The walls of China are going up: three scenarios for UK AI sovereignty

On 7 July, reporting emerged that Beijing's Ministry of Commerce had spent the past month in talks with Alibaba, ByteDance and Z.ai about curbing overseas access to China's most capable AI models. The proposal is a tiered regime: basic open-source tools cleared with a filing, more advanced systems facing security review, and the most powerful frontier models barred from public release.

Read it slowly, because it is the end of an era written in the language of paperwork. Since DeepSeek's R1 last year, the cheap, capable, open Chinese models became the floor under an enormous amount of the world's AI, including much of what British startups build on when they cannot afford or cannot reach the American frontier. Those models were the fallback, the thing you could always download and run yourself. Beijing is now moving to close that door.

Washington has already fenced its side. I wrote here recently about Pax Silica and the American bills that would treat even remote cloud access to a controlled GPU as an export, and about the June morning when two frontier models were switched off worldwide in ninety minutes. Both walls are going up at once, from opposite directions, around the same thing.

Most of the commentary is filing this under trade. That is the mistake. This is not a trade story, it is an arms-control story, and until Britain starts reading it that way we will keep making small, sensible, doomed decisions about a very large thing. So let me make the case properly, because UK AI sovereignty is about to stop being a conference panel and start being the whole game.

The weapon nobody will call a weapon

The superpowers are not racing to ship a better chatbot. They are racing to build the defining strategic capability of the century, and they know it even when they dress it up as a product launch. Frontier AI, and the artificial general intelligence the leading labs are openly aiming at, is not the internet or the smartphone. It is closer to the atom.

Hold the comparison for a second, because it is more exact than it is comfortable. A nation without its own nuclear capability during the Cold War could sit at the table. It could speak, trade, ally, object. What it could not do was set the terms, because the terms were written by the powers who could end the conversation. The five permanent members of the Security Council are, to this day, the five recognised nuclear states. Capability wrote the rulebook, and everyone else negotiated inside rules they had no hand in.

Frontier AI is becoming the same kind of instrument. It compounds, it is dual-use to the point of absurdity, and the lead it confers may not be catchable once it is large enough, because the system that is ahead helps build the next system that is further ahead. A country that owns a frontier lab will write the rules of the coming order. A country that rents one will live inside rules written by others, and will call the arrangement a partnership because the alternative is admitting what it actually is. That is the stake behind the phrase UK AI sovereignty. Not data protection, not chatbot safety. Whether Britain is a rule-writer or a rule-taker in the order that AGI is about to impose.

The metaphor is imperfect, and I know it. Nobody vaporises a city with a language model. But strategically, in the way it concentrates power, rewrites who sets terms, and punishes the nations that arrived late, AI is behaving far more like the bomb than like any consumer technology we have lived through. We should start treating it with the seriousness we once reserved for the thing it most resembles.

What Russia would teach us about seriousness

Here is the part I find most revealing about our own psychology.

Right now the race is a two-hand game, America and China, and the West is treating it as commerce because both leaders are, in the end, powers we trade with and partly trust. The complacency is not really about the technology. It is about who currently holds it. We are relaxed because the frontrunners are a rival we do business with and an ally we depend on.

History is blunt about what changes that. We did not build the atom bomb because it was scientifically interesting. We built it, in a terrified sprint, because we believed a regime we feared would build it first. The nuclear age became real to the public not when fission was discovered but when an enemy could return the favour. Fear of the adversary, not the wonder of the science, is what turns a technology into a national priority.

So ask the uncomfortable question. What happens the day Russia enters the frontier AI race in earnest, alone or as a Moscow-Beijing axis pooling compute and weights beyond the reach of any Western control? My honest bet is that this is the moment Britain suddenly discovers that AI sovereignty was a first-order national-security matter all along. The urgency that is missing today is missing precisely because the two leaders are not, yet, an enemy we lie awake over. Add one that is, and the entire Western posture flips inside a week. The indictment of our current thinking is that we are waiting for the threat to become personal before we act on a race that is already running at full speed. Serious countries do not wait for the enemy to arm before they arm. We are, at the moment, waiting.

The gunpowder script

If you want to know how this ends for the nation that invents a weapon and fails to keep control of it, we have the case study. We just have to be willing to be the China of it.

In the ninth century, Tang dynasty alchemists looking for an elixir of immortality found gunpowder instead. China did not merely make fireworks with it, whatever the schoolbook says. It built fire lances, bombs and rockets, and used them in war for centuries. But it was Europe that took the powder, married it to metallurgy and drill and industrial iron, and turned it into the cannon and the massed musket. And in the nineteenth century, in the Opium Wars, those European guns were turned on China itself and used to force the country that invented gunpowder to open its own doors at gunpoint. The inventor became the victim of the invention. The script did not just flip. It was flipped, deliberately, by the powers that industrialised what China had discovered.

Now stand in front of the mirror. Britain invented the modern AI lab. DeepMind was founded in London in 2010, and its lineage runs directly into Gemini and Deep Think, arguably the most powerful reasoning engine alive on the planet today. We discovered the gunpowder of this century. And then we handed the foundry to California. The powder is ours by origin; the cannon, the pricing, the terms of use and the off-switch belong to somebody else. If the script flips the way it flipped on the Chinese, and history is not shy about repeating this exact manoeuvre, the weapon Britain invented ends up owned abroad and, in the wrong scenario, aimed squarely back at Britain's own sovereignty. Not through malice. Through ownership. A capability you do not control is a capability that can be priced, withheld or turned against you, and we invented this one and gave it away with a press release and a sense of national pride.

That is the frame. Now the roads.

Scenario one: the squeeze

Britain does what Britain does with a slow strategic threat. It commissions a review, holds a summit, publishes a blueprint with "pro-innovation" in the title, and waits.

Both walls finish rising. Beijing restricts its frontier models to domestic use and puts the mid-tier ones behind security reviews no foreign startup will clear. Washington prices allied access to the American frontier as a favour rather than a market good. The cheap open floor a British seed-stage company built on is gone; the expensive American ceiling is rationed and conditional. Compute costs climb, because the competitive discipline the cheap open models imposed on the American labs has evaporated. Startups that priced around near-free inference find their margins were borrowed from a geopolitical accident that has now been corrected. The most ambitious founders relocate to where the models and the compute actually live, which is not here. We keep the accelerators and the panels and the papers, and become a nation that consumes AI rather than one that builds it. Nobody decides this. It happens one deferred decision at a time, which is how sovereignty is always lost.

This is the default. It costs nothing and requires nobody to be brave, which is exactly why it is the likeliest.

Scenario two: the vassal's bargain

The second road is more comfortable and, for that reason, more dangerous, because it feels like winning.

Britain leans all the way into the American bloc. We are already a Pax Silica signatory; here we make ourselves the most loyal and useful member of the club, and in return we are granted preferential access to the American frontier. British companies build on the best closed systems in the world on generous terms, and for a few years it feels like a triumph. Look, we say, our startups have access most nations would envy.

The price is paid in sovereignty, quietly. Preferential access is not control, and a favour granted is a favour that can be withdrawn. We would be building the digital economy of the country on infrastructure owned in another nation, governed by that nation's export law, switchable by that nation's politics, the same politics that pulled two models offline in ninety minutes last month. We would become a well-treated dependency and abandon any serious ambition to build our own, because renting the best was easier than building the merely good. It is the most seductive of the three roads. It is also the one where we wake up one morning to discover that everything we depend on answers to a government we do not elect.

Scenario three: the sprint

The third road is the optimistic one, and it is not a fantasy. It is open to us, and this week's news is exactly the shock that could push us onto it.

In the sprint, Britain reads the double closure correctly and sees the opening inside it. If China fences its open models and America fences its closed ones, a gap opens in the middle of the world for a trusted, capable, openly available frontier model built by a democracy. Nobody credible is filling it. Britain could.

We hold more of the pieces than most people realise. Cosine has built a British frontier model and is training Lumen Sovereign on home soil. The Sovereign AI Fund exists and has already put serious compute behind British companies. The talent is here, much of it currently working for American labs from a King's Cross office. What we have lacked is urgency and the decision to treat this as the priority rather than one programme among many. The sprint means scaling sovereign compute at the pace of a country that knows it is running out of time, backing a British open-weight frontier model and releasing it deliberately to allies and Commonwealth partners as an act of soft power, and building sovereignty clauses into the IP we fund so the next DeepMind cannot be quietly bought and shipped abroad. In this future Britain is neither tenant nor vassal. It is the third pole: the open, democratic anchor between a closed China and a rationed America, exporting trust as a product at the exact moment the other two are exporting restriction. That is a far bigger prize than preferential access to someone else's model, and it is the only road on which UK AI sovereignty is something we hold rather than something we rent.

The wildcard

None of those three roads is disruptive enough to contain the move I actually want to see, which is why it deserves its own name. Call it the wildcard.

The wildcard is that Britain stops being polite about the foreign ownership of its own crown jewel and treats it as the national-security matter it plainly is. DeepMind, and the Deep Think capability descended from it, is the finest AI lineage this country ever produced, and it is owned outright by an American company that decides who may use it, on what terms, and when. A serious nation would at least be asking the unaskable question: what would it take to bring that kind of capability home, to write sovereignty conditions into British AI IP before it is sold, to treat "founded in London, owned in California, switchable from there" as the strategic vulnerability it obviously is rather than a proud line in a pitch deck. Reclaiming the foundry is the one move that flips the gunpowder script back in our favour. It is not in the tidy scenario set because it demands a kind of nerve British technology policy has not shown in a generation. But wildcards are precisely the plays that change the game, and this is the one sitting in plain sight, waiting for someone with the nerve to name it.

We are choosing now

The nuclear age needed a single mushroom cloud to become real to everyone at once. The AGI age is being more generous with its warnings. Fable's ninety-minute shutdown, Pax Silica, and now China bolting its own door are the warning shots, fired in sequence, asking us plainly whether we intend to build our own capability or spend the next two years as a customer of everyone else's.

This week China began closing its door. America closed its own weeks ago. The open era is ending, and no amount of British politeness will hold it open. The only question left is whether we treat the most strategically important technology since the atom like the arms race it has already become, or wait for an adversary to make the point for us. I know which country I want to be a founder in. It is the one that built its own, not the one that watched both doors close and decided the dignified thing to do was wait.

Louis O'Connell-Bristow is a co-founder of Moonlabs, the operator-led AI incubator and academy, and previously built the home.co.uk, Homemove and homedata.co.uk stack. Moonlabs builds and funds AI-first companies in the UK. Site: moonlab.ventures.