Somewhere, right now, an AI agent is paying another AI agent a fraction of a cent for the right to query a database. No bank is involved. No card network is processing the transaction. No human approved it. The payment settles in roughly two seconds on a blockchain, costs less than one-tenth of a cent in fees, and leaves a permanent, auditable record on-chain.
This is not science fiction. It is infrastructure, and it is already running at scale.
A new report from market intelligence firm Keyrock — titled “Who Pays the Agent?” — has put hard numbers on a trend that blockchain analysts have been forecasting for years. Between May 2025 and April 2026, AI agents settled $73 million across 176 million blockchain transactions. The overwhelming majority of those payments were denominated in USDC. The overwhelming majority of them would have been economically impossible on Visa or Mastercard.
The implications for blockchain’s role in the global economy are difficult to overstate.
The 30-Cent Wall That Card Networks Built
To understand why AI agents are defaulting to crypto rails, you need to understand a number that has quietly defined the economics of digital payments for decades: 30 cents.
That is the approximate fixed-fee floor on most card transactions. When you buy a $50 item online and Visa charges the merchant 2.9% + $0.30, the $0.30 is the structural minimum that makes the economics of card processing work. It covers fraud protection, chargebacks, authorization infrastructure, and the intermediary layers that sit between your card and the merchant’s bank.
Thirty cents is fine when you are buying a coffee. It is catastrophic when you are an AI agent paying one-tenth of a cent to retrieve a weather data point, a quarter-cent for a single API call, or three cents for a block of processing time.
The Keyrock report found that 76% of all AI agent transactions fall below that 30-cent floor. Most payments in the dataset ranged between one and ten cents. These are not edge cases — they are the dominant transaction type in machine-to-machine commerce. Every single one of them is uneconomical on traditional card rails before the payment even begins.
Card networks were designed by humans, for humans, transacting at human scale. They are structurally incapable of supporting the micropayment economy that autonomous AI systems require.
What the Keyrock Report Actually Found
The “Who Pays the Agent?” report is the most comprehensive data survey of agentic payments published to date, and its findings deserve careful attention.
$73 million across 176 million transactions in one year sounds modest against the backdrop of Visa’s $14.5 trillion in annual processing volume. But context matters: this is the first year of serious agentic payment infrastructure. The on-chain agent ecosystem barely existed at the scale required to generate this data twelve months ago.
The concentration is striking. 98.6% of all machine payments in the dataset settled in USDC — Circle’s dollar-pegged stablecoin. That is not a market with competing alternatives finding equilibrium. That is a market that has made a decisive, near-unanimous choice about settlement currency.
The cost differential explains much of it. Settlement in USDC on Base or Stripe’s Tempo blockchain costs fractions of a cent — the Nevermined research team puts the figure at under $0.0001 per transaction. Sub-second finality means an agent does not wait for banking hours or clearing windows. The payment pushes directly from sender to receiver, settles on-chain, and the workflow continues.
For context on scale: stablecoin transaction volume across all use cases reached $33 trillion in 2025, growing 72% year-over-year. AI agent payments are currently a rounding error against that number. The interesting question is how long that remains true.
Three Companies Racing to Own Machine Money
The commercial stakes of becoming the default payment layer for AI agents are enormous, and three of the world’s largest technology companies are already building competing infrastructure. Their approaches differ in ways that reveal fundamentally different visions of how machine commerce should work.
Coinbase and the x402 Protocol
Coinbase’s entry is the most crypto-native of the three. Its x402 protocol extends the HTTP 402 “Payment Required” status code — a standard that has existed in web specifications since 1996 but was never practically implemented — into a functional payment primitive for AI agents.
The mechanism is elegant: when an AI agent requests a resource, the server can respond with an HTTP 402 header specifying a USDC amount and a blockchain address. The agent pays, the server verifies the on-chain transaction, and the resource is delivered. No account creation. No subscription setup. No API key management. Pay per use, programmatically, in seconds.
Nevermined’s statistics show x402 has already processed over 35 million transactions on Solana, handling roughly $600 million in annualized volume. In a single 30-day window, $24 million flowed through the protocol across 94,000 buyers and 22,000 sellers — with 24,000 agents registered via the ERC-8004 identity standard since January 2026.
Stripe’s Machine Payments Protocol
Stripe is approaching the problem from its position as the incumbent infrastructure layer for internet commerce. Its answer — the Machine Payments Protocol (MPP), built on its Tempo blockchain — is designed to translate agentic payments into the financial framework that enterprises already trust.
Where x402 is raw and cryptographic, MPP layers in compliance hooks, spending controls, and audit trails that make it more palatable to corporate finance teams. Stripe is essentially arguing that AI agents will need enterprise payment rails, not just crypto rails — and that those two things can coexist on blockchain infrastructure.
Google’s AP2 System
Google’s AP2 takes a third approach: delegated spending authorization. Rather than letting agents pay autonomously, AP2 focuses on controlling what agents can spend and on whose behalf — solving the liability and oversight problem that sits at the edge of every conversation about autonomous machine payments.
The reach is significant. Google has already attracted 60+ partner organizations to AP2, including PayPal, Coinbase, and Mastercard. That coalition represents a bet that the future of machine payments runs through credentialed identities and delegated permission, not anonymous crypto wallets.
Visa’s Tokenized Credentials
Visa’s play is different in kind. Rather than building a competing protocol, Visa is experimenting with tokenized credentials — essentially bringing card-network trust anchors into the blockchain environment. The thesis: enterprises that already trust Visa infrastructure will adopt agentic payments faster if those payments carry familiar compliance guarantees.
Whether tokenized card credentials can compete with the raw economics of native stablecoin settlement remains an open question. The fee structure of card networks does not change just because the credentials are on-chain.
Why USDC Dominates — And Why That Matters
The 98.6% USDC dominance figure in the Keyrock report is one of the most commercially significant data points in the document, and it cuts in two directions simultaneously.
For Circle, it is an extraordinary competitive moat. The stablecoin market is crowded — USDT accounts for $13.3 trillion in annual volume, and the overall stablecoin supply has reached $310 billion — but in the specific domain of machine payments, USDC has established near-total lock-in. When Coinbase builds x402, it settles in USDC. When institutional payment flows move on-chain, they predominantly move in USDC.
This has structural logic: USDC is regulated, audited, and dollar-backed in a way that makes it acceptable to enterprise compliance functions. AI agents operating inside corporate environments need a stablecoin that can survive a board-level review, and USDC’s regulatory posture — particularly its positioning under the incoming U.S. GENIUS Act — gives it an edge that pure-crypto alternatives lack.
The risk is concentration. A settlement layer where 98.6% of activity runs through a single issuer introduces systemic dependency that the broader blockchain ecosystem should take seriously. If Circle encounters a regulatory or operational crisis, the machine payment layer built on its infrastructure faces a single point of failure. Diversification across USDT, DAI, or emerging alternatives would create resilience — but the data suggests that diversification has not happened yet.
The $15 Trillion Prize
The dollar figures in the current Keyrock dataset are meaningful, but they are not the reason this story matters. The reason this story matters is what the projections say about where machine commerce is heading.
Gartner estimates that AI agents could intermediate $15 trillion in purchases by 2028. To calibrate that number: it represents roughly the current annual GDP of China. It is more than Visa and Mastercard combined process in a year.
McKinsey puts the retail agentic commerce opportunity at $3 to $5 trillion by 2030 — a narrower estimate of the consumer-facing segment alone, excluding enterprise and B2B machine-to-machine flows.
The AI agents market itself is growing at a 49.6% compound annual growth rate, projected to expand from $7.63 billion in 2025 to nearly $183 billion by 2033. Every new agent that needs to pay for compute, data, or services adds to the demand for payment infrastructure that can handle micropayments at machine speed.
B2B stablecoin payments have already grown 733% year-over-year, hitting $226 billion annually in 2025. The infrastructure for sending money natively on-chain is maturing faster than regulators and incumbents anticipated. The American Banker’s payment analysts have estimated that AI agents and stablecoins could displace 20% of traditional card-based settlement volume by end of 2026 — a projection that seemed aggressive six months ago but looks more credible against the Keyrock data.
At these growth rates, the $73 million settled in the Keyrock study period is a rounding error on the transaction volumes projected for 2028. The infrastructure that wins machine payments in 2026 is the infrastructure that captures trillions of dollars in settlement by the end of the decade.
Regulation Has Not Caught Up — And That Is a Risk
Three major regulatory frameworks are converging around mid-2026 with direct relevance to stablecoins and AI systems: the EU’s MiCA regulation, the U.S. GENIUS Act, and the EU AI Act. None of them, according to the Keyrock report, directly address autonomous machine-to-machine transactions.
The gaps are consequential. When an AI agent pays another AI agent for a service, several foundational legal questions remain unanswered: Who is the legal counterparty? Who bears liability if a transaction is fraudulent? How does KYC/AML compliance apply to a non-human entity initiating payments? Is a machine-executed payment subject to the same consumer protection rules as a human-initiated one?
These are not hypothetical questions. They are live operational risks for any enterprise deploying AI agents with payment capabilities today. The absence of regulatory clarity creates legal exposure that may slow institutional adoption even as the technical infrastructure matures.
Google’s AP2 framework — with its emphasis on delegated authorization and identity — can be read partly as a hedge against this regulatory uncertainty. If an agent always acts under explicit human delegation, the liability question has a cleaner answer: the delegating human is responsible. That may prove to be a powerful selling point for risk-averse enterprises once regulators begin drawing lines.
The GENIUS Act, which establishes a federal licensing framework for stablecoin issuers in the U.S., will clarify some of this by requiring issuers to meet reserve and compliance standards. USDC’s dominant position in machine payments may partly reflect anticipation of that framework — Circle has spent years building the compliance posture that GENIUS codifies.
Blockchain’s Structural Advantage, Explained
For readers who follow blockchain’s technical evolution, the AI agent payment story confirms something that the ecosystem has argued about for years but struggled to demonstrate at scale: blockchains are not primarily useful as money for humans. They are useful as money for machines.
The properties that make blockchain payments awkward for everyday retail — the requirement to hold a wallet, the irreversibility of transactions, the absence of chargebacks, the strangeness of interacting with cryptographic keys — are precisely the properties that make blockchain payments ideal for autonomous systems.
AI agents do not need chargebacks. They do not need fraud protection designed for humans who might lose their cards. They do not need banking hours. They need deterministic settlement, programmable payment logic, sub-cent transaction costs, and a settlement layer that operates at machine speed without human intermediaries. Blockchain provides all four.
The x402 protocol’s elegance lies in how naturally it maps blockchain payments onto the HTTP communication layer that the internet already uses. An agent requesting a resource and receiving a payment challenge is not doing something exotic — it is extending a protocol that web servers and clients have used for decades, now powered by on-chain settlement rather than card networks.
What the Keyrock report documents is not a blockchain use case being forced onto AI systems. It is AI systems organically choosing blockchain because blockchain is the infrastructure that fits the job.
What Comes Next
The machine payment layer is still early. The $73 million in the Keyrock dataset is real, but it is also fragmented across multiple protocols, chains, and agent architectures. The ecosystem that captures the multi-trillion-dollar machine economy will need to solve several problems that remain open:
Agent identity is the most foundational. For machine payments to scale into enterprise environments, agents need verifiable identities that comply with KYC frameworks without requiring human-speed verification for every transaction. ERC-8004 and Google’s AP2 are early attempts at this, but no standard has achieved broad adoption.
Interoperability between competing protocols — x402, MPP, AP2, and Visa’s tokenized credentials — remains unsolved. An agent ecosystem fragmented across four incompatible payment layers creates friction that slows adoption. The market will likely converge, but which protocol wins is not yet clear.
The USDC concentration risk will need to be addressed before the machine payment layer can be considered robust. A settlement layer with a single issuer commanding 98.6% share is a systemic risk, not a feature.
Regulatory clarity will come — it always does, eventually. The question is whether it arrives before or after a significant incident forces legislators to act reactively.
The direction, however, is not in question. Card networks were built for humans making occasional purchases in a world where payment infrastructure ran on phone lines and magnetic stripes. That world is not coming back. The economy that AI agents are building operates at a different speed, a different granularity, and a different scale — and blockchain is the infrastructure that has the right physics for the job.
The $73 million is just the beginning.
SGNChain covers blockchain infrastructure, DeFi protocols, and the on-chain economy. For more analysis on stablecoin markets and payment layer evolution, explore our DeFi topic hub and Real-World Assets coverage.