Cosine unites UK giants to create a sovereign AI model.

      The banks, telecom companies, and weapons manufacturers in Britain share a new concern: the AI systems they increasingly rely on are developed, owned, and controlled in the United States. A startup that's only three years old is wagering that these companies will invest in a solution.

      Cosine, a UK frontier-AI lab, has brought together a coalition of major British entities to collaboratively design Lumen Sovereign, which it claims is Britain’s first sovereign frontier AI model.

      The partners include prominent organizations such as BT, HSBC, Lloyds, NatWest, BAE Systems, Babcock, LSEG, PwC, Thales UK, Leonardo UK, and Telefónica Tech. Each has signed a memorandum of understanding to assist in defining the model’s applications, security needs, and governance.

      This announcement coincided with London Tech Week, during which Prime Minister Keir Starmer unveiled a more proactive AI strategy and announced around £400 million in new funding for specialized AI chips to enhance the country’s computing capabilities. He noted that the upcoming AI leaders should "start here, scale here and stay here."

      The main selling point is control. Lumen Sovereign will be exclusively trained on Isambard-AI, a supercomputer powered by Nvidia that ranks among the most powerful in Europe, utilizing computing resources allocated through the government's £500 million Sovereign AI program, which recognized Cosine in its initial cohort in April.

      Cosine asserts that the model will have no reliance on foreign infrastructure at any point and can be implemented within a client’s systems, including isolated environments that are not connected to external networks.

      This is the crux of their proposal. For many businesses, selecting an AI model is a straightforward procurement choice; however, for a defense contractor, a bank managing anti-money-laundering checks, or a critical infrastructure operator, it primarily revolves around legal and security considerations. Transferring classified systems, AML alerts, or clinical information to a server in a US data center is often prohibited.

      “Companies are increasingly aware of the risks associated with being entirely dependent on foreign providers,” stated Cosine co-founder and CEO Alistair Pullen, who believes that vendor dependency introduces “security risks, dependency risks, and cost escalation risks.”

      Cosine may seem like an atypical yet credible candidate for the project. Founded in 2022 by Pullen, Yang Li, and Sam Stenner, the Y Combinator-supported lab has secured just $8 million from backers including Lakestar, yet its coding models have excelled in independent benchmarks against OpenAI, Anthropic, Mistral, and DeepSeek for two consecutive years. It accommodates over 38 programming languages, including COBOL, Fortran, and Ada, which continue to be vital for Britain’s defense systems and financial operations, and which most AI tools struggle with.

      Lumen Sovereign, they claim, will be developed from the ground up on proprietary datasets that encompass more than 30 regulated workflows, rather than being fine-tuned from an existing open-source model, with a deployment goal set for the end of 2026.

      The coalition partners are clear about their motivations. “Cosine has provided us with a pathway to a fully UK-native and highly customizable AI framework,” noted Peter Passaro, director of AI and data at Babcock, highlighting “the very complex defense environments in which we operate.”

      Key applications include cybersecurity, KYC and AML investigations, legal document reviews, and healthcare administration—areas where AI adoption in the UK has faced challenges primarily due to security issues.

      Cosine’s initiative represents the forefront of a broader British push for AI independence, a movement that continually grapples with a challenging reality. The government has invested heavily in domestic computing and chip production and established sovereign-infrastructure agreements with Nscale and Nvidia, even while prominent US projects like OpenAI’s Stargate UK have faced difficulties.

      Nonetheless, “sovereign” AI in Britain still predominantly relies on American chips, including Isambard-AI.

      For the moment, these remain memoranda, not contracts, and Lumen Sovereign is yet to be realized. Building a frontier model from scratch that meets the rigorous assurance standards required by the defense and finance sectors and completing it by the end of the year is a significant undertaking, especially since Britain has previously announced sovereign-AI ambitions.

      What sets this initiative apart is the coalition: a group of institutions viewing sovereign AI not merely as a concept, but as a necessity in their procurement process.