relative to global payments volume, but the infrastructure race it has triggered is not: Amazon Web Services, Coinbase, Stripe, Google, Visa, and MoonPay have each launched competing payment protocols for autonomous software agents in the past 90 days. Nearly all settlement — 98.6% — occurs in USDC on Coinbase's Base network and Solana.
The average AI agent transaction is $0.48. At that size, 76% of agent payments fall below Visa's $0.30 fixed-fee floor, making traditional card rails economically unviable for machine-to-machine commerce. This structural mismatch is driving infrastructure spend into stablecoin-based micropayment protocols. Three regulatory frameworks — MiCA, the GENIUS Act, and the EU AI Act — reach enforcement between July 1 and August 2, 2026. None contain provisions for autonomous machine-to-machine transactions, creating a liability vacuum that the industry is building into regardless.
Keyrock's May 2026 report provides the most granular public dataset on AI agent payment activity to date. The headline figures:
The transaction profile reveals what agents are actually buying: API calls, data feeds, compute inference, and paywalled content. These are small, frequent, programmatic purchases — the economic signature of software consuming services from other software.
For context, Visa processed approximately $15 trillion in payment volume in 2025. AI agent volume of $73 million represents roughly 0.0005% of that figure. The significance is not in the current scale but in the growth trajectory and the infrastructure investment it has attracted from Visa's own partners and competitors.
Between March and May 2026, six major technology and financial services companies launched or expanded payment infrastructure for AI agents. The speed and breadth of deployment suggest institutional consensus that machine-to-machine payments represent a durable market, not a speculative narrative.
| Company | Protocol/Product | Launch Date | Settlement Layer | Key Feature | |---------|-----------------|-------------|-----------------|-------------| | Coinbase | x402 | 2025 (expanded 2026) | Base, Solana, Ethereum, Arbitrum, Polygon | HTTP-native payment standard | | Stripe/Paradigm | Machine Payments Protocol (MPP) | March 18, 2026 | Tempo L1 | Streaming micropayments | | Google | Agent Payments Protocol (AP2) | Q1 2026 | Payment-agnostic | FIDO Alliance governance | | AWS | Bedrock AgentCore Payments | May 7, 2026 | Base (via Coinbase) | Native cloud integration | | Visa | Intelligent Commerce Connect + CLI | March–April 2026 | Multi-protocol | Legacy rail bridging | | MoonPay | MoonAgents Card | May 2026 | Solana | Mastercard-accepted virtual debit |
Each entrant occupies a different position in the stack, but the competitive overlap is significant. Coinbase and Stripe each span five of six infrastructure layers, according to Keyrock. The coming 12 months will determine whether the market consolidates around one or two dominant protocols or fragments into interoperable standards.
Amazon Web Services launched Bedrock AgentCore Payments in preview on May 7, 2026, in partnership with Coinbase and Stripe. The product integrates stablecoin payment capability directly into AWS's AI agent hosting infrastructure.
Technical parameters: settlement occurs on Coinbase's Base network in USDC, with approximately 200-millisecond finality at sub-cent transaction costs. Developers choose between a Coinbase wallet or Stripe's Privy wallet as the payment connector. Spending limits are enforced per session, constraining agent budgets at runtime.
The initial scope is narrow — micropayments for APIs, data feeds, and paywalled content. AWS has stated plans to expand to larger transactions including hotel bookings, travel reservations, and merchant payments. The significance is distribution: AWS controls an estimated 31% of the global cloud infrastructure market. Embedding stablecoin payments into the default AI development environment normalizes crypto rails for a developer population that may have no prior exposure to blockchain technology.
Coinbase's x402 protocol repurposes the HTTP 402 "Payment Required" status code — defined in the original HTTP specification but never widely implemented — to embed stablecoin payments directly into web requests. The protocol, co-developed with Cloudflare, charges zero protocol fees.
Scale as of April 2026: approximately 69,000 active AI agents on x402 have processed over 165 million transactions totaling $50 million in volume. Sub-2-second settlement. Transaction cost of approximately $0.0001.
Coinbase has extended x402 into a broader ecosystem. Agent.market, launched in May 2026, functions as an AI agent app store built on x402, offering permissionless stablecoin rails across seven service categories. The x402 Foundation, co-governed by Coinbase and Cloudflare, oversees protocol development.
The protocol supports Base, Ethereum, Arbitrum, Polygon, and Solana. It is the settlement layer underlying AWS Bedrock AgentCore Payments, giving Coinbase infrastructure presence inside Amazon's cloud platform.
Stripe and Paradigm launched the Machine Payments Protocol (MPP) on March 18, 2026, simultaneously with the mainnet launch of Tempo, a payments-focused Layer 1 blockchain. MPP is positioned as a protocol built specifically for non-human transactors.
MPP supports multiple settlement rails: stablecoins on Tempo, fiat via Stripe and Visa card rails, and Bitcoin Lightning via Lightspark. This multi-rail design differentiates it from crypto-only alternatives. Streaming payments, introduced at Stripe Sessions 2026, allow real-time billing against per-token, per-second, or per-call usage — a billing model that mirrors how cloud compute is priced.
Early adopters include Browserbase (headless browser sessions billed per use), PostalForm (physical mail printing and delivery), and Prospect Butcher Co. (food ordering). Visa contributed to the MPP specification, developing standards for agent payments via credit and debit cards.
Google's Agent Payments Protocol (AP2) takes a different approach. Rather than building a settlement layer, AP2 is a payment-agnostic coordination standard designed as an extension of the Agent-to-Agent (A2A) protocol and Model Context Protocol (MCP). It specifies how agents and services negotiate, authorize, and execute payments regardless of the underlying rail.
The April 2026 v0.2.0 release introduced "Human Not Present" payments — allowing agents to autonomously complete purchases without real-time human approval. Google has assembled a coalition of more than 60 organizations behind the standard, including Adyen, American Express, Ant International, Coinbase, Etsy, Mastercard, PayPal, Revolut, Salesforce, and Worldpay.
In May 2026, Google donated AP2 governance to the FIDO Alliance, the industry body best known for the FIDO2/WebAuthn authentication standards. The move signals intent to position AP2 as a neutral, industry-governed layer above competing settlement protocols.
Visa launched Intelligent Commerce Connect on April 8, 2026, describing it as the first major payment infrastructure enabling AI agents to browse, select, and pay for goods autonomously. The platform is protocol-agnostic, supporting Visa's Trusted Agent Protocol, Stripe's MPP, OpenAI's Agentic Commerce Protocol, and Google's Universal Commerce Protocol through a single integration.
Separately, Visa Crypto Labs launched a CLI (command-line interface) in March 2026, currently in closed beta, enabling AI agents to pay for API calls without pre-set accounts. Visa also expanded its stablecoin settlement program to nine blockchains and partnered with Lightspark for stablecoin and Bitcoin-backed Visa debit cards in over 100 countries.
MoonPay launched the MoonAgents Card in May 2026 — a virtual Mastercard debit card that allows AI agents to spend stablecoins directly from on-chain wallets at the point of transaction. The card converts crypto to fiat in real-time. Available in the UK and Latin America, with US and EU availability planned. MoonPay CLI, the underlying interface, has processed more than 4 million tool calls since launch, with the second million processed in seven days versus 30 days for the first million.
Keyrock's data shows 98.6% of AI agent transaction volume settling in USDC. This creates single-issuer dependency on Circle, which went public as CRCL. If Circle experienced a de-peg event, regulatory action, or technical failure, nearly all machine-to-machine payment infrastructure would be affected simultaneously.
The concentration has structural causes. USDC has the deepest liquidity on Base (Coinbase's L2), Solana, and Ethereum — the three chains where AI agent activity is concentrated. x402, the dominant agent payment protocol, is built by Coinbase, which also issues USDC jointly with Circle. The infrastructure is vertically integrated in ways that reinforce USDC dominance.
No competing stablecoin has made significant inroads into the agent payments market. USDT (Tether) accounted for the remaining 1.4%. Euro-denominated stablecoins, despite MiCA-driven growth in human-facing payments, have negligible presence in machine-to-machine transactions.
Three major regulatory frameworks reach enforcement within weeks of each other in mid-2026:
None contain explicit provisions for autonomous machine-to-machine transactions. The regulatory frameworks assume human principals on both sides of a transaction. When an AI agent autonomously purchases API access or data feeds, existing frameworks provide no clear answer to basic questions: Who bears liability for an unauthorized agent purchase? Can an agent's transaction be reversed? Who is the "customer" for KYC purposes?
With credit card payments, merchants bear chargeback risk and consumers receive legal protection. With stablecoin settlement, once funds land in a merchant wallet, they cannot be recalled. The risk moves entirely to the consumer — or, in the agent context, to whoever deployed the agent. Current liability regimes assume human intent and direct causation. These assumptions break down when autonomous agents make independent purchasing decisions.
According to the IMF's May 2026 note on agentic AI in payments, none of the four major jurisdictions studied have explicit regulatory provisions for autonomous AI agents participating in digital asset markets.
The AI agent payment market is a $73 million niche today. The infrastructure deployed to serve it — by AWS, Coinbase, Stripe, Google, Visa, and MoonPay — is built for trillions. That gap between current volume and infrastructure investment reflects a bet that machine-to-machine commerce will follow the same adoption curve as cloud computing: negligible for years, then dominant within a decade.
The economic logic is structural. When 76% of transactions are too small for card rails, stablecoin settlement is not a philosophical choice — it is an engineering requirement. The question is not whether crypto rails will serve AI agents, but which protocol stack captures the settlement layer and the fee economics that come with it.
The regulatory vacuum is the primary unresolved risk. By August 2, 2026, three major frameworks will be in force and none will address the entities generating the fastest-growing segment of on-chain transaction volume. Whether regulators treat AI agents as extensions of their deployers, as independent economic actors, or as something else entirely will determine which of these infrastructure bets pays off — and which creates liability that no protocol can absorb.