Sovereign AI development runs into the same constraint every national model project eventually meets: compute scale. Training a language model capable of contending at the frontier demands accelerator access and capital that most governments do not procure directly. London startup Cosine, a 30-person team with UK government backing, is attempting exactly that, positioning itself as Britain's domestic answer to OpenAI.
The sovereign model mandate
The term "sovereign AI" describes a system trained and governed inside a single national jurisdiction, without routing data through infrastructure subject to another country's legal or corporate control. For UK policymakers, that distinction matters. The frontier models government ministries might otherwise procure are American, and Cosine's pitch is that Britain should not have to depend on foreign-controlled endpoints for sensitive workloads. The UK government has decided to back that argument.
A gap measured in resources
The source describes Cosine's competitors as far better resourced. OpenAI, the direct comparison the source draws, has assembled engineering headcount and raised capital at a scale no 30-person team can replicate from a standing start. UK government backing may improve Cosine's access to compute procurement channels it could not reach alone. The capability gap, though, is ultimately determined by what the training budget can sustain across successive model runs.
Thirty people against the frontier
Thirty people is enough for a disciplined specialist research program. Whether it supports a general-purpose frontier model that can genuinely compete with OpenAI depends on what the government backing unlocks in practice, and the source does not detail the terms. What the UK government has publicly signaled is that a 30-person London startup is its chosen vehicle for sovereign model development. Cosine's headcount stays at thirty.