London‑based Cosine has UK government backing to build a sovereign AI model, but its modest $15mn raise puts it at a disadvantage compared with rivals like OpenAI and Anthropic. Cosine, founded by Alistair Pullen and Yang Li, is developing Lumen Sovereign, a frontier model that will run on the Isambard supercomputer and UK data centres. The company’s strategy is to adapt and train large language models for highly regulated sectors, such as finance and defence, rather than compete directly with the largest AI labs.

Cosine’s approach uses a mixture‑of‑experts framework inspired by China’s DeepSeek and is roughly twice the size of Mistral’s Large 3 model. While the Isambard supercomputer will supply about 80 % of the required compute, Cosine plans to license Lumen Sovereign to others instead of operating it itself, avoiding the massive costs of running AI at scale.

The startup aims to release checkpoint snapshots in September and a wider public rollout by October, with a fundraising round scheduled later this year to finance development and hiring. Critics note the need for hundreds of millions more to compete, but Pullen stresses a “healthy chip on the shoulder” among his small team.

Cosine hopes its focus on sovereign, on‑premise AI will appeal to engineers currently pigeonholed in larger organisations.