Morning All,

A little under two weeks ago, the US government led by Agent Orange, reached across the oceans and hit the kill switch on the world's access to the best AI model.

One government, one leader, one click of the fingers, and the lights went out for the rest of us.

Commerce officials ordered Anthropic, one of the most valuable AI companies on earth, to cut off access to two of its most capable models for every foreign national, whether they were sitting in Paris, Lagos, or at a desk inside the US itself. The company couldn't reliably tell who was who, so it pulled the plug on both models altogether. The exact trigger is still in dispute, and I won't pretend to know the exact motivation behind the decision, but the headline consequence isn't in question.

If you run anything important on somebody else's AI, that should be a huge wake up call. It just proved, in front of the whole world, that as we rush to add AI into every crevice of our working and personal lives, we might not be building on the most solid of foundations. That's what we're discussing today, and why "sovereignty" might be the wrong thing to be chasing in the first place.

Only two countries own the building

Let's define the word everyone keeps throwing around. Real AI sovereignty "is the ability of a country or region to develop, control and govern AI without depending on foreign powers". Which means owning the whole stack, the chips, the compute, the models, the data, the energy to run it all, and the rules on top. Not renting any of it. Owning it.

By that measure, the Sovereign AI club has exactly two members. The United States and China. Everyone else is a guest.

Look at the gap and it's hard to argue otherwise. OpenAI is worth close to a trillion dollars, the big American tech firms are spending something like $450 billion a year on AI infrastructure, and a single site in Memphis runs more than half a million GPUs. Europe's answer has mostly been to write rules. Decent rules, sure, but you cannot fully enforce rules on an industry you don't actually have or control.

So here's the uncomfortable bit, and I'll say it plainly. For every other single country on the planet, true AI sovereignty is a fantasy. The required cheque is too big, the head start too long, and that gap grows wider every month. Chasing it is like a Sunday league side announcing it'll win the Champions League next season. Love the ambition, doesn't change the fact it's a delusional goal that's not going to happen.

The honest goal for everyone outside the big two isn't sovereignty. It's some degree of self-determination. Europe's relationship to AI right now is having their flag flying on a rented building. That isn't ownership. Having your own front door key, and a generator in the basement, changes everything. More on that shortly.

A tenant's two problems

Renting your intelligence feels brilliant right up until it doesn't. It's cheap, it's instant, and somebody else does the maintenance. But underneath all that convenience, you have two problems.

The first is the kill switch. We just watched one version of it, but it wears many costumes. A US export directive is one. American law applies to data held by US companies wherever in the world it sits. It gets more physical too, because roughly 99% of the world's internet travels through undersea cables, and the UK leans on around 60 of them. Cut the cord, for whatever reason, and the cloud stops being a cloud.

The second problem is the rent, it's not cheap and it's going up. The prices we pay for AI right now aren't real. They're subsidised by investors betting on the future. Sam Altman admitted as much when he said OpenAI was "currently losing money on openai pro subscriptions" because people used them far more than expected, and that was the $200-a-month tier. Running these models reportedly cost around $700,000 a day at one point. That bill doesn't vanish into thin air. One day it lands on your desk.

Both problems share a single root. You don't own the thing your business runs on. And a tool you don't control isn't really a tool, it's a favour, and favours can be withdrawn.

The gap closed while we argued about the frontier

Here's the part that flips the script, and it's genuinely good news.

While everyone was transfixed by the race at the very top, the open models quietly caught up with the pack. You can now download a capable model, for free, run it on your own machine, and get real work done. Not toy work either. IBM who are hardly a band of hippies, reckon the gap between open models and the frontier giants has narrowed to the point where the difference is "not practically meaningful" for everyday tasks. Better still, a smaller model fine-tuned on your own data will often beat a giant generalist at your specific job.

Small, specialised models like Google's Gemma, IBM's Granite, and Alibaba's Qwen can summarise documents, answer customer questions, and power little agents while running on a laptop, sometimes even a phone, with the internet switched off completely.

The hardware caught up too. A pair of high-end consumer graphics cards can now roughly match a data-centre chip that used to cost a fortune, and tools like Ollama make getting a model running locally closer to installing an app than running a science experiment. The economics flip completely. Instead of a meter that ticks with every word, you pay once for the kit and the running cost drops close to zero. For comparison, if five years of heavy cloud use comes in at around £120,000, the local AI setup would set you back £18,000 to do the same work yourself.

This is already happening in the wild. A London law firm pointed a local model at confidential contracts, cut review time by about 60%, and kept every sensitive document inside its own walls. A health startup ran a small model fine-tuned on anonymised NHS data, hit its compliance bar comfortably, and got faster results because nothing had to make the round trip to a foreign server. The dependency, in other words, has quietly become a choice.

"But aren't the best open models Chinese?"

Yeah, this is the fair question. A lot of the strongest open-weight models do come out of China, names like Qwen, DeepSeek, and GLM. So doesn't going local just swap an American landlord for a Chinese one?

No, and the reason matters more than almost anything else here. An API you call over the internet is a tap that someone else controls; they can turn it off, meter it, and watch what flows through it. A set of open weights you've downloaded is more like a book on your shelf. Once it's sitting on hardware you own, you can unplug the network completely and it still works, exactly the same, for as long as you like.

Nobody can reach across an ocean and delete a book you already own. So the model's passport stops mattering the moment it's running on your own machine. You aren't trusting the people who made it, because you've quietly removed their ability to do anything to you at all. That, in a single sentence, is the whole difference between renting and owning.

Aim for self-determination, not sovereignty

So let's bring it back to you, whether "you" is a country, a company, or an individual reading this with their morning coffee.

Think of it like a backup generator. You're not trying to out-produce the national grid, that's a fool's errand, and you don't need to. You just need enough power of your own that a blackout doesn't finish you off. The aim isn't winning the AI race; it's making sure no one else can end yours.

In practice that means going hybrid, which is the grown-up answer anyway. Run the bulk of your work, especially anything touching sensitive data, on a local model you control. Then reach for an expensive frontier model only for the rare, genuinely hard problem where you need its extra muscle. You get the best of both, and you're never fully exposed.

The wind is at your back here in the UK, too. The government has put £500 million into a Sovereign AI Unit, built the Isambard-AI supercomputer in Bristol, and passed a Data Act that nudges firms toward keeping sensitive information at home. The tools, the hardware, and the political will are finally pointing the same way.

Let's be clear, none of this is effort-free. Someone has to maintain the kit, the skills take time to build, and the very top-end reasoning still belongs to the frontier for now. Self-determination is a balance you set, not a switch you flip once. The trick is to map your assets and your risks honestly, then decide how much you're comfortable renting and how much you want to own. The point is that, for the first time, that balance is yours to set, rather than one handed down in a boardroom on another continent.

If you remember nothing else

You'll probably never be fully AI-sovereign, and the liberating part is that you don't need to be. Work out the level of self-determination your situation actually calls for, match it to how much risk you can stomach, and then go and look properly at what a local setup would mean for you. This week, list the AI tasks that touch your most sensitive data, or that would hurt most if the price jumped or the access vanished, then try running just one of them on a free local model and feel how close "good enough" has quietly become. Own a little of your intelligence, before you're reminded that you rent all of it.

See you next week.