UK backs Cosine with GPU hours for sovereign AI push

The UK government has selected Cosine as a partner in its Sovereign AI programme, giving the British startup 500,000 GPU hours on the Isambard-AI supercomputer.

The allocation will allow Cosine to train what it describes as an AI coding model built, owned and deployed entirely in the UK. The programme has a budget of GBP £500 million, and its venture arm has secured an option to participate in Cosine's next funding round.

Cosine, a London-based AI startup founded in 2022 by Alistair Pullen and Yang Li, says its coding models have outperformed systems from OpenAI, Anthropic, Mistral and DeepSeek on independent coding benchmarks for two consecutive years.

The company focuses on customers in defence, critical infrastructure and other regulated sectors, where data-handling rules can restrict the use of foreign-hosted AI tools. Its software is designed to run within a customer's own infrastructure, without an internet connection or data leaving the site.

UK focus

The award gives Cosine access to one of Europe's most powerful supercomputers through the UK's AI Research Resource. The company describes the compute allocation as worth millions of pounds in infrastructure value and says it closes a remaining gap in its claim to full sovereign AI development.

That is particularly relevant in sectors that depend on older software environments. Cosine says its platform supports more than 38 programming languages, including Fortran, COBOL, Ada and Verilog, which are still used in defence systems, nuclear infrastructure and parts of financial services.

Many commercial coding assistants have gained traction with mainstream software teams, but their use can be limited in environments handling classified or sensitive systems. In those organisations, where code cannot be sent to foreign-managed servers, on-site deployment is often a basic requirement rather than a preference.

Cosine says it is already working with UK defence prime contractors and critical national infrastructure operators, including organisations involved in the country's nuclear deterrent programmes and next-generation defence platforms. It did not identify those customers.

In comments on the announcement, the company drew a clear distinction between mainstream coding assistants and products designed for classified settings. "Cursor and Claude Code are outstanding products. They are also legally off the table for a lot of our customers," said Alistair Pullen, Chief Executive Officer and Co-Founder of Cosine. "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, and the people we work with know the difference. What we are building now goes further: a truly British AI lab, producing sovereign models for Britain's most critical use cases, owned and controlled by the UK, for the benefit of the UK."

Funding and growth

The government backing comes as European policymakers and startups push for greater domestic control over AI infrastructure and models. Concerns about dependence on US providers have grown alongside rising demand for AI systems in defence, public services and strategically important industries.

Cosine has raised USD $8 million from investors including Lakestar, SOMA Capital and Gaingels. It is also advised by Michael Seibel and Tom Blomfield, and took part in Y Combinator during its early development.

According to the company, Pullen has worked on AI products since 2018. Li previously helped scale bike-sharing company Mobike to 220 million users across four continents before its USD $55 billion acquisition.

Cosine presents itself as both an AI lab and a product company, developing its own models while also selling software built on top of them. That combination may become increasingly important for governments seeking domestic AI suppliers that can offer both model development and deployable tools for public and regulated sectors.

The compute award also reflects a broader UK effort to use state-backed infrastructure to support home-grown AI companies. Access to large-scale compute has become one of the main bottlenecks for smaller model developers trying to compete with larger international rivals.

For Cosine, the immediate significance is that it can now say both training and deployment take place on UK infrastructure. "For two years we've been telling defence primes and critical infrastructure operators that we can do what no one else can: air-gapped, on-premise, trained on the legacy code that runs Britain's most sensitive systems," said Yang Li, Chief Operating Officer and Co-Founder of Cosine. "The one thing we couldn't say was that the model itself was trained on sovereign infrastructure. The AIRR grant completes that picture. The UK should be an AI maker, not an AI taker - and this is what that looks like in practice."