Sovereign AI opens with a Callosum stake and six compute allocations

On 16 April the Department for Science, Innovation and Technology (DSIT) launched Sovereign AI, a £500m unit that it describes as different from any previous government-backed body, “acting like a venture capital fund with the muscle of the state behind it”. Its first equity investment is in Callosum, a London startup building software that makes different chip architectures work together as one system. Six more startups – Prima Mente, Cosine, Cursive, Doubleword, Twig Bio and Odyssey – get access to the AI Research Resource (AIRR), the government’s network of supercomputers.

Cash is only part of the offer. Each backed startup can get fully funded time on the UK’s largest AI supercomputers, up to 1 million graphics processing unit (GPU) hours, according to DSIT. Companies that take investment get visa decisions within a working day and an initial 10 cost-free visas for research and development staff. The unit also promises help with “access to data, early procurement opportunities, independent product validation and routes into new approaches to regulation”.

The question for UK organizations building or buying AI is whether a fund of this size can change which British suppliers exist in three years’ time – and whether being picked by the state should count for anything when choosing one.

What the state has bought so far

Callosum was founded by two Cambridge PhDs, Danyal Akarca and Jascha Achterberg. DSIT’s announcement does not give the size of the stake. Speaking at the launch, held at the British self-driving company Wayve, the technology secretary Liz Kendall said the unit had already made two direct investments, with the second “to be announced soon”. She added that it had signed five ‘right of first refusal’ deals, which give it the option to invest when those companies next raise money.

The compute allocations are the more visible commitment. DSIT says the six recipients were chosen through an open, competitive call and assessed on strategic relevance, technical quality, scaling potential and “material compute need”. It says the unit is in talks with around 30 more firms about AIRR access and expects to allocate compute worth tens of millions of pounds to British startups over the year. A first funding call from its £282m research and development offer will pay for new datasets and other assets.

Kendall stressed speed and independence. The unit is chaired by James Wise, with Josephine Kant as head of ventures, and she said its investment committee makes its own decisions “free from political interference”. She also said it would pay the legal fees of any startup it works with that wants to ‘flip’ into a UK limited company.

Why compute rather than cash

The design reflects where UK startups are short. Crunchbase data published on 14 April shows that European venture funding reached $17.6bn (£13bn) in the first quarter of 2026, up nearly 30% year on year, with AI companies taking more than half for the first time. UK startups raised $7.4bn (£5.5bn), more than in the same quarter a year earlier.

The headline growth hides a narrowing market. Crunchbase counted 40% fewer European deals than a year before. Seed deal volume fell 44% and early-stage volume 30%, while early-stage funding dropped by around 20% to $5.3bn (£3.9bn). Late-stage funding, by contrast, rose 91% to $9.2bn (£6.8bn) across 83 deals. More money is going into fewer companies, and more than half of it went to late-stage rounds.

Beside those numbers, £500m in total is modest: UK startups raised roughly 11 times that in one quarter. What the unit can offer that private funds struggle to match is hardware. DSIT describes access to large-scale infrastructure as “the binding constraint on progress” for the companies it has chosen, and GPU time on national machines is something no seed fund can supply at the same speed. Alex DePledge, the chancellor’s entrepreneurship advisor, framed the government’s diagnosis as “a scale problem” rather than a shortage of talent.

There are limits to that logic. A state investor with a national-security mandate will favor some categories over others, and the first cohort shows it. DSIT says the six span biological foundation models, world simulation, sovereign inference infrastructure, agentic AI, engineering biology and AI for national security. Kendall’s own line at the launch was “go big or go home”. That suits frontier research. It says less about the applied AI companies that sell to ordinary UK businesses.

What it means for organizations buying AI

For buyers, the more relevant fact is what several of the backed companies sell. DSIT says Doubleword provides inference and model governance tooling that lets regulated and government users run and audit models inside their own secure environments “rather than depending on foreign cloud platforms”. Cosine builds coding agents for defense, national security and regulated industries, and DSIT says its platform can run entirely within a customer’s own infrastructure. The announcement says Callosum’s software can cut the cost and raise the performance of running models across mixed hardware. Akarca said in the release that “the future of compute is heterogeneous, and making that complexity usable is the next frontier.”

Those capabilities matter to mid-sized firms in regulated sectors and in the public-sector supply chain that want models they can host, inspect and keep under UK control. State backing may give these suppliers a longer runway than a thin seed market would, and the promise of “early procurement opportunities” is likely to put some of them in front of government buyers first. A public-sector reference customer would make them easier to buy for everyone else.

State backing does not do the buyer’s due diligence. Right of first refusal is an option to invest, not a guarantee of follow-on money, and the unit’s total budget could be absorbed by a handful of later rounds. A startup on national compute today still has to win a priced round tomorrow. Buyers should also expect the sovereignty language to spread to suppliers with no state backing at all, which makes the specifics – where the model runs, who holds the keys and where support staff sit – more important than the label.

The sensible course for UK organizations weighing a British supplier for sensitive AI workloads is to treat Sovereign AI backing as evidence of runway, not of quality. Test where the model runs, who can audit it and what the exit terms are if the next round does not close. The state has answered the last of those questions only in part.