Squeezed between the gigantic AI race being fought between the US, China and, to a lesser extent the EU, the United Kingdom, once a powerhouse country inside the EU, faces a difficult choice.
Prostrating itself before frontier AI companies from the US is not an edifying prospect, and it’s desparate to reduce its dependence on foreign AI infrastructure.
But building sovereign AI means a little more than wacking an open-source AI model onto some server in the British Isles.
UK startup Cosine is now setting out to train Lumen Sovereign, its own frontier AI model, and has won Uk government backing to do it. Its being developed from the ground-up on the Isambard-AI, one of Europe’s most powerful supercomputers, based in Bristol, using compute power awarded through the UK Government’s £500 million Sovereign AI programme. Cosine was named among the programme’s first cohort of supported startups in April 2026.
Cosine has now recruited some of the country’s largest defence, banking and infrastructure groups to help build a frontier AI model designed to operate without dependence on foreign technology or infrastructure.
These include BT, Babcock, Lloyds Banking Group, HSBC, LSEG, NatWest, BAE Systems, PwC, Thales and other major British companies and institutions. All have signed memoranda of understanding to participate in the design of Lumen Sovereign.
Speaking to Pathfounders, Cosine co-founder Yang Li said the government support included access to compute, introductions across the public sector and a right of first refusal to participate in the company’s next funding round.
Lumen Sovereign will use a mixture-of-experts architecture with roughly 1.35 trillion parameters, according to Li. While it may not be the largest publicly available model, he said its scale would put it “within range” of the frontier category.
Coalition members will help define the model’s use cases, security requirements and governance standards. Cosine is targeting deployment readiness by the end of 2026.
The model is being designed to run entirely inside customers’ own infrastructure, including air-gapped systems with no connection to external networks.
“Our USP here is that we can deploy the entire stack on premises, fully air-gapped, so there’s no data going in or out,” Li told Pathfounders.
Customers would be able to run inference and continue training the model internally, he said, “without data ever going to a third party, whether that’s a hyperscaler or an American provider”.
The initial applications are focused on highly regulated and security-sensitive work, including cybersecurity testing, anti-money-laundering and know-your-customer investigations, clinical trial coordination, legal document review and healthcare administration.
These are areas where using cloud-based AI is not feasible for security reasons. Defence contractors, banks and critical-infrastructure operators may be legally or operationally unable to send classified, financial or clinical information to externally controlled data centres.
Cosine argues that installing GPUs on customers’ premises could also make heavy AI use cheaper. Once the initial hardware cost has been absorbed, Li said, “the more you use the models, the cost per token decreases, because it’s a fixed cost”.
Founded in 2022 by Alistair Pullen, Yang Li and Sam Stenner, Cosine has raised only $8 million for far from investors including Lakestar, SOMA Capital and Gaingels. The Y Combinator-backed company previously developed Genie, an autonomous AI software engineer.
Cosine claims its models have beaten systems from OpenAI, Anthropic, Mistral and DeepSeek on independent coding benchmarks for two consecutive years. Its technology supports more than 38 programming languages, including Fortran, COBOL and Ada, legacy languages still used across defence, nuclear infrastructure and financial services.
Lumen Sovereign will be trained using proprietary datasets developed in-house and covering more than 30 regulated-industry workflows, rather than being adapted from an existing open-source model.
Data supplied by coalition members will be handled case by case. Li said some organisations may contribute data to improve the overall model, but the level of participation will depend on their security requirements.
“We are still working organisation by organisation to understand exactly what types of data are shareable,” he said.
Cosine says its focus will distinguish Lumen Sovereign from more general-purpose models, which can write poetry. Customers were instead asking for secure-by-design systems, modernisation of legacy codebases and smaller models that could run at the edge.
“Enterprises are increasingly waking up to the risk of being wholly dependent on foreign providers,” said Cosine CEO and co-founder Alistair Pullen in a statement. He said vendor lock-in created security, dependency and cost-escalation risks, and claimed Lumen Sovereign could operate at a lower price than alternatives from OpenAI and Anthropic.
Li said the larger US labs could theoretically offer fully on-premises versions of their models, but had so far chosen not to.
“There’s nothing stopping Anthropic or OpenAI from deploying fully on premises,” he said. “They can send their model weights and their IP into an organisation, but so far they’ve chosen not to. It’s a business decision.”
He argued that sovereignty should not be treated solely as a national-industrial policy.
“Don’t think of it as sovereignty. Think of it as control over your own destiny,” he said. “Control over what data goes in, control over how data is processed, and ultimately control over cost.”
Babcock said the model could provide a UK-native and customisable AI stack for complex defence environments, while LSEG pointed to secure deployment, trusted data and governance as prerequisites for adoption in regulated industries.
BAE Systems group CTO Rob Merryweather said in a statement that working with startups such as Cosine could strengthen the UK’s sovereign AI capability and create a “more dynamic and resilient UK industrial base”.
Cosine is also positioning the technology beyond Britain. Li said the company was already speaking to Swedish banks, Polish cloud providers and public-sector organisations elsewhere in the European Union.
The involvement of major prospective customers gives Cosine’s project more credibility than a purely government-led sovereign AI announcement, given that these institutions have genuine restrictions that cloud-delivered AI models may struggle to satisfy.
