Liz Kendall, UK Technology Secretary, formally launched the £500m Sovereign AI Unit on 16 April 2026 and named all seven first-cohort investees , : Callosum (multi-model interoperability), Prima Mente (neuroscience foundation models), Doubleword (sovereign inference), Cosine (agentic coding AI), Cursive (foundation-model infrastructure), Odyssey (world models for physical AI) and Twig Bio (AI-enabled bioproducts). Each investee receives up to £20m equity, 1 million GPU-hours on the AI Research Resource, ten fast-tracked visas and procurement access. Prior coverage had listed only six firms with compute access; Kendall's announcement confirmed Callosum as the equity recipient and completed the cohort.
All seven firms sit at the infrastructure or foundation-model layer; no application-layer firm was selected, a pattern that echoes Brussels' sovereign cloud award and the Digital Europe Programme calls opened three days earlier that fund AI applications in health and safety rather than frontier model work. The policy bet is coherent with itself: DSIT is funding the plumbing on the view that frontier training remains compute-gated and that application-layer equity is premature while UK grid constraints keep training-scale compute offshore. The rhetoric attached to the fund, however, is sovereign AI rather than sovereign infrastructure, and the two are not the same thing.
Anthropic's London AI engineering packages run from roughly £225k to £340k, a salary band no investee can match under a £20m equity stake. That creates a structural hiring asymmetry: the Sovereign AI Unit cohort will compete for senior technical staff against US frontier labs whose London compensation sits above three times the funded range. DSIT's investees will recruit on research prestige, equity upside and mission rather than on pay. The implication is that the UK's sovereign AI project is being wired as a research and infrastructure base that trains British talent for a wider market, rather than as a vertical sovereign-model stack capable of retaining that talent domestically.
Whether DSIT expands the second cohort to include an application-layer firm, or doubles down on infrastructure bets, will be the clearest signal of how the Sovereign AI Unit reads its own remit. A compute allocation of 1 million GPU-hours per investee is adequate for research iteration; it is orders of magnitude short of what frontier pre-training runs now consume. Within that constraint DSIT is building a credible research-layer instrument. The sovereignty framing will need to narrow itself to match, or the funding layer will keep outrunning what the UK grid and salary market can actually support.
