What I believe the UK government and UK businesses should actually do about it, and why they should start today.
Like many other early adopters of AI, we were bright-eyed and excited evangelists, preaching the gospel of the LLM to any who would listen, and to one or two who would have preferred not to.
While most businesses were still trying to work out whether the whole thing was just a short-lived fad, we had already incorporated tools from OpenAI and Anthropic into our day-to-day work here at Katatomic.
Whether we were perfecting custom code, drawing up technical architectures, brainstorming web solutions, or researching whether the gravitational differences between Earth and Mars would afford a significant maximum height difference to a squirrel on a pogo stick, the efficiency gains were clear.
We welcomed the change, structured every aspect of our operations around it, and promoted it to all our friends and peers in business. We soon saw our productivity rise exponentially. But the more dependent we became, the more an inner voice, tiny at first but bit by bit more vociferous, began to trouble me.
For a business now dependent on AI, operating in a nation increasingly dependent on AI, what would happen if access to this wondrous jinn were suddenly denied us?
This internal musing ceased being hypothetical back in June, when the US government issued an export control directive requiring Anthropic to cut off access to its most powerful models, Fable and Mythos, for every foreign national on the planet.
A few weeks later access was restored, but by then a clear lesson had been learned. Namely, that one nation could, without warning, dictate exactly who worldwide would be allowed to use a piece of software. The kill switch was no longer a hypothetical idea for hyperbolic journalism, but a demonstrated reality. This meant that any country lacking its own capabilities could, in principle, be switched off, pressured, or gradually reduced to the rank of a vassal state at another nation’s discretion.
My initial logic was flawed
As soon as my niggling fear became reality, my initial conclusion was that the only real solution was for the UK to develop its own sovereign AI. My reasoning was simple. As a nation, we could not afford to remain under the control of capricious foreign governments or unreliable overseas technology monopolies, since we would be at the mercy of whichever administration was in power and whatever mood it happened to be in. If AI is the new electricity, or the new nuclear weapon, then it follows that the United Kingdom should have its own domestic, state-supported version.
However, when I looked more carefully into what that would really involve, my initial conclusion had to change.
I had originally thought that creating a British frontier model was mainly a software problem with some political will attached. In fact, the obstacles are enormous and mostly physical. To compete at the very frontier, the UK would need hundreds of thousands of specialised chips, billions of pounds of initial capital, and an ongoing financial commitment far greater than anything the public purse could plausibly afford. As the Tony Blair Institute pointed out recently, trying to run that race from a standing start is little more than a very costly assurance of finishing second.
Why, regrettably, we need to learn to love the data centre and the power plant
The physical limitations are even more alarming, because artificial intelligence is not weightless software floating around in some abstract cloud. It is as much heavy industry as steel or shipbuilding.
Training and running models on this scale requires enormous data centres that consume electricity measured in gigawatts, and our national grid cannot currently cope with this scale of demand. The government has now gone as far as designating data centres as Critical National Infrastructure, on the same level as the grid and the water supply, and has set a target of tripling our capacity to about six gigawatts by 2030. But achieving this aim will involve years of planning disputes, the construction of new substations and extensive civil engineering work before a single competitive model can be trained.
There’s also the question of heat. When you pack the required number of processors together, they give off so much heat that standard air cooling fails, which is why the latest machines, such as the Zenith system being introduced in Cambridge, need specialised liquid cooling and the water supply to go with it.
A more pragmatic kind of sovereignty
As a result, I had to reconsider what true resilience really entails, and I now think that our national salvation does not lie in building a world-beating national model from scratch. The practical approach is instead to build our sovereign capabilities on existing open-source foundations.
Open models such as Llama, DeepSeek and Mistral have improved at a genuinely surprising pace. For the vast majority of everyday business tasks, the gap between them and the closed US systems has shrunk from a chasm to a narrowing crack.
If our national strategy were to host, fine-tune and protect these open models here in the UK, it would be possible to achieve real sovereignty without emptying the Treasury or waiting ten years.
The key, then, is not building the models themselves but building a national data infrastructure to serve our own customised and enhanced versions of open source models through a UK cloud beholden to no other nation.
What your business can do this week
Whitehall operates at Whitehall’s own pace, and in a financially stretched nation, the vision and investment required, however critical to the national interest, are unlikely to arrive with sufficient speed or scale.
For those of us in business, it would therefore be unwise simply to wait for the government to formulate a national solution before taking steps to protect our own operations.
If your business currently cannot operate efficiently or competitively without a single US chat window, then the risk to you is not hypothetical. What you need to do with a degree of alacrity is reduce your dependence on the major closed platforms and start building some redundancy into your daily work processes.
In reality, this is much easier than it sounds. All you have to do is become independent of any particular model, so that if one service evaporates, triples its price, or decides that the UK is no longer a priority market, you can simply shrug and make the switch. There are three good reasons why you should start doing this today.
The first is money.
The biggest misconception about open-source AI is that you need racks of extremely expensive graphics cards humming away in the cellar. You don’t. Hosted services such as ‘Groq’, ‘DeepInfra’ and ‘Together AI’ run these models on highly efficient hardware and charge per token, usually at a fraction of the cost of the mainstream frontier models.
Those savings go straight to your bottom line.
The second is data.
Each time your business pastes a client brief, your proprietary code, or your commercial strategy into a standard chat box, you place your trust in Ts&Cs you’ve probably never read, and therefore quite possibly allow that information to be used to train someone else’s model. If you run open-source models on a service that prioritises privacy, your commercial secrets remain yours, which is most likely exactly what you and your clients would expect.
The third is IT savviness.
None of this requires an engineering team or command line prompting. Open interfaces such as ‘LibreChat’, ‘Open WebUI’ and ‘HuggingChat’ give you and your team the same familiar chat window, while quietly connecting to cheaper, more private models behind the scenes. The experience barely changes, but the independence, improved privacy, and lower costs are all yours.
Become your own digital nation state
I used to think that securing our digital future was something to be done from the top down by the state, with someone in Westminster commissioning a massive British artificial intelligence.
I have now reached the opposite conclusion. Real sovereignty is built from the bottom up, by adopting habits that don’t depend on any particular model, by supporting open technologies, and by refusing to rely entirely on a single overseas supplier.
There is no need to abandon the big commercial platforms for our most difficult problems. We just need to stop seeing them as the only option.