Cosine rallies UK giants to build a sovereign AI model

Britain’s banks, telecoms, and weapons-makers have a new shared anxiety: that the AI they increasingly run on is built, owned, and controlled in the United States. A startup barely three years old is betting they will pay to fix it.

Cosine, a UK frontier-AI lab, has assembled a coalition of blue-chip British institutions to co-design Lumen Sovereign, what it bills as Britain’s first sovereign frontier AI model.

The signatories read like a roll-call of the national economy: BT, HSBC, Lloyds, NatWest, BAE Systems, Babcock, LSEG, PwC, Thales UK, Leonardo UK, and Telefónica Tech, each signing a memorandum of understanding to help shape the model’s use cases, security requirements, and governance.

The reveal was timed for the opening of London Tech Week, where Prime Minister Keir Starmer set out a more interventionist AI strategy and announced roughly £400mn of new spending on specialist AI chips to expand the country’s computing capacity. Britain’s next AI champions, he said, should “start here, scale here and stay here.”

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

The selling point is control. Lumen Sovereign will be trained entirely on Isambard-AI, the Nvidia-powered Bristol supercomputer that ranks among Europe’s most powerful, using compute awarded under the government’s £500mn Sovereign AI programme, which named Cosine in its first cohort in April.

Cosine says the model will carry no dependence on foreign infrastructure at any stage, and can be deployed inside a customer’s own systems, including air-gapped environments with no connection to outside networks.

That is the crux of the pitch.

For most companies, choosing an AI model is a procurement decision; for a defence contractor, a bank running anti-money-laundering checks, or an operator of critical infrastructure, it is a legal and security one first. Sending classified systems, AML alerts, or clinical data to a server in a US data centre is often simply not allowed.

“Enterprises are increasingly waking up to the risk of being wholly dependent on foreign providers,” said Cosine co-founder and chief executive Alistair Pullen, who argues that vendor lock-in brings “security risk, dependency risk, and cost escalation risks.”

Cosine is an unusual candidate for the job, and a credible one.

Founded in 2022 by Pullen, Yang Li, and Sam Stenner, the Y Combinator-backed lab has raised just $8mn from investors including Lakestar, yet its coding models have topped independent benchmarks against OpenAI, Anthropic, Mistral, and DeepSeek for two years running. It supports more than 38 programming languages, among them the COBOL, Fortran, and Ada that still run Britain’s defence systems and financial plumbing, and that most AI tools handle badly.

Lumen Sovereign, it says, will be built from scratch on proprietary datasets spanning more than 30 regulated workflows, rather than fine-tuned from an open-source model, with deployment targeted for the end of 2026.

The coalition partners are blunt about why they are in. “Cosine has offered us a path to a completely UK native and highly customisable AI stack,” said Peter Passaro, director of AI and data at Babcock, pointing to “the very complex defence environments we operate in.”

Priority uses span cybersecurity, KYC and AML investigations, legal document review, and healthcare administration, areas where UK AI adoption has stalled on exactly these security concerns.

Cosine’s effort is the sharp end of a wider British scramble for AI independence, one that keeps colliding with an awkward fact. The government has poured money into homegrown compute and chip supply, and struck sovereign-infrastructure deals with Nscale and Nvidia, even as flagship US projects such as OpenAI’s Stargate UK have stumbled.

Yet “sovereign” AI in Britain still tends to run on American chips, Isambard-AI among them.

For now, these are memoranda, not contracts, and Lumen Sovereign is still to be built. Training a frontier model from scratch to the assurance standards demanded by defence and finance, and doing it by the end of the year, is a formidable task, and Britain has announced sovereign-AI ambitions before.

What is different this time is the coalition: a row of institutions treating sovereign AI not as a slogan but as a procurement requirement.