Britain Bets on a 30-Person Startup, Cosine, in Its Push for Sovereign AI

Britain Bets on a 30-Person Startup, Cosine, in Its Push for Sovereign AI

When governments talk about "sovereign AI," they usually mean spending billions on data centres and courting the same handful of American labs everyone else is courting. Britain has taken a different bet. It has handed a slice of its most powerful supercomputer to Cosine, a London startup of roughly 30 people, and asked it to build a frontier AI model that lives, trains and runs entirely on UK soil.

It is a wager on focus over scale, and it has quietly become a cornerstone of the country's industrial strategy for AI.

The backing: compute, not just cash

The core of the government's support is not a headline grant but something scarcer: machine time. Under its £500 million Sovereign AI programme, the UK selected Cosine as a partner and awarded it 500,000 GPU hours on Isambard-AI, one of Europe's most powerful supercomputers, accessed through the national AI Research Resource. Cosine describes the allocation as worth "millions of pounds in infrastructure value."

The point is not the money. It is provenance. Compute has become the main bottleneck for smaller model developers trying to compete with far better-funded rivals, and access to state-backed infrastructure lets Cosine make a claim almost no one else in Britain can: that both training and deployment happen on domestic hardware. The government's venture arm has also secured an option to participate in Cosine's next funding round, tying the state's interest directly to the company's commercial fortunes.

Cosine was founded in 2022 by chief executive Alistair Pullen and chief operating officer Yang Li, alongside chief information officer Sam Stenner. It went through Y Combinator, has raised about $8 million from investors including Lakestar, SOMA Capital and Gaingels, and counts Michael Seibel and Monzo founder Tom Blomfield among its advisers. Its coding agent, Genie, has topped independent software-engineering benchmarks for two consecutive years, beating systems from OpenAI, Anthropic, Mistral and DeepSeek. That track record, more than its headcount, is what earned it a seat at the sovereign-AI table.

Why "sovereign" matters to this buyer

Cosine's pitch is aimed squarely at customers for whom mainstream AI tools are simply unavailable. Its software is built to run inside a customer's own infrastructure, air-gapped, with no internet connection and no data leaving the site. It supports more than 38 programming languages, including legacy ones like Fortran, COBOL, Ada and Verilog that still run defence systems, nuclear infrastructure and parts of financial services.

That design is the whole argument. "Cursor and Claude Code are outstanding products. They are also legally off the table for a lot of our customers," Pullen said when the compute award was announced. "The moment your work touches classified infrastructure, you need an AI that lives entirely inside your walls. We built Cosine that way from the start, because retrofitting security onto a cloud product is not the same thing."

His co-founder framed it as national policy. "The UK should be an AI maker, not an AI taker," Yang Li said, "and this is what that looks like in practice."

In June, that thesis grew teeth. Cosine unveiled a coalition of major UK institutions to co-design what it calls Britain's first sovereign frontier model, Lumen Sovereign. The signatories, who have signed memoranda of understanding to shape the model's design phase, read like a roll call of critical national infrastructure: Babcock, BT, Lloyds Banking Group, LSEG, NatWest, PwC, Thales UK, BAE Systems, Leonardo UK and Telefónica Tech UK&I. Lumen Sovereign is being trained entirely on Isambard-AI, built from proprietary in-house datasets spanning more than 30 regulated industry workflows rather than fine-tuned from an open-source checkpoint, with deployment readiness targeted for the end of 2026.

The bigger bet: dependency as strategy

The sovereign-AI logic mirrors older anxieties about semiconductors, energy and cloud computing. Just as institutions spent years reducing reliance on single energy suppliers or foreign-controlled communications networks, a growing number of Britain's defence, financial and public-sector organisations are reaching the same conclusion about AI: that outsourcing your most strategic technology to a handful of overseas providers is a risk, not a convenience.

"Enterprises are increasingly waking up to the risk of being wholly dependent on foreign providers," Pullen said in June. "Vendor lock-in creates security risk, dependency risk and cost escalation risks."

The obvious counter-argument is scale. OpenAI, Anthropic, Google DeepMind and Meta are spending sums that dwarf Cosine's entire funding history on single training runs. Can 30 people out-engineer that? Cosine's answer is that it does not need to win the general race. It needs to win a narrow, defensible one, where the requirement to run air-gapped and on legacy code disqualifies the giants before performance is even measured. In that segment, being trained on sovereign infrastructure is not a marketing line but a procurement checkbox. Focus, in other words, can beat scale when the buyer's constraints rule scale out.

The risk is equally clear. Frontier models are expensive to keep current, and a small team betting on a single flagship model is exposed if timelines slip or if a well-funded rival ships a "sovereign" tier of its own.

What to watch next

Three markers will show whether the bet is paying off. First, whether Lumen Sovereign hits its end-of-2026 deployment target and how its performance compares once it is running. Second, whether those MOU signatories convert into paying customers or quietly walk away. And third, whether the government's option on Cosine's next round is exercised, the clearest signal of how much the state is willing to double down on its 30-person champion.