Real AI sovereignty lies in funding the problems, not the products

Looking back, the moment I knew Britain had an investment problem came during our application for a grant from Innovate UK, the UK’s national innovation agency.

It demanded more detail than any VC due diligence I had ever seen – Gantt charts, multi-stage risk matrices, and an endless loop of residual risk mitigation. A reasonable ask for a PhD thesis; less so for a pre-revenue startup trying to get to market. It is a process built neither for ambition nor speed.

After two rejections from Innovate UK, and years of unsuccessful pitching to UK investors, I took my startup, Safe Sign Technologies, to North America and raised just over £2 million. Two years later, Thomson Reuters acquired the company, in its first pre-revenue acquisition.

It is a story we will see more of during the current AI boom. Britain builds world-leading AI companies, then lets them go.

Everyone reaches for DeepMind in this context. Founded in London in 2010, and bought by Google in 2014, it is the archetypal example of us being unable to scale brilliant technology. What happened afterwards is more revealing. Research by Evertrace found that 112 DeepMind alumni have founded startups or moved into stealth roles in the last eighteen months. 70 are in the United States; 28 are here.

Britain’s “founder factory” works; it reliably produces globally competitive AI companies. Tax, regulation and talent costs then push those same companies elsewhere.

According to DWF, late-stage funding accounted for roughly 20% of total UK venture investment in 2024, against 35% in the US. That is the capital that turbocharges growth after product-market fit.

Of the late-stage money that does reach the UK, over 60% comes from overseas investors, and 42% from America alone. When a British AI lab needs conviction capital to scale, the decisive cheque is unlikely to be written in London.

What we have actually bought

The Sovereign AI Fund has announced five investments since April. Four of the five are infrastructure or applications built on research that is already largely settled. Only one, Ineffable Intelligence, is a bet on an unsolved scientific question, and there the fund took a slice of well under 1% of a $1.1 billion round that Sequoia and Lightspeed had already filled.

This is not a failure of execution, but of the mandate. James Wise, who chairs the unit, has said plainly that it is “not a grant-making institution” and is “here to invest our resources and capital on a commercial basis”. The fund is doing what it said it would: conventional venture capital, backing revenue, in markets that already exist, on science somebody else has already de-risked. Even for investors who only care about the return, that edge narrows as capability built on top of a model gets captured by whoever owns the model underneath. A sovereign fund should instead be pioneering and owning the AI equivalent of critical national infrastructure. Not just the data centres and the chips, but the unsolved layers everything above them depends on.

Sovereignty is owning the unsolved parts

AI Minister Kanishka Narayan said that if Britain wants to lead in AI we need to back the technologies that sit underneath it. He is right, but the fund’s portfolio does not yet reflect it.

A handful of scientific problems block the entire AI frontier, and the fund’s portfolio barely touches them. Models cannot learn continually, retraining from scratch at enormous cost each time. They cannot be verified or controlled well enough to be trusted in medicine, aviation, energy or law. Trust and safety is one of the fund’s advertised priority sectors with no investment in it at all. What’s on offer is evaluation and audit tooling, not the formal verification that would let a model fly an aircraft or read a scan.

Causal and world models – systems that understand cause and can act on the physical world – don’t appear in the fund’s portfolio. Closed-loop scientific discovery, where AI and robotics run experiments and learn from them, likewise remains unachieved and unmentioned.

Nothing built above those layers works properly until they are solved, and whoever solves them owns a position that is hard to dislodge. Britain has, at various points, owned exactly such positions: ARM in instruction sets, DeepMind in reinforcement learning. In each case the value came from holding a layer everyone else had to build on. These are the foundations that will determine what AI can do in ten years. Britain has an academic lead in several of them, but no capital pointed at any.

The basics, which have worsened

Business Asset Disposal Relief, the reduced rate founders pay when they sell, went from 10% to 14%, then again to 18% since April this year, while the lifetime cap has been stuck at £1 million since 2020. It’s the clearest tax incentive Britain has for founders to build and exit here, but it’s moving in the wrong direction.

See Also

Metal Morph raises pre-seed to recover valuable resources from wastewater

Employer National Insurance has gone up, and the threshold at which it starts fell by almost half to £5,000. Support claimed by SMEs under the R&D tax relief scheme fell 29%, while total support across all schemes fell only 2%, because the relief was rebalanced toward larger companies. At the same time, the cost of a visa for a researcher coming to Britain is 22 times the average of seventeen other leading science-focused nations.

The politics have not helped either. Just a day after the new Prime Minister took office, the department behind the Sovereign AI Fund, DSIT, was abolished. Kanishka Narayan, the first AI minister to attend Cabinet, has called AI “the most significant technology in human history”, but it is too early to say whether the reorganisation will leave anyone with the authority to act on that.

What would actually change things

In the two years I spent trying to keep Safe Sign British, there wasn’t a shortage of money. It was a shortage of belief, and the technical judgement that belief has to rest on.

Investors were comfortable backing applications of models that already worked, and almost nobody was equipped to assess a team proposing to build a new one. The result is that capital arrives only once an area has become legible to generalists, by which point the price already reflects it.

Fixing that is not primarily a policy question, though the policy asks are worth making: restore the exit incentive, make it cheaper to hire, and replace grant processes built around risk-mitigation paperwork with ones built around technical assessment. The deeper fix is that somebody has to be able to tell, early and with confidence, whether a hard technical claim is true. That means putting people who understand the science inside investment decisions, and holding those judgements closer to the standard of peer review than of a partner meeting. It is the only thing that works in a market that misprices what it cannot evaluate.

Britain can produce top-tier AI companies; it does, repeatedly. Whether it also owns the layers those companies are built on depends on whether anyone here is willing to fund the unsolved problems and can judge them before somebody in San Francisco does.