Slowing down is a product decision.

AI stocks fell this morning after the people building the most advanced models asked the industry to slow down. Dario Amodei published a long essay arguing for a slower pace so guardrails can catch up. Sam Altman said OpenAI would not go public this year — “given everything happening with safety, right now would be an ill-advised moment to go public.”

The market read it immediately. Nvidia off more than 3%, Intel, AMD and Marvell between 5% and 6%, the Philadelphia Semiconductor Index down almost 6%, SoftBank closing nearly 11% lower in Tokyo. But Meta, Microsoft and Alphabet rose — investors beginning to separate the companies that need endless capex from the ones that could profit by cutting it.

In the same week, the payments industry ran hard in the opposite direction. Visa announced Agent Scoring, an Agentic Registry and stablecoin settlement. Stripe shipped an agentic commerce suite letting agents settle over a machine payments protocol. PayPal launched a platform for issuing custom stablecoins and moved nine figures through it inside a fortnight. Visa, Mastercard and Coinbase are in open competition over the rails by which an autonomous agent spends money.

So: the labs asked for a slowdown, and the financial infrastructure to let software spend on your behalf shipped anyway.

The debate is happening one layer too high

Read the slowdown argument closely and it is almost entirely about capability. How fast should frontier models improve. How much compute should go into the next training run. Whether the curve should be bent.

That is a real question and we are not equipped to answer it. We do not train frontier models. What we do is build products on top of them, which is the layer where capability meets an actual person — a nine-year-old writing up an experiment, someone logging a symptom, someone moving money they cannot afford to lose.

Almost nobody is discussing the pace of deployment. And deployment pace is the one users experience. A model’s capabilities are latent until somebody wires them to a button. The wiring is a product decision, made by product people, on a schedule nobody is debating on X.

You can hold both of these at once: the frontier may or may not need to slow down, and the application layer almost certainly does.

What slowing down looks like when you actually ship

Restraint at this layer is unglamorous and specific. It is not a manifesto. It is four or five decisions that each cost you something real.

Do not give the model the keys. When we put an LLM over health data, it reached that data through a handful of narrow, purpose-built tools rather than a database connection. Every capability had to be designed and named, which is slower and more annoying than handing it broad access and writing a careful prompt. The industry has since catalogued what happens when you do the convenient thing instead — tool poisoning, rug pulls, confused deputies, and scan results showing the overwhelming majority of exposed tool servers failing basic security review.

Do not let a demo look more real than it is. SEAM is a wallet aimed at people the financial system excludes. It has no licences, no bank partner, and has collected no money — and every screen that could be mistaken for a working service says so on its own face, not in a footer. That is deliberately worse marketing. It is the only honest option when the audience has more to lose than you do.

Do not claim something is live because it deploys. Nothing counts as shipped until it writes a real production row. We currently decline to claim a measurement milestone on a product that is already taking payments, because real usage worth reading does not exist yet. Holding that line costs the satisfying announcement and buys a roadmap where every green item is actually green.

Be willing to stop. We took a consumer health product through App Review, a paid subscription tier and real market testing, then retired it when the usage did not justify running the backend. Market testing only means something if you have agreed in advance that the answer might be no.

None of that required anyone at a frontier lab to slow anything down. All of it was available to us at the speed the models already moved.

The payments race is the live test

Which brings the two halves of this week together.

Agentic payments are arriving fast, and the infrastructure is real — this is no longer a demo. Pre-authorised permissions, tokenised credentials, agent registries, settlement rails that split between cards for consumer purchases and stablecoins for machine-to-machine flows. Serious institutions have built serious plumbing.

The question is not whether an agent can spend money. It plainly can. The question is which products should let one, and how soon, and what the person on the other end loses when it goes wrong. For a wallet serving people whose immigration status is entangled with their finances, our answer for now is: not yet, and not from us. We are building in that space and deliberately keeping agents away from money movement.

That is a restraint position, not a capability claim, and it should be read as exactly that. We have not solved agentic payments. We have decided not to ship them into a population that cannot absorb the failure mode.

The part worth keeping

The slowdown conversation will resolve however it resolves, in rooms we are not in. What it should not do is convince everyone downstream that pace is somebody else’s variable.

The frontier sets the pace of capability. Product builders set the pace of deployment. The second one is closer to users, moves faster than the first, and is being decided this week by people who are not writing essays about it.