A week ago the agentic checkout was a slide deck. As of last night it is a working stack with named partners, public protocol commitments, and a CEO statement of intent. The pace tells the story. Five business days produced four protocols, two BNPL activations, one supercomputer deal, one cloud distribution pact, and one declared competitive strategy. None of these are speculative. All of them are live or contractually committed.

The question is not whether agentic commerce arrives. It is whether the layer the card networks occupy continues to expand alongside it, or whether the volume routes around them on both sides.

The five days that closed the bracket

The events run in tight sequence.

May 7. The Solana Foundation and Google Cloud opened Pay.sh on the agent stablecoin rail, letting agents pay for enterprise APIs via the x402 protocol. Machine-to-machine. Stablecoin-final.

May 11. AWS, Stripe, and Coinbase shipped the Merchant Payments Protocol for agent-to-agent and agent-to-merchant commerce, again on stablecoin rails with no chargeback model. The same day, Anthropic took all of SpaceX's Colossus 1 supercomputer, bringing the contracted Anthropic compute stack to more than 11 gigawatts across hyperscalers.

May 12. Anthropic launched Claude Platform on AWS, putting the lab inside the AWS Marketplace procurement layer that most large enterprises already buy through. Google and Klarna announced that Klarna's buy-now-pay-later button would appear across Gemini, Google Search AI Mode, Google Store, Google Play, and Google Cloud, activated through Google's Universal Commerce Protocol.

May 13. Less than 18 hours later, Affirm joined Klarna inside Gemini. Same UCP. Same surfaces. Same instant.

Then Affirm's investor forum, the same day. CEO Max Levchin told investors the company's strategic goal is to be "as important to merchants as Visa, Mastercard, American Express." The supporting numbers are not abstract: $150 billion in cumulative volume, 2.3 billion repayments tracked over 14 years, 4.4 million Affirm Card users spending roughly $2,400 each annually, and a 30 percent checkout conversion lift on the merchant side. Affirm is also pursuing an industrial bank charter to hold a larger share of its loan book.

Eight items. Five business days. Two ends of the same stack. The middle was not where any of these announcements landed.

What "bracketed" means

The card networks sell two things to merchants: rails and dispute infrastructure. The rails move money from a buyer's funding source to a seller's account, governed by network rules and protected by chargeback rights that the networks have built and refined over decades. Interchange funds that infrastructure. Issuers underwrite the credit. The networks orchestrate the rest.

The events of the past week pressure both ends of that arrangement.

On the machine-to-machine side, Pay.sh and the Merchant Payments Protocol route agent payments around the card rails entirely. An agent calling an API, paying for an MCP server, or settling a content fee does not need card-network arbitration. The transaction finalizes on chain. Chargebacks are not part of the model because they are not part of the use case.

On the consumer side, BNPL providers inside the agent compete for the payment-method pick at the moment of recommendation. When an agent in Gemini suggests a product and the buyer accepts a Klarna or Affirm financing plan, the card network does not see that transaction unless the BNPL provider routes the underlying credit to a card credential. Affirm's closed-loop architecture, in which it operates as issuer, transmitter, acquirer, and risk manager simultaneously, does not require that routing. Klarna's pattern is similar. The agent picked. The BNPL underwrote. The merchant got paid. The card network was not in the flow.

The middle, where interchange and dispute infrastructure live, depends on volume that the agent-mediated transactions are starting to bypass. The bracket is not a forecast. It is what last week looked like in production.

Affirm's declaration

The Affirm investor forum is the part of the week most worth reading in full.

Levchin's framing was direct. The company's ambition is to occupy the role merchants currently assign to the card networks. The competitive moat he described has two components. The first is data and underwriting: 14 years of repayment history, 2.3 billion observations, real-time per-transaction evaluation rather than static credit lines. Levchin called this data "rocket fuel." The second is the closed-loop network itself, where Affirm controls each step of the transaction rather than splitting roles across issuer, network, and acquirer.

The growth metrics back the positioning. Transaction frequency rose 50 percent from the previous investor forum, to 6.7 transactions per active consumer annually. Affirm Card spend grew 130 percent year over year. Wallet volume grew 155 percent year over year. The card holder spends roughly three times what the non-card holder does.

The industrial bank charter is the part most likely to be underestimated. Affirm wants to originate 40 to 50 percent of its loan volume and hold about 10 percent on balance sheet. That changes the math on what Affirm can offer merchants and how the company captures the spread that has historically gone to issuers.

None of this is hostile commentary. Affirm stated the plan in public to its own investors. The question now is how the rest of the stack responds.

The protocol stack now visible

The agentic checkout has three layers, and each layer has named contenders.

The consumer-facing protocol layer. Google's Universal Commerce Protocol is the open-source play, with Klarna and Affirm activated this week. The closed-network protocols are Mastercard's Verifiable Intent and Visa's Trusted Agent Protocol. The technical work on commitment governance and credential verification is real, and both protocols ride the network's existing dispute and fraud infrastructure. The distribution gap is the problem. Google now has two consumer-surface integrations through UCP. The closed-network protocols are still building their integration sets.

The machine-to-machine protocol layer. Pay.sh on Solana and Google Cloud carries x402-based agent payments. MPP on AWS, Stripe, and Coinbase covers agent-to-agent and agent-to-merchant flows. Both settle finally on stablecoin rails. Both bypass the card networks structurally.

The distribution layer. Gemini Intelligence on Android ships on Samsung Galaxy S26 and Pixel 10 this summer, with Chrome, Gboard, Auto, and laptop endpoints to follow. The agent moves to the OS, not the app. OpenAI's ad infrastructure for ChatGPT is moving the same direction from a different starting point. Anthropic is distributing through AWS Marketplace at the enterprise procurement layer. The agent surfaces are multiplying faster than any single payment standard can claim them all.

The pattern across these three layers is the same. The open-source and closed-network protocols are competing for the same buyer relationship. The machine-to-machine protocols are operating on different rails from either. The card networks have credible work in flight at the consumer layer and structural exposure at the machine layer.

What the card networks can do

There are three serious plays available, and the strong case is that they need to happen at the same time.

The first is interoperability. Mastercard and Visa can publish a UCP interop posture that lets their protocols ride the same agent integrations Google is shipping. The card-network advantage in that posture is dispute infrastructure, which is the part UCP and the BNPL providers do not yet replicate at scale. Being the rail underneath someone else's protocol is not a defeat. It is the position Visa and Mastercard have held for decades against every direct-to-consumer payment scheme that came before agents.

The second is acquisition or partnership at the agent surface. The networks already work with merchants, issuers, and processors. The new counterparty type is the AI lab and the consumer-agent platform. Direct integrations with OpenAI, Anthropic, Google, and the next tier of agent-runtime companies would put network-grade dispute coverage inside the conversational checkout where Klarna and Affirm currently sit alone.

The third is sharpening the dispute infrastructure as the differentiator. The protocols address commitment governance, which is the prevention side. They do not address repudiation. When the agent recommended the purchase, the consumer approved the BNPL terms, and the goods never arrived, the consumer-recourse path is not yet defined inside UCP. The card networks built that infrastructure over 50 years and sell access to it as a meaningful part of what an issuer pays for. The repositioning is to make agent-aware dispute the product the networks publish a protocol for. G2's Enterprise AI Agents Report found that 95 percent of IT leaders cite integration as the primary hurdle in AI adoption. Trust infrastructure follows integration in the buying decision. The networks are well-positioned for the second, less so for the first.

These three plays are not exclusive. The strong case is that the response has to happen on all three at once because the window is shorter than the protocols are old.

What to watch next

Five concrete things to track over the next 30 days.

Whether PayPal, Sezzle, or Zip announce UCP integrations. The Klarna and Affirm activations are the marquee credentials. The next wave is where the open-source standard either keeps pulling fintechs in or pauses.

Whether Apple Pay or Cash App appear inside Gemini or any other agent surface. Wallet integrations would expand the consumer-protocol footprint beyond BNPL.

Whether Mastercard or Visa publish a UCP interop position inside Q2. The closed-network protocols have technical merit; the question is distribution.

The Affirm industrial bank charter timing. An approval shifts the unit economics of merchant pricing and the spread Affirm can capture.

Whether the Pay.sh and MPP rails see agent-mediated consumer transactions, not just machine-to-machine. The line between those use cases is starting to thin, and the protocols may not stay neatly separated.

The week that just ended is the kind of week that gets quoted in retrospect. Whether the card networks remain at the center of the merchant's checkout, or whether they become a rail underneath someone else's protocol, is decided by what happens in the next few quarters.

Sources

When the agent picks the payment method, the BNPL underwrites, and the stablecoin handles the rail, what part of the merchant's checkout still flows through the card network?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.