Ten trillion tokens a day.
That is what OpenRouter was pushing through its gateway — 400-plus models, 80-plus providers, one chokepoint — when Stripe wrote a $7.5 billion cheque for it. Not a payment. Not a settlement. A river of raw intent, flowing at machine speed, twenty-four hours a day.
I keep a folder of AI-crypto hybrids — fifty of them, tracked across two years, updated at 2am the way other people water houseplants — and I have learned to distrust any number that large arriving without a revenue line stapled to it. So I went hunting for the smaller numbers. The ones that didn't make the announcement.
A dozen Shopify merchants, sales "close to zero." An Instant Checkout product that shipped in spring and was switched off by autumn. A chief executive who, asked directly for volume, chose not to answer.
The signal is silent. In a bull market, silence is the loudest thing you'll hear all quarter.
April 2026: Stripe released 288 agentic commerce products in a single push. Read that as an engineer rather than a headline — 288 SKUs in one release is not a demand curve, it is a land grab for the definition of a standard. The Agentic Commerce Suite was built to cover OpenAI, Microsoft, Meta and Google simultaneously. Link wallets. Shared payment tokens. A "Chain of Trust" demo establishing that an agent is who it claims to be, still running in Stripe test mode.
Around it, the cast: Meta's Muse at 2.18 million daily actives — the first product where a user hands over a purchase entirely, no confirmation screen, no final tap. Shopify's merchants on one side, OpenRouter's model routing on the other. And a protocol stack still arguing with itself: PAP, then Visa's TAP, then UCP, each built on the same choreography — identity first, money second.
Two numbers from the same reporting cycle should sit next to each other until they hurt. Ninety-three percent of merchants believe the agent side should carry liability for a purchase gone wrong. Fourteen percent of consumers say they would trust an agent to buy on their behalf.
That gap — 93 against 14 — is the entire business, and nobody is pricing it.
Start with the thing almost nobody in this discussion is naming: Stripe is not a bank. It is an aggregator and acquirer, running on state money-transmitter licences and foreign EMI permissions, with issuing and treasury pushed through partner banks. That is a solid base for card-not-present commerce. It is a useless base for what comes next.
The regulatory question in agentic commerce is not "do you have a licence." It is "does an AI agent's instruction constitute a valid payment order" — and no money-transmitter framework on earth answers that. When a person taps a button, intent, authentication and authorization collapse into one auditable instant. When an agent executes, those three separate, and the law has a hole exactly the size of the separation.
Which is why "still in test mode" deserves more scrutiny than it is getting. Test mode is a technical status. It is also a compliance posture — a sandbox that generates demos instead of filings, defers the question of who is regulated, and buys eighteen months of ambiguity at the cost of zero live revenue. In 2024 I built a narrative translation guide for a Cape Town fund, mapping crypto sectors onto traditional asset classes for clients who had avoided the space over "narrative risk." The question they asked most often was not about volatility. It was: who is liable if the software is wrong? Two years later, nobody has answered them.
I have watched this move before. In the Layer 2 wars, "decentralized sequencing" was a slide deck for two years while every sequencer on the market was a single node with a nicer logo. The agentic equivalent — "decentralized agent authorization" — is currently the same slide, and the industry has learned nothing except better typography.
Then there is OpenRouter, and this is where my interest sharpens, because it is the part of the deal that isn't about money at all.
Owning both the payment layer and the model-routing layer gives Stripe something no competitor has: a continuous read on the path from prompt to purchase. Not transaction data. Intent data. What a person wanted, in their own words, and whether the machine delivered it. That is the deepest moat in the stack — and the least provable one. You cannot show a regulator a data panel they are not permitted to see and call it a competitive advantage. You can only call it an audit finding.
It also drags a payment company into terrain it never trained for. Four hundred models from dozens of jurisdictions, routed through one gateway, means Stripe inherited AI and cross-border data exposure entirely unrelated to acquiring. And intent data sits closer to thought than to commerce. GDPR, CCPA, PIPL — all three treat inferred preference with a wariness that transaction logs never attract. The moat and the liability are the same asset. That is not a coincidence. That is the trade.
And here is what should worry the bull case most: Stripe controls the pipe, not the valve.
It owns the protocols, the token routing, the developer surface. What it failed to hold is the checkout — the place where money actually moves, where a human or an agent makes the final call. Instant Checkout lived roughly six months and was retired. The dozen Shopify merchants who tried it reported sales near zero. Not low conversion. Zero.
That is the hardest disconfirming evidence in the entire narrative, and it kills the "vertical takeoff" story on its own. If agentic commerce had genuinely lifted off, small merchants would be the first to feel it — shortest feedback loop, least ability to fake it. Zero is not a funnel problem. Zero is a demand problem wearing a funnel's clothes.
Which brings me to crypto, and to why I'm writing this on a blockchain beat instead of a payments one.
Every property that makes agentic commerce hard — high frequency, tiny ticket size, programmable, machine-to-machine, no human in the loop — is a property stablecoin rails already have. Stripe bought Bridge for precisely this reason, and if a GENIUS-style stablecoin framework settles in the US, Stripe becomes the junction where agent payments meet programmable settlement. That is the genuinely interesting position, and it matters more than anything in the CBDC conversation, because CBDCs are a state project and stablecoins are a market one.
But crypto does not escape the problem. Agent payment protocols on-chain inherit the identical authorization vacuum — a signed transaction is not proof of intent, and prompt injection does not care whether the rail is a card network or a rollup. The AML models are what I would worry about most. A manipulated agent produces a transaction that is neither fraudulent nor legitimate — legitimately executed and genuinely unwanted — and every monitoring system on the market today has no category for it. Ninety-three percent of merchants wanting the agent to carry liability is not a demand for fairness. It is a demand to outsource a risk they cannot model.
The consensus reading is that regulation will decide who eats the loss, and that Stripe's fate turns on the answer. I think that is backwards.
The liability question is not unresolved because regulators are slow. It stays unresolved because openness is profitable for the incumbent infrastructure. A test-mode sandbox is a hedge that pays for itself in optionality — and it also freezes the market, because no merchant scales an agent checkout channel while loss allocation is undefined. The vacuum is not a gap in the system. It is a moat dressed as a gap.
Second: everyone is watching Adyen and PayPal. Wrong fight. The real contest is between Visa's TAP, the AI platforms' own protocols, and whoever else gets to define what a valid agent instruction looks like. Stripe's neutral-integrator position is genuinely valuable and genuinely fragile, because Meta already owns the entry point through Muse. Building infrastructure for four giants that each have a motive and a capability to replace you is not a partnership strategy. It is a countdown.
Third, and the one I keep circling: the Chain of Trust work is aimed at the wrong adversary. Verifying that an agent is a legitimate agent solves impersonation. It does nothing about an agent that is authentic, credentialed, and wrong. Trust chains authenticate identity. They cannot authenticate intent. Everyone is building the first thing because it is the thing that demos well.
Watch for the first piece of European guidance on agent-error liability allocation — PSD3 revisions are where this gets named first, and the direction of that sentence will matter more than any product launch this year. Watch, too, whether stablecoin settlement becomes the default rail for sub-cent agent payments before the card networks notice they have been routed around.
The crash is just a chapter, not the end — but the chapter being written right now is about who gets to define an instruction. Alchemy is just storytelling with better chemistry, and the chemistry here is unstable.
If ninety-three percent of merchants already know who should pay when an agent gets it wrong, what exactly is the industry waiting for — a ruling, or a permission slip?