The compute fence: Pax Silica, cloud export controls, and where UK founders now sit

A fortnight ago I wrote here about the moment a government took two frontier models offline in ninety minutes, and what that revealed about building on capability you do not control. That was the model layer. This week the story moved down a layer, to the thing underneath every model, and it is a bigger story.

On 27 June the second Pax Silica summit closed in Washington with the US-led semiconductor and AI alliance expanded to twenty-four formal signatories, and thirty-five countries signing a joint AI Opportunity Statement. In the same window, a US bill called the Remote Access Security Act has been working through Congress that would treat giving a foreign person remote access to a controlled GPU, over the cloud, as an export transaction subject to the same licensing as shipping the physical chip.

Read those two developments together and the picture is stark. Access to advanced AI is being fenced not at the model, and not even at the physical chip, but at the compute itself, including compute you only ever touch through a browser. For anyone building in Britain, that changes where the risk now lives.

From the model to the metal

The June shutdown taught founders that a model is a permission, not a possession. The lesson of this week is that the permission goes all the way down.

Most UK AI companies never touch a GPU directly. They rent capability from an American cloud, through an API or a hosted instance, and treat the compute beneath as weather, always there, someone else's problem. What the Remote Access Security Act proposes is that the act of a non-US person using that remote compute is itself an exportable event. The physical chip never leaves an Oregon data centre. The export, in this framing, is the access.

If that becomes law in anything like its current shape, the implication for a British founder is direct. The cloud GPU capacity your product runs on is no longer a neutral utility priced by the hour. It is a controlled item, and your ability to use it from outside the United States becomes a question of licensing and alignment rather than a question of your credit card clearing. This is not a tail scenario dreamed up by pessimists. It passed the House by 369 votes to 22 in January. The direction of travel is not subtle.

What Pax Silica actually is

Pax Silica is the other half of the same move, and it is the half that determines which side of the fence you end up on.

Launched in December 2025 and coordinated by the US State Department, it is a framework for aligning partner countries on semiconductors, compute, critical minerals and the infrastructure around them. The second summit on 27 June brought in the European Union, Germany, the Netherlands and others, and produced concrete machinery: a credentialing pilot in Panama, nicknamed Pax Pass, to fast-track vetted shipments of chips and AI infrastructure, and a Stanford-linked workforce programme for member economies. It is not a talking shop. It is the beginning of an operating system for who gets compute and on what terms.

The organising principle, stated more or less openly, is that access to AI infrastructure is conditional on political alignment. China is outside the fence by design. The members are inside it, and inside it they get faster access, co-investment, and a seat at the table where the rules are written. The good news for us is simple: the UK is a signatory. Britain is inside the tent.

The uncomfortable news is what being inside the tent actually means when you are not the one holding the pen.

Inside the tent, holding someone else's pen

The UK's own AI posture is pro-innovation and light-touch: sandboxes, sectoral rules, a growth-first blueprint rather than heavy legislation. That is a reasonable stance for regulating what British companies build. It does almost nothing to address where the compute they build on comes from, or who controls the licensing regime that governs it.

Because on the compute question, Britain is largely a rule-taker. The advanced chips are designed by an American company and made by a Taiwanese one. The alliance that governs their distribution is run from Washington. The export regime that may soon treat cloud access as a licensable act is US law. The UK can be a loyal and valued member of Pax Silica and still find that the terms of its members' compute access are set by another government's domestic politics, the same politics that switched off Fable in June.

Being inside the tent protects you from the worst outcome, which is being outside it. It does not make you sovereign. It makes you a well-treated dependent. And the history of dependents is that the terms are generous right up until the moment they are not, at which point you discover how little of the thing you actually controlled.

What this means for founders building now

I am not raising this to alarm anyone into inaction. Pax Silica membership is good for the UK, and the sensible move for almost every founder is still to use the best compute available and ship. But the risk register has changed, and pretending it has not is how you get caught.

Three things follow for the companies we build and back. The first is that compute jurisdiction is now a design decision, not an afterthought. Where your GPUs physically sit, whose legal regime governs your access to them, and whether you could move to a different provider in a different jurisdiction are questions worth answering before you are forced to. The second is that a credible fallback matters more than ever, and increasingly that fallback is domestic. This is exactly why the UK's Sovereign AI Fund, the compute it has put behind Cosine on Isambard-AI, and the wider push for British-controlled infrastructure stopped being industrial-policy theatre and became something a founder should actually care about. Compute you can reach without a foreign licence is worth a premium now. The third is that alignment has become a live variable in a way it never was. The question a serious investor asks is no longer only whether your model can be switched off, but whether the compute beneath it sits on the right side of a fence that is still being built.

None of this argues for retreat from the frontier. It argues for building with your eyes open about what you are standing on. The founders who treat compute as weather will be fine, right up until the weather becomes policy. The ones who treat it as infrastructure they need to understand, diversify and where possible domesticate will be the ones still shipping when the terms change.

The layer under the layer

Every few weeks now the fence moves down a level. First it was which countries could buy chips. Then it was which models could be served. Now it is whether a British engineer accessing an American cloud GPU counts as an export. Each step makes the same point more loudly: the compute that AI runs on is the most contested resource in the world economy, and the era of treating it as a neutral, borderless utility is over.

Britain has played this reasonably well. We are inside Pax Silica, we have a sovereign fund putting serious compute behind companies that need it, and we have a pragmatic regulatory posture. But inside the tent is not the same as holding the pen, and a founder who understands the difference will make better decisions over the next two years than one who assumes the compute will always just be there. It will be there. The only question is on whose terms, and that question now has a geopolitics all of its own.

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.