Someone just pressed the off switch on the future
The US just pulled the world’s most capable AI model offline, overnight. Here’s why Britain can’t keep betting its future on infrastructure it doesn’t own.
On 12th June 2026, the US government ordered Anthropic to pull its two most capable AI models, Fable 5 and Mythos 5, offline for every foreign national on the planet. Anthropic couldn’t reliably separate foreign users from everyone else in real time, so to comply it switched both models off for all of us. (Anthropic’s statement is here.)
These weren’t minor releases. Fable 5 was the frontier: major jumps in reasoning, and the kind of huge-context work nothing else on the market gets close to. And as far as anyone can tell, this is the first time a government has reached in and switched off a frontier AI model that was already deployed and in people’s hands. The stated reason? “National Security.”
Read that again. AI models are now being treated like arms, sitting in the same export-control bracket as missile guidance systems and fighter jets. Here’s the difference. Nobody is betting the British economy on missile guidance systems. We are betting it on AI: to 100x our productivity, grow the economy, save the NHS, keep us safer on the streets. We’ve pinned the entire future to a technology that another government can switch off with a letter sent at 5pm on a Friday.
And it gets worse, because the “national security” reason is disputed. Anthropic says it’s a misunderstanding, that the capability the government is worried about is already available in other models including OpenAI’s, and that it’s working to turn the models back on. They may well be right. But that’s almost the point. You don’t even need a real threat to lose access. You need one person, in another country, to decide. Their hand is on the switch. Not yours.
Yet so far we’ve been happy to hand over the keys to our economy, our defence and potentially our health, to companies another nation can switch off at a moment’s notice (I wrote about this in “Are we sleepwalking into an economy we don’t own?”). This has to end. We cannot build the whole future on a bet that other nations will keep being nice to us. The honest assumption is that the US will keep doing this as frontier models get more powerful, and that Britain, Europe and most of the world will keep getting locked out of the best of it. We’ll be worse off for it.
It’s the same logic as not bothering with a standing army. Without sovereign AI infrastructure we actually control, including models matured here in the UK, we have no way to defend our own economy against the people who dominate the space. We end up at their mercy because we didn’t prepare. (Some would say that’s the British way these days.)
The solution
Honestly, the solution isn’t complicated, though there’s nuance to it. Britain needs to invest far more heavily in AI infrastructure built to be resilient to outside forces: systems designed to run intelligently, locally and at scale. That’s exactly why we built Locai. And we need to pair that with serious investment in building AI models we own.
But the move isn’t to throw £100bn at building the next Anthropic (though if anyone fancies ponying up the cash, do let me know). The way to win this race is to be smarter about it. Three things:
Swallow the ego. We need to get comfortable building on the latest Chinese and open-source models as a starting point. DeepSeek V4 shipped in April as open weights under an MIT licence, self-hostable, sitting within a few months of the closed frontier. That’s a phenomenal base. It saves years of research and spend, closes the gap fast, and gives us a real backstop capability. I can’t tell you how much it frustrates me to watch companies boast about building models that are “100% British.” Why? Beyond a nice line in a pitch deck, there’s no reason for it. Use what’s already out there and build from there - you better believe thats what Locai is doing.
Bet on new architectures. The next phases of this race will almost certainly move beyond the transformer architecture everything runs on today, and that’s where we can lead rather than chase. We should be putting real money into becoming global leaders in models that are ultra power-efficient and run across a far broader range of hardware, narrow and focused enough to collaborate with each other, and small enough at nano-scale to live directly on sensors.
Control the infrastructure (the pitch — obviously I was going to get one in). This is where Locai comes in. We need national infrastructure for running and managing AI that’s completely independent of outside influence. Think of it almost like building the nation’s own private servers: infrastructure designed to deploy and manage AI models with no need for connectivity, working in complete isolation from anyone else’s control. We don’t just need to airgap our businesses’ AI. We need to airgap the nation’s AI.
The switch exists, and we’ve now watched it get used. I started Locai because I’d rather we built the off switch the people can control themselves rather than having us flinching every time someone else reaches for theirs. But strong words are not enough enough. Britain has to get serious about what actually works, because this problem isn’t arriving in five years. It landed on a Friday afternoon this week.
We’re not going to out-build the people who already dominate this, and chasing our own frontier-scale model is the kind of fantasy that gets us cut off all over again while we’re still raising the round to attempt it. We don’t have the time, and we don’t have the compute capability. The practical answer is sitting right in front of us: take the best open models on the table today, ego be damned, and own the infrastructure they run on, and leverage off-cloud compute solutions, so nobody else’s hand gets to decide whether our lights stay on. One of those is a strategy. The other is a prayer and ego.
The next letter cutting us off further is likely already being written right now.