Why AI Sovereignty Doesn't Matter

We're building a British sovereign AI model at Cosine. Perhaps that makes me an unlikely person to say the following: sometimes sovereignty really doesn’t matter.

There's a common and perhaps growing view that AI models are somehow better, safer or more worthy of our support simply because they're British. Train them on British soil, give them a British name (Sovereign?), change "elevator" to "lift" and "patronize" to "patronise", and, v̶o̶i̶l̶à!̶ Bob’s your uncle, our technological future is secured.

I'm not convinced.

Nor am I convinced by the opposite argument: that Britain should give up on building its own AI capability and outsource the whole business to the US and China (or, worst of all, to France!*). *To be clear and avoid a diplomatic incident, that is a joke.

AI sovereignty is important. But not in all the ways people think. Here's my view.

Three reasons AI sovereignty doesn't matter

1. Nobody wants the world's second-best AI, even if it's British

Let's start with something obvious that gets lost in the national enthusiasm.

AI is a global competition. The technology does not become inherently more useful because it was developed in Bristol rather than Beijing.

If a British model is materially worse than its American or Chinese competitors, customers shouldn't be expected to use it out of patriotic duty. Nor should governments spend vast sums protecting mediocre technology from competition.

The ambition has to be to build something genuinely world-class. That doesn't mean every model needs to be able to solve Navier-Stokes. But it does need to be state-of-the-art (or close to it) at something customers genuinely care about. Being the best British model is not, in itself, a particularly meaningful achievement.

At Cosine, we want to compete with the world's leading AI labs. We know we’re not there yet. But it’s our long-term goal. And we believe that sovereignty should be a consequence of building excellent technology, not an excuse for building inferior technology.

2. Where a model is trained is considerably less important than people pretend

Some people focus on the physical location of AI training. Was the model trained in Britain? On British infrastructure? Using British power?

These questions aren't entirely irrelevant. There are good reasons for wanting domestic computing capacity, not least the jurisdiction under which sensitive information is processed. And, of course, we are training our Lumen Sovereign models in Bristol, on Isambard-AI.

But a model doesn't acquire a mystical British quality because its training GPUs happen to be sitting in Bristol.

What matters much more is who controls the model. Who owns the weights? Who can modify it? Who decides how it behaves, where it can be deployed and who gets access?

A model trained in the UK but controlled by an overseas company might offer less meaningful sovereignty than a model originally developed abroad but owned, operated and adapted independently by its British users.

3. Complete technological sovereignty is a fantasy

It is definitively absurd that the UK can become entirely self-sufficient in AI.

We don't manufacture the leading AI chips. Good luck finding enough silicon, gallium and germanium in British soil to build an entire semiconductor industry. Our infrastructure relies on global supply chains. The research community is international. So are the customers, the capital and, quite rightly, the talent.

I'm an economist by background. Ricardo's theory of comparative advantage is one of the most widely accepted principles in economics, even if there are debates about its application. Countries should specialise in what they're relatively good at, rather than attempting to produce everything themselves. Trying to recreate the entire AI supply chain within Britain's borders would be an extraordinarily expensive way to ignore 200 years of economic thinking.

On the other hand, we should analyse where that dependence is acceptable, and where it becomes dangerous. Which brings me to the other side of the argument.

Three reasons AI sovereignty does matter

1. Because control matters

Imagine you're a defence organisation, a major bank or a government department.

You're increasingly relying on AI for work that is fundamental to your operations. Not just summarising emails, but writing software, interpreting sensitive information, operating critical systems and making consequential decisions.

How comfortable are you knowing that the intelligence behind those systems is something you don't really understand or control?

That someone else can change the model, withdraw access, alter the terms, refuse to disclose how the model was trained, increase the price, modify its behaviour without explanation, or decide that your particular application is no longer acceptable?

For everyday consumer AI, this may not matter much. For an organisation that has built its critical operations around a model, switching may be considerably harder.

This is why sovereignty matters at the customer level. It isn't principally about the nationality of the people who built the model. It's about having the ability to deploy, inspect, modify and operate it on your own terms. For sensitive workflows, that can include air-gapped AI deployment, keeping data, prompts and outputs inside an organisation’s security perimeter.

The distinction is between using someone else's intelligence and having meaningful control over your own.

2. Because the global AI market is becoming an uncomfortable duopoly

Consider the choices facing organisations looking for serious AI capability today.

At one end, you have a handful of American companies producing extraordinarily capable, largely proprietary frontier models. OpenAI, Anthropic, Google.

At the other, you have a rapidly advancing collection of Chinese model developers, including DeepSeek, Kimi and Qwen, many of whom have made powerful models available with open weights.

There are important exceptions, including European and American open-weight models. But the broad direction is unmistakable.

The frontier is increasingly dominated by two countries, with very different approaches to technology, commerce and the relationship between the state and private enterprise.

Both approaches have merit. American labs have produced astonishing breakthroughs. Chinese open-weight models have helped make powerful AI much more accessible.

But should the rest of the world really be content with choosing between them?

This isn't an argument against American or Chinese technology. We should use the best technology available, wherever it comes from.

It's an argument for having a third option.

Competition is good for customers. Diversity is good for resilience. And an AI ecosystem in which almost every consequential technical decision is ultimately made in either Washington's or Beijing's orbit is not a particularly healthy one.

3. Because Britain cannot outsource its national security

This is the argument I find most compelling.

For decades, Britain has accepted substantial dependence on allies for important technologies. Much of that has been entirely sensible. We are not going to build our own alternative to every American defence system or semiconductor.

But AI is different in one important respect.

It is increasingly becoming a foundational capability, not simply another product we purchase.

It will underpin how intelligence is analysed, how cyber threats are identified, how military systems operate, how critical infrastructure is maintained and, potentially, how entire government departments function. These concerns emerged when we asked what critical sectors need from sovereign AI, with input from defence, finance, energy, and healthcare organisations.

Do we really want all of that to depend on technology whose continued availability is ultimately determined elsewhere?

The United States is our closest strategic ally. That doesn't mean our national interests will always be identical. In June 2026, the US Government ordered Anthropic to suspend foreign-national access to two of its most advanced models. Later that month, OpenAI restricted the initial rollout of its latest models to selected US-based partners at the government’s request. Export controls, commercial decisions and political priorities can change, sometimes very quickly.

Nor is this an argument for technological autarky. Britain should continue working closely with American, European and other partners.

But there is an enormous difference between choosing to depend on an ally and having no alternative.

Sovereignty, in this context, is a form of insurance. You hope never to need the independence it provides. But the consequences of not having it could be enormous.

What meaningful AI sovereignty looks like

Sovereignty isn't a binary characteristic. It isn't something a model acquires simply by being trained on British soil, or by having a British company name on the documentation.

It's a question of how much control we have, where that control sits, and whether we could continue operating if circumstances changed. Our practitioner’s guide to sovereign AI explores how to assess that control across the AI stack.

For some applications, the answer is that we don't need much control at all. The world's best commercial model, accessed through an API, will do perfectly well.

For others, particularly in defence, national security and critical infrastructure, the answer is very different.

Britain doesn't need to win every part of the global AI race. But it would be extraordinary to decide now that the only sensible future is one in which the US and China build the intelligence on which everyone else depends.

Britain, alongside other like-minded democracies, still has an important role to play in shaping what comes next. Not by building a protected domestic industry of second-rate models, but by developing technology that is genuinely competitive and gives organisations around the world a meaningful alternative.

This is why we've chosen to build Cosine in the UK. We're dedicated to giving our clients control and security over the AI they use. We're proud to work with organisations across British and European defence, healthcare and critical infrastructure, helping them adopt AI without surrendering control of their most sensitive systems.

But our ambition goes beyond Europe. We want to help build a world in which countries and organisations have genuine choices about the technology they depend on; where access to the most powerful AI isn't determined by a handful of companies or governments; and where excellence and independence aren't mutually exclusive.

We are not making British AI because AI needs to be British. We are making brilliant AI that happens to be British, and that organisations can control.

To my mind, that is an ambition worth pursuing, and we’re dedicated to giving our clients control and security over the AI they use.