What happens when your AI gets switched off?

Martin Wheatley

Founder, The AI Leader Lab · CIMA Qualified · Managing Director, Rawson Ellis

For nineteen days in June, businesses all over the world lost access to one of the most capable AI models on the market.

Not an outage, not a billing row, but a mandate from the US government.

Anthropic, the American company behind Claude, got a letter at 5:21 on a Friday evening, New York time, on 12 June. The government had applied export controls to two of its most powerful models, Fable 5 and Mythos 5, barring their use by foreign nationals anywhere in the world. Anthropic had no way of checking anybody's nationality in real time. So it turned both models off. For everyone, everywhere.

Access came back on 1 July.

If you're sitting anywhere outside of the USA, read that middle bit again. The restriction was aimed at foreign nationals. From where that order was written, that's you and it's me.

A fortnight later something similar happened at the other big lab. The White House asked OpenAI, the company behind ChatGPT, to hold back its next model, GPT-5.6, and release it only to a small group of government-approved partners first. Sam Altman, CEO OpenAI, told his staff the arrangement "is not our preferred long term model". The model went out publicly a couple of weeks later. Axios reported it as the first time the US government had asked an American AI company to restrict a launch before it happened.

So here's the question I've been putting to people since. If the AI your business had quietly come to depend on went dark on a Friday evening, and nobody could tell you when it was coming back, what would you actually do on Monday morning?

And the answer isn't always easy.

This isn't an IT question

A lot of people may say that this is purely an IT question, but that's not the case.

If you're in financial services you already have a name for this. Operational resilience. Since March 2025 firms in scope have had to be able to run their important business services within what the regulator calls an impact tolerance, defined as "the maximum tolerable disruption to your important business services without causing harm to consumers, firms or markets". And the accountability doesn't transfer. If a supplier is the thing that failed, it's still your problem, formally.

Nineteen days is a long time inside an impact tolerance.

Here's what I find interesting. In May the Bank of England, the FCA and HM Treasury published a joint statement on frontier AI. It's a serious document and it's worth a read. It covers AI as a source of cyber threat, board understanding, vulnerability management, supply chain security. What it doesn't cover is what happens if the AI model your business runs on simply becomes unavailable.

And as we have seen; a month later, one did.

So the regulators have been thinking hard about AI as a threat, and rather less about AI as a dependency. That gap will close, and there's a marker in the diary already. New incident reporting and third party notification requirements come into force on 18 March 2027. The twelve-month preparation window is already running.

If you're not in a regulated sector, none of that applies to you formally, but the logic should. Something in your business is becoming extremely load-bearing, and nobody has written down what happens if it stops.

Which brings us to a phrase you'll keep hearing: Open Source AI.

It's usually raised as a cost question. After June, I think the more interesting version is a control question, so it's worth understanding properly.

There are two kinds of AI model.

The ones you rent (Closed Weight Models). ChatGPT, Claude, Copilot, Gemini. You pay a subscription or you pay for what you use. You can't look inside them, you can't put them on your own computers, and when something breaks it's somebody else's job to fix it. This is where almost every business currently sits. It's also, as June showed, where somebody else's government gets a say.

The ones you can download (Open Weight Models). These get called open source, although open weights is the more accurate term. A lab publishes the finished model as a file, free to take. You run it on your own machines, or in whichever cloud you already use. Nobody charges you for the model itself. You pay to run it, and that distinction matters more than it sounds.

The difference is closer to renting an office versus buying a building than it is to paid versus free. One has a monthly invoice and a landlord. The other has a maintenance bill and a phone that rings when the boiler goes.

One more thing about the downloadable ones. You get the finished model, not the ingredients. You can use it and adapt it, but you can't see what went into it or how it was built. Cake, not recipe.

And the point that matters after June. A model sitting on your own machines can't be switched off by anybody. Not by a supplier, not by a government, not by a change in someone else's foreign policy. That's a genuinely strong argument, and it's a better one than just a focus on pricing.

So why isn't everyone doing it

Plenty of people are, as it happens. Just not with the big ones.

Running a smaller AI model on your own hardware has got dramatically easier over the past year. Two minutes of setup on a decent laptop. A MacBook with 16GB of memory will happily run a model that would have needed a paid cloud subscription eighteen months ago, and for summarising documents, answering questions against your own files or drafting routine copy, it's good enough. Free to run, private by default, and nobody can switch it off.

That end of it gets easier by the month.

The frontier end went the other way.

A Chinese lab called Moonshot, in the same corner of the market as DeepSeek, released the most capable downloadable model anyone has produced (Kimi K3). The file is free, as promised but running it is not.

This is the bit that trips people up. A downloadable model costs nothing to obtain and quite a lot to operate, and the operating cost scales with the size of the model. Moonshot's new one is enormous. Whoever runs it, whether that's you or a cloud provider, has to buy a serious amount of computing power, and that cost reaches you one way or another.

Moonshot also rents access to it, the same way Anthropic rents out Claude, and that rental price is the clearest signal of what the thing costs to run. It's the most expensive model Moonshot has ever charged for. Roughly fifty times the price of the cheaper downloadable options, and only about a third below what Anthropic charges for its very best model.

Free to download. Not cheap to use. The gap between those two is where a lot of business cases quietly fall over.

If you wanted to skip the rental and run it in your own building instead, you'd need something in the region of sixty-four high-end graphics processors wired together. A data centre, not a cupboard with a server in it.

Which leaves an awkward middle. The downloadable model you can comfortably run on your own kit won't replace Claude for your hardest work. The one that might, you can't run at all, so you'd end up renting it from Moonshot, which sends your data to Beijing. Having just spent a section worrying about Washington, that's an uncomfortable place to land.

For what it's worth, on whether open weight has caught up, it depends what you measure. On coding work the best downloadable models are within touching distance. Across the board the researchers at Epoch AI put the lag between frontier to open weight at roughly four months, stretching to six on a stricter test.

What's actually happening instead

The move I'd pay attention to came from Microsoft.

They announced that a downloadable model from Mistral, a French AI company and the closest thing Europe has to its own OpenAI, now sits inside Azure. Aimed specifically at regulated industries, with the option to run it completely cut off from the internet. You get a cheaper model and control over your data without owning hardware or hiring anyone to look after it.

That's the route I'd expect most mid-sized businesses to take. Not downloading anything. Just reaching a different model through a supplier you already have a contract with.

There's a British version too. HSBC, Lloyds and NatWest have signed up to help design a UK sovereign AI model, built to run without American cloud providers at all. Worth noticing what they didn't do. They didn't download something and stand it up themselves. They partnered with a specialist.

Three banks with more money and more engineers than almost anyone reading this, and they still chose to buy rather than build.

What a fallback actually looks like

Back to the Monday morning question if the model you rely on has disappeared.

A fallback isn't a second AI subscription sitting in a drawer. It starts a long way before that, with three things, and none of them are technical.

One. What's actually running in the business. Not what you bought, but what's in use. Which tools, in which teams, doing what, set up by whom. In businesses, this list easily becomes longer than the leadership team expects and nobody owns it centrally. Somebody in finance has built something clever in ChatGPT that three people now depend on, and it isn't written down anywhere.

Two. What happens if it stops. Take each thing on that list and ask what actually breaks. Does a process slow down, or does it stop? Does a customer notice? How long before it hurts? Most of what you find will be inconvenient rather than serious, which is useful to know, because the two or three things that aren't are where all your attention belongs.

Three. What moving would actually take. For those two or three, could you switch to something else, and how long would it take? Not in theory. Who would do it, what would they have to rebuild, and would the output be good enough. There's research suggesting nearly nine in ten executives believe they could switch AI suppliers inside a month, and fewer than half of those who tried found it went smoothly. That gap is the whole risk.

Do those three and you'll have something most businesses don't, which is a written, honest picture of where AI has become load-carrying without anyone deciding it should.

I'd be lying if I said this was solved

It isn't, and I'd be wary of anyone who tells you otherwise.

Running open weight models inside a business is genuinely early. The tooling is improving fast, the costs move monthly, and the good practice is being worked out in public by people who are, frankly, guessing intelligently. What I've written above is where my thinking has got to in July, and I'd expect parts of it to look naive by Christmas.

But the three questions don't depend on any of that being settled. They'd have been worth asking before June and they'll still be worth asking when this generation of models is a footnote. The businesses that came out of June calmly weren't the ones with the cleverest AI strategy. They were the ones who knew what they had.

One small ask

Hit reply and tell me the one AI-shaped thing in your business you'd least want to lose on a Friday afternoon.

If the honest answer is "no idea, and that's rather the problem", say that. It's the most common answer I get, and it's a better starting point than a confident one that turns out to be wrong.

I read every reply. I'm also turning those three steps into a single page you could take into a leadership meeting without needing to explain any of the technology, so tell me if you'd like a copy and I'll send it over when it's done.

♻️ Enjoy this? Repost to your network if you found this useful.

🔔 Follow Martin Wheatley for more AI for leadership.

✅ 𝙃𝙚𝙡𝙥𝙞𝙣𝙜 𝙡𝙚𝙖𝙙𝙚𝙧𝙨 𝙗𝙚𝙘𝙤𝙢𝙚 𝘼𝙄-𝙘𝙤𝙣𝙛𝙞𝙙𝙚𝙣𝙩. Putting AI to work with your people, without replacing what makes them great.

⭐️ If that's the conversation you're interested in having, leave a comment, drop a DM, book a call, or visit my website. All links in my profile.