AI agents need to spend money. That much is clear. CB Insights projects agentic commerce at $135B today, growing to $4.7T by 2030. Morgan Stanley expects agents to capture 10 to 20% of US ecommerce by the end of the decade.

But how do they pay?

There are three paths. Cards. Account to account. Stablecoins. They are not mutually exclusive. Each has a different cost structure, a different trust model, and a different set of winners. The interesting question is not which one wins. It is which one wins for which use case. And who captures the economics.

The market map

47 companies across 10 categories are building the agentic commerce stack. The payment rails layer splits cleanly into three columns. The infrastructure layers above and below span all three.

Cards: paying for safety

The card rail is the one everyone knows. Visa and Mastercard touch roughly 99% of merchants. They carry six decades of consumer protection law. If an agent goes rogue and buys 500 toaster ovens, you chargeback. That is the value proposition.

But you pay for it.

Schemes, issuers, and acquirers all skim off the top. Interchange sits at roughly 1 to 3%. Consumers get rebates off their cards. Cashback, points, miles. These are funded by the spread the card network charges merchants for reliability and dispute management. The consumer thinks the card is free. The merchant pays. This model has worked for decades and it will keep working for agent commerce too.

Visa shipped the Trusted Agent Protocol (TAP) to cryptographically verify that an agent is legitimate before a transaction goes through. Mastercard launched Agent Pay, completing Europe's first live AI-agent payment with Santander in early 2026. Both create a "card-like" token for agents. Same rails, same dispute resolution, new actor.

The startups here are Basis Theory ($33M Series B, leads the 20-member Agentic Commerce Consortium) and Nekuda ($5M seed backed by both Amex Ventures and Visa Ventures). Both card networks writing seed checks into the same startup tells you something.

Modern acquirers like Stripe are poised to gain the most share on this rail. Stripe co-authored the Agentic Commerce Protocol (ACP) with OpenAI. ChatGPT Instant Checkout ran on Stripe rails. When the agent commerce wave hits mainstream consumer transactions, Stripe is the default checkout layer.

The weakness is structural. Cards cannot price sub-cent micropayments. An agent that needs to pay $0.001 for an API call cannot use a card. The authentication flows assume a human is present. For high-value, human-authorized purchases, cards will dominate. For everything else, you need a different rail.

Bank accounts: the inevitable shift

Account-to-account payments bypass cards entirely. Open banking in Europe. ACH and FedNow in the US. UPI in India. Pix in Brazil. The infrastructure is available today.

The economics tell the story. Cards charge 1 to 3% interchange. A2A costs near zero. For any business where margins matter, and margins always matter, removing interchange is an obvious move.

This rail gets less attention than cards or crypto. But it is already well-capitalized. Plaid ($734M raised), GoCardless ($540M), Trustly ($970M), TrueLayer ($272M), Tink (acquired by Visa for EUR 1.8B). GoCardless launched an MCP server in February 2026, the first A2A player with native LLM integration. Agents can initiate bank payments using natural language.

Variable Recurring Payments (VRPs) in UK open banking are a natural primitive for agent authorization. They are essentially bank-native mandates. An agent gets a pre-authorized spending envelope from the user's bank, then transacts within those limits without per-transaction consent. This is the same concept as Google's AP2 protocol.

The bank account is still needed. It is a fiat wallet. You need somewhere to store funds, and banks and e-money institutions are the regulated entities that hold them. This is where fintechs with worldwide banking and e-money licenses are poised to win. At Airwallex, we hold licenses across 20+ markets. Revolut and Wise have similar global footprints. These companies already move money across borders at near-zero cost. Extending that to agent-initiated payments is a natural step.

The weakness: no chargebacks. If an agent makes a bad purchase via bank transfer, the money is gone. Geography fragmentation is real. Consent flows are still maturing. But for B2B payments, where chargebacks are less critical and fee sensitivity is high, A2A has a strong case.

The shift from cards to A2A is inevitable. Not because cards will disappear. They will not. But because the economics are too compelling to ignore for any transaction where the consumer does not need the insurance that card networks provide.

Stablecoins: the account and rail of the future

The stablecoin route is new. The infrastructure is being built right now.

Circle reported that AI agents completed 140 million payments worth $43 million in nine months. Almost all on USDC. Circle launched Nanopayments on testnet in March 2026, enabling gas-free USDC transfers as small as $0.000001. Coinbase's x402 protocol moved to the Linux Foundation in September 2025 with backing from Google, Stripe, Visa, Cloudflare, AWS, Anthropic, and NEAR.

Stripe's Tempo L1 went live on March 18, 2026, with Visa, Mastercard, Deutsche Bank, Shopify, OpenAI, and Anthropic as design partners. The Machine Payments Protocol (MPP) introduces "sessions." An agent authorizes a spending limit upfront and streams micropayments continuously. Rail-agnostic by design. Visa extended MPP for cards, Lightspark for Bitcoin Lightning, Tempo for stablecoins. Same protocol, any settlement layer.

This is the ideal model. Here is why.

With stablecoins, the wallet can be owned by the individual. Not by a bank. Not by a card network. Not by a fintech. By you. The agent holds a wallet with funds you loaded, spends within limits you set, and the transaction settles on a public ledger you can audit. No intermediary skimming 1 to 3%. No bank holding your float. True ownership of the payment layer.

But three things need to happen before this scales.

First, on-off ramp costs need to go to near zero. Today, converting fiat to stablecoins costs 0.5 to 1.5%. That fee erases the savings. Crypto players like Coinbase and Circle are best positioned here. They have the liquidity for on-off ramp operations and the active consumer base already comfortable with wallets and tokens.

Second, AML and transaction monitoring need to work natively on-chain. Card networks and banks have decades of compliance infrastructure. Blockchain transactions are pseudonymous by default. For stablecoin rails to gain regulatory trust, there must be a system for flagging and pushing out bad actors in real time. Without that, regulators will not let this scale to consumer payments. Companies like Chainalysis and Elliptic do this today for crypto exchanges. The same capability needs to be embedded into the agent payment flow itself.

Third, merchant and consumer adoption needs to reach critical mass. Stablecoins are winning machine-to-machine payments today. Agent-to-agent compute purchases. Pay-per-API-call. Pay-per-inference. These transactions are too small for cards and too fast for bank rails. But consumer adoption is a different challenge. Most people do not hold USDC. Most merchants do not accept it.

Wallet infrastructure is the bridge. Stripe acquired Privy in 2025 to own the agent-wallet primitive. Turnkey raised a $30M Series B from Bain Capital Crypto and Sequoia. Whoever provisions the agent's wallet controls the spend flow.

Who wins where

These three rails are not competing to replace each other. They are sorting into different layers.

Cards keep the human-authorized consumer purchase layer. You want chargebacks. You want dispute resolution. You want the insurance. Stripe and the card networks win here.

Bank accounts win the cost-sensitive layer. B2B payments, cross-border transactions, recurring commerce where 1 to 3% interchange is an unacceptable cost. Fintechs with global banking licenses win here.

Stablecoins win the machine-to-machine layer. Sub-cent micropayments, agent-to-agent compute, pay-per-use APIs. Everyone is getting into this space and it remains to be seen.

The protocol war (ACP, AP2, UCP, TAP, x402, MPP, and counting) will consolidate. Stripe already owns card rails, stablecoin settlement via Tempo, and agent wallets via Privy. Visa owns Tink for A2A, TAP for cards, and is a Tempo design partner. Everyone is hedging everything.

The real question is not which rail wins. It is whether the stablecoin rail matures fast enough to capture the machine economy before the card networks and banks figure out how to extend their existing rails into that space. That race is the one worth watching.