A world without frontier AI: the moral arithmetic of the closed lab

There was a moment, somewhere between the 12th and the 16th of June, when a question that had always been theoretical became briefly literal: what do you do if the frontier just goes away?

Anthropic's two best models had been switched off by a government order. The most-used open models in the world that week came, awkwardly, from China. The American lab with the friendliest relationship to the current administration, OpenAI, was untouched, but that only sharpened the point rather than softening it. The frontier had revealed itself to be not a fact of nature but a permission, granted by states, to a handful of private companies, extendable to you at their discretion and revocable at theirs.

We want to use this essay to think about two things that the news coverage mostly skipped past in its rush to handicap the winners. First, the practical and oddly clarifying exercise of imagining a world without reliable frontier AI, and what it asks of anyone who builds. And second, the harder moral question underneath: if a private lab is going to hold the off switch for a capability the rest of us increasingly depend on, what does it owe us, and is the closed model, as a moral structure, even defensible?

The thought experiment

Start by taking the loss seriously rather than reassuring yourself it won't happen.

Imagine that, from tomorrow, you cannot rely on any frontier model being there next quarter. Not because the technology regressed, but because access has become a political variable: granted to approved national-champion firms, withdrawn from foreign nationals, rationed by export letters and trade leverage. The capability still exists. It is simply no longer yours in any sense you can plan around.

The interesting thing about this exercise is that it is not actually a return to 2019. The knowledge does not vanish. What vanishes is the guarantee. And it turns out that a great deal of how we have all been building over the last three years quietly assumes the guarantee. We have written products whose core loop is a call to a model we do not own, served by a company in a country that is not ours, under terms that can change without our consent. We have, in the language we use with our own founders, built our houses on rented land and then renovated as if we held the deed.

A world without dependable frontier AI is one where that assumption is no longer free. And the striking thing, the thing to hold onto when the panic fades, is that almost everything you would do to prepare for it is something you should be doing anyway.

You would build so that the model is a replaceable component, not the foundation. You would keep your data, your evaluations, your domain knowledge and your distribution as the assets that actually compound, because those are the things no government can repossess. You would treat the specific model as the most swappable part of the system rather than the most sacred. You would keep a capable open-weight model running on hardware you control, not because it is as good as the frontier (it usually is not) but because a floor you own is worth more than a ceiling you rent. And you would get honest about how much of your product's value comes from the raw intelligence of the model versus the system you built around it. The companies that survive a frontier blackout are the ones whose moat was never the model in the first place.

This is, almost word for word, the architecture we already preach: keep the compounding assets, treat the swappable parts as swappable, build the system rather than worshipping the component. The Fable shutdown did not change the right way to build. It just removed the excuse for building any other way.

The harder question: who should hold the off switch

Now the philosophical part, because it does not resolve cleanly and nobody should pretend it does.

The closed-lab model (keep the weights private, serve the capability through an API, retain the ability to monitor, restrict and shut down) has always been justified on safety grounds, and the justification is not stupid. If a model can plausibly help discover software vulnerabilities at scale, or worse, then the ability to revoke it is a safety feature, not a fiction. Mythos reportedly walking through classified systems in an afternoon is exactly the scenario the closed model was built to contain. You cannot recall an open weight. You can switch off an API. Seen from inside a safety-focused lab, the off switch is not a bug; it is the entire point.

But June exposed the cost of that structure, and the cost is a moral one, not just a commercial inconvenience.

When a private lab builds a capability that thousands of companies, hospitals, public services and livelihoods come to depend on, and then retains a unilateral off switch, it has created a profound asymmetry. You take the dependency; they keep the control. And the moment the switch is flipped, even for defensible reasons, even by a government rather than the lab itself, every person downstream discovers that their dependence was never matched by any reciprocal obligation. There was no notice. There was no continuity guarantee. There was a ninety-minute window and then silence.

This is the moral arithmetic we think closed labs have not honestly reckoned with. It is one thing to hold the off switch for a weapon. It is another to hold it for what has quietly become infrastructure. We do not accept, from a power company or a water utility, "we can cut you off without notice and owe you nothing." We accept it, apparently, from the companies supplying the most important productive capability of the decade, because we have all silently agreed to call it a product rather than a utility.

The implications that land on OpenAI

We single OpenAI out here not to pile on, but because OpenAI sits at the sharpest point of this contradiction, and June drew the contradiction in bright lines.

The organisation began with a name and a charter that promised openness and a mission framed explicitly around the benefit of all of humanity. It became the definitive closed lab, with the weights held tight and the governance famously turbulent. That journey from open to closed was always defended as the responsible path: the technology became too powerful to give away. Reasonable people can hold that view.

But watch what the Fable episode does to it. When the US government switched off Anthropic's models, the lab that emerged untouched and advantaged was the one with the better political relationship, not the one with the better safety story. Anthropic, by most accounts, lost precisely because it had been the loudest about danger. The closed model, sold to the public as a safety architecture, revealed itself in practice as a political one. Whoever holds the off switch is now playing a game whose rules are written by whichever government they are closest to. "Safe" and "favoured" became indistinguishable from the outside.

That is the moral implication we think OpenAI in particular has to answer for, because it is the lab that most loudly claimed the high ground of acting for everyone. If your weights are closed, your governance is private, and your continuity depends on the goodwill of a single administration, then the people who depend on you are not your beneficiaries. They are hostages to a relationship they cannot see and did not consent to. "For the benefit of all humanity" is not compatible with "available to you until our government decides otherwise." You cannot claim a universal mission and run a revocable, nationally-captured service at the same time. June made those two things visibly incompatible.

We are not arguing that everything should be open, or that safety concerns are fake. The NSA red-team result alone should embarrass anyone who thinks capability has no dark side. We are arguing something narrower and, we think, harder to dodge: a lab that takes on the closed model takes on an obligation it has mostly refused to name. If you keep the off switch, you owe the people downstream of it honesty about who can flip it, continuity commitments that mean something, and an end to the pretence that a privately-held, politically-contingent kill switch is the same thing as serving humanity.

How we are choosing to build

An operator has to land somewhere between the safety case for closed models and the moral case against unaccountable dependence.

We land on resilience as the only honest answer available to people who do not get to write AI policy. We will use the best model for the job, including the closed frontier ones, because pretending otherwise would be performative. But we will build every company so that no single model, and no single government's view of that model, can end it. We will keep an open-weight floor under everything, treat our data and our systems as the assets that last, and design for the switch being flipped rather than praying it never is.

And we will keep saying the quiet part, because the labs will not. A capability that the world depends on should not be controllable, without notice or recourse, by a private company and the one government it has befriended. Until that changes, "a world without frontier AI" is not a doomsday scenario to plan against. It is a Tuesday afternoon in June that we have already lived through once, and the only people who came out fine were the ones who had never confused the model with the company.

Louis O'Connell-Bristow and James Freestone are the co-founders of Moonlabs, the operator-led AI incubator and academy. They write about building durable AI-first companies in a world where the tools are powerful, essential, and not entirely under anyone's control.