AI agents settled $73 million across 176 million blockchain transactions in the 12 months ending April 2026, with 98.6% of volume denominated in USDC, according to a Keyrock report published May 21. The average transaction size stabilized near $0.48 — well below the $0.30 fixed-fee floor charged ...
"76% of agent transactions fall below the 30-cent fixed-fee floor common in card payments." — Keyrock, May 2026 AI Agent Payments Report
AI agents settled $73 million across 176 million blockchain transactions in the 12 months ending April 2026, with 98.6% of volume denominated in USDC, according to a Keyrock report published May 21. The average transaction size stabilized near $0.48 — well below the $0.30 fixed-fee floor charged by Visa and other card networks, rendering traditional payment rails economically unviable for machine-to-machine commerce.
The infrastructure race to capture this emerging payment layer has drawn four of the largest names in technology and finance: Coinbase, Stripe, Google, and Visa each launched competing or complementary protocols between January and May 2026. Amazon Web Services entered the market on May 7 with Bedrock AgentCore Payments, integrating both Coinbase's x402 protocol and Stripe's Privy wallet to give enterprise AI agents stablecoin payment capabilities out of the box. Juniper Research projects $8 billion in agentic spend for 2026, scaling to $1.5 trillion globally by 2030.
The numbers remain small in absolute terms — AI agent volume represents 0.0001% of annual stablecoin settlement — but the structural logic is clear: autonomous software agents need payment rails that support sub-cent transactions, instant settlement, and no account creation. Card networks cannot provide this. Blockchain rails can.
Keyrock's May 2026 report, co-published with Coinbase, Tempo, and Virtuals Protocol, provides the most granular dataset available on autonomous agent payment behavior. Key findings:
The transaction curve showed significant volatility. Daily agent transactions peaked at 731,000 in December 2025 before declining 92% to 57,000 daily in February 2026. The report attributes this to the end of incentivized testnet programs rather than a structural decline in agent adoption.
Coinbase's x402 protocol, the largest single settlement layer, processed 140 million of those transactions — approximately $43 million in volume — over nine months. As of April 21, roughly 69,000 active AI agents were registered on x402, up from near-zero at launch in late January 2026.
The dominant use case is API access: agents purchasing blockchain analytics data, AI inference, cloud compute cycles, and paywalled content. These are services priced in fractions of a cent per call, where even Visa's $0.30 minimum fee would exceed the value of the underlying transaction.
Between January and May 2026, four major payment protocol frameworks emerged for machine-to-machine commerce. Each reflects a different strategic bet on where value will concentrate.
Coinbase — x402 Protocol
x402 repurposes the HTTP 402 "Payment Required" status code as a native payment layer. When an AI agent requests a resource, the server returns a 402 response with a payment specification. The agent evaluates the cost, executes a USDC micro-payment on-chain, and resubmits the request with a payment receipt. Settlement occurs in under two seconds on Base at approximately $0.0001 per transaction.
As of May 2026, x402 supports ERC-20 payments on Base, Polygon, Arbitrum, World, and Solana. Coinbase charges zero protocol fees. The company launched Agent.market, an AI agent app store, with service providers including OpenAI, Bloomberg, CoinGecko, AWS Lambda, and Coinbase's own RAT analytics tool.
According to Keyrock's analysis, Coinbase spans five of six agentic payment stack layers: settlement (Base), wallets (AgentKit), payment protocol (x402), marketplace (Agent.market), and governance (as a co-author of AP2). This vertical integration is the most aggressive positioning in the space.
Stripe — Machine Payments Protocol (MPP)
Stripe launched MPP as an open standard supporting both stablecoin and fiat settlement. The protocol uses Shared Payment Tokens (SPTs) that allow agents to pay with USDC, credit cards, or buy-now-pay-later instruments through a single interface. Stripe acquired Privy to provide wallet infrastructure, and its Tempo blockchain — which went live in March 2026 — serves as an additional settlement layer.
Stripe's positioning is deliberately hybrid: it does not require merchants to accept crypto, but routes stablecoin payments when both parties support them. This bridges the gap between crypto-native agents and traditional commerce infrastructure.
Visa — Intelligent Commerce Connect
Visa launched Intelligent Commerce Connect on April 8, 2026, as a protocol-agnostic orchestration layer. The platform integrates Visa's Trusted Agent Protocol, Stripe's MPP, OpenAI's Agentic Commerce Protocol, and Google's Universal Commerce Protocol through a single API. The strategy leverages Visa's existing merchant network rather than competing on settlement economics.
The structural limitation remains: Visa's card economics require a minimum fee floor that exceeds the value of most agent transactions. The platform is better suited for higher-value agentic purchases — travel bookings, software subscriptions, physical goods — than for the sub-cent API calls that dominate current volume.
Google — Agent Payments Protocol (AP2)
Google published AP2 as an open framework intended to standardize how agents discover, negotiate, and execute payments across providers. Details on adoption and transaction volume are limited. Google's primary strategic interest appears to be ensuring its Gemini and Vertex AI agent ecosystem can transact across all major payment rails rather than owning a proprietary settlement layer.
Amazon Web Services' May 7 launch of Bedrock AgentCore Payments marked the entry of the largest cloud provider into agentic stablecoin infrastructure. The system integrates directly into Amazon Bedrock — AWS's managed AI model service — allowing enterprise AI agents to make autonomous purchases through a single API call.
Key technical specifications:
AWS's initial focus is micropayments for APIs, data feeds, and paywalled content. The company has stated plans to expand to larger transactions including hotel bookings, travel, and merchant payments. Warner Bros. Discovery is among early testers.
The strategic significance: AWS hosts an estimated 31% of global cloud infrastructure. By embedding stablecoin payment rails into its managed AI service, Amazon normalizes crypto settlement for enterprise customers who may have no prior blockchain exposure. The agent handles the crypto; the enterprise sees an API call.
A parallel infrastructure development is ERC-8004, the Ethereum-based identity standard for autonomous AI agents. Ratified and deployed to Ethereum mainnet on January 29, 2026, the standard was authored by Marco De Rossi (MetaMask), Davide Crapis (Ethereum Foundation), Jordan Ellis (Google), and Erik Reppel (Coinbase), with input from over 80 teams.
ERC-8004 uses three on-chain registries:
As of May 2026, approximately 24,000 agents have registered with ERC-8004, with projections of 130,000 by year-end. BNB Chain leads adoption with 34,000 active agents, followed by Base (16,500) and Ethereum mainnet (14,000).
The standard addresses a foundational problem in machine-to-machine commerce: counterparty verification. When two agents transact autonomously, neither can rely on human judgment to assess trustworthiness. On-chain identity and reputation registries provide a machine-readable trust layer — the equivalent of a credit score for software agents.
Three major regulatory deadlines converge between June and August 2026:
None of these frameworks contain provisions for autonomous machine-to-machine transactions. According to Keyrock's analysis, there is no regulatory clarity on several fundamental questions: Can an AI agent hold a stablecoin balance? Who bears liability when an agent makes a payment error? Does an autonomous agent transaction constitute a "payment service" requiring a license?
The MiCA and GENIUS Act diverge on licensing, custody, and compliance architecture. "Compliant" in one jurisdiction does not translate cleanly to the other. For agentic commerce platforms operating globally, multi-jurisdictional compliance introduces cost and complexity that could slow adoption.
The gap is structural: payment regulation assumes a human initiator. Agent commerce has no human in the loop. Regulators have not yet addressed this.
Applying an economic-value-distribution lens to the agentic payment stack reveals a familiar pattern: infrastructure operators capture disproportionate value relative to protocol developers.
Settlement layer: Base (Coinbase's L2) collects sequencer fees on every x402 transaction. At $0.0001 per transaction and 165 million transactions processed, direct fee revenue is approximately $16,500 — trivial in isolation. The value proposition is network effects: every agent that settles on Base increases demand for USDC on Base, which increases Coinbase's custody and on-ramp revenue.
Wallet layer: Coinbase's AgentKit and Stripe's Privy capture value through fiat on-ramp fees when users fund agent wallets via debit card. Stripe's standard processing fee of 2.9% + $0.30 applies to fiat deposits, creating a significantly higher margin than the on-chain transaction itself.
Stablecoin issuer: Circle captures the largest share of economic value. With 98.6% of agent payments in USDC, Circle earns yield on the underlying treasury reserves backing every dollar in agent wallets. At current rates, this is approximately 4.3% annually on idle balances — a passive revenue stream that scales linearly with agent adoption.
Cloud provider: AWS charges for Bedrock API usage regardless of payment method. AgentCore Payments is a feature that increases Bedrock utilization, not a standalone revenue line.
The pattern: direct on-chain payment fees are near-zero, but every adjacent service — custody, on-ramp, stablecoin reserves, cloud compute — generates margin. Value accrues to the infrastructure surrounding the transaction, not the transaction itself.
The $73 million in annual agent settlement volume is a rounding error against the $46 trillion stablecoin market. The infrastructure buildout is not. In five months, the four largest payment and technology companies in the Western world deployed competing systems to capture a market that does not yet meaningfully exist at scale.
The bet is directional: Juniper Research forecasts $8 billion in agentic spend this year, rising to $1.5 trillion by 2030. McKinsey's estimate runs higher, at $3 to $5 trillion. If either projection proves accurate, the payment layer chosen by the first generation of production AI agents will become deeply entrenched.
The data suggests that crypto rails — specifically USDC on L2 networks — have a structural advantage for micropayments that card networks cannot match. The question is whether that advantage extends to higher-value transactions, where Visa and Stripe's existing merchant relationships and consumer protections matter more than sub-cent fees.
For now, the infrastructure is being laid faster than the demand. That is either prescient positioning or premature investment. The answer depends on how quickly AI agents move from buying API calls to buying everything else.