27MAY

UK names first Sovereign AI investees

4 min read

15:19UTC

DSIT took a direct equity stake in Cambridge chip-interoperability startup Callosum and awarded 500,000 GPU hours each on Isambard-AI to six British firms on 16 April. The fund moved from launch to placements in seven days.

← Back to Brussels slips sovereignty law a third time Jump to analysis ↓

TechnologyDeveloping

Key takeaway

Cosine, air-gapped for nuclear, now has a cornerstone claim on the British sovereign AI stack.

The Department for Science, Innovation and Technology (DSIT), Britain's tech ministry, named the first investees of its Sovereign AI Fund on 16 April 2026: a direct equity stake in the Cambridge chip-interoperability startup Callosum, founded by Danyal Akarca and Jascha Achterberg, plus 500,000 GPU hours each on the Bristol-based Isambard-AI supercomputer for Cosine, Prima Mente, Cursive, Doubleword, Twig Bio and Odyssey ¹. Each firm also receives same-day super-priority visa decisions, ten free R&D visas, legal-fee cover for UK incorporation and access to government procurement pipelines ¹. Roughly thirty further companies are in discussions for AIRR, Britain's public AI compute programme ¹.

James Wise, the Balderton Capital partner chairing the fund, said the programme was moving at a pace "Craig David would blush at" ². Josephine Kant, the fund's Head of Ventures, runs deal flow ². Technology Secretary Liz Kendall told an audience on 16 April that AI was "non-negotiable for our national security" and that the investment committee would operate free from political interference ³. Chancellor Rachel Reeves listed a "thriving domestic AI sector" among her three economic priorities ¹.

Founded in 2022 by Alistair Pullen and Yang Li, Cosine took the cornerstone role and has outperformed OpenAI, Anthropic, Mistral AI and DeepSeek on independent coding benchmarks for two consecutive years ⁴. Its system is air-gapped and on-premise, built for classified infrastructure, covering 38 programming languages including Fortran and COBOL for nuclear, defence and financial services clients ⁴. The fund also took an option to follow on in Cosine's next funding round ⁴. Alongside Odyssey, which builds world models for autonomous defence systems, the portfolio tilts firmly toward regulated and defence applications.

Mistral's $830m debt raise in March and the French Ministry of Defence framework requiring French-only infrastructure remain larger single commitments than the first UK tranche. Paris has already bought a national champion; DSIT has bought a portfolio, compute access and visa tooling on a seven-day clock instead.

Deep Analysis

In plain English

Training an artificial intelligence system requires enormous amounts of computing power, running for weeks or months. That computing power lives in specialist chips called GPUs (graphics processing units), which are expensive to buy and costly to rent. Britain's Sovereign AI Fund is giving British AI companies free access to a powerful government-owned supercomputer called Isambard-AI, based in Bristol, plus equity investment in some cases. The aim is to help British AI companies reach a competitive stage without having to rely entirely on American cloud providers or take money from American investors who might eventually move the technology out of the country. The seven companies named on 16 April 2026 are the first beneficiaries. Most are focused on areas where the UK government itself wants domestic AI: defence systems, medical research, and software for critical national infrastructure.

Deep Analysis

Root Causes

GPU compute is the rate-limiting input for training and fine-tuning frontier AI models. A single H100 GPU costs approximately $30,000 to purchase or $2 to $3 per hour to rent from a US hyperscaler. A startup needing 500,000 GPU hours of training compute faces a $1-1.5m cost barrier before writing a single line of product code. Access to Isambard-AI's compute at zero marginal cost removes that specific barrier without requiring equity dilution or US cloud dependency.

The deeper structural cause is that European AI investment has historically concentrated at Series B and above, after US or Asian co-investors are already on the cap table with governance rights.

The Sovereign AI Fund's earliest-stage intervention, at pre-Series A for most investees, is designed to set UK governance conditions before US capital can impose its own. Callosum's direct equity position is the clearest expression of this logic: the government wants a seat at the table before any private lead investor.

What could happen next?

Consequence
The ~30 additional companies in AIRR discussions suggest a second cohort announcement before July 2026, which will test whether DSIT can sustain portfolio-bet velocity or reverts to bureaucratic procurement timelines.
Short term · 0.75
Risk
Without anti-acquisition conditions on equity investees, US AI labs can offer 10 to 20x early-stage valuations to capture the portfolio companies once they demonstrate product-market fit on UK sovereign infrastructure.
Medium term · 0.68
Opportunity
If Cosine's air-gapped, classified-infrastructure positioning attracts NATO or Five Eyes procurement contracts, the UK gains a sovereign AI income stream that no EU programme can access due to non-membership.
Long term · 0.6

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: Israel's Tamir programme of the 1990s, which provided government compute contracts and procurement guarantees to early defence-technology firms including Check Point and Amdocs, seeded a commercial technology sector that now accounts for roughly 10% of Israeli GDP.

The mechanism was identical: institutional anchor contracts plus visa facilitation creating a talent-dense environment where founders retained equity but had a government customer from day one. Cosine's air-gapped, nuclear-and-defence positioning mirrors the Tamir model almost exactly.

Counter-parallel: Canada's strategic investment in Cohere through federal procurement and the National Research Council produced Europe's best-capitalised language model company, but Cohere is now in merger talks that would see it absorbed into a German sovereign AI structure under conditions Berlin, not Ottawa, is setting.

Government anchor investment that does not include governance conditions or anti-acquisition clauses creates assets that can be captured by better-capitalised foreign actors at the next funding round.

The DSIT fund allocates £500m across a portfolio of investments ranging from £1m to £20m+ in equity per company, plus compute grants valued at approximately £1.5m per firm (500,000 GPU hours at a realistic Isambard-AI unit rate). The first cohort of seven firms represents a small fraction of the total fund deployment.

For comparison, Mistral AI's $830m debt raise purchased 13,800 Nvidia GB300 GPUs to build a 44MW data centre. The UK's compute allocation to six firms totals 3,000,000 GPU hours, roughly equivalent to 18 months of uninterrupted compute on a modest cluster.

France's approach builds permanent infrastructure; Britain's compute grants create access windows that must be renewed. The British government is moving at VC speed, where capital recycling and follow-on rights matter more than anchor infrastructure.

Consensus view: Chatham House's digital sovereignty programme and TechUK both argue the portfolio-bet model is the correct UK approach at this stage. A single national champion strategy (the French path with Mistral) requires an existing company of sufficient scale; Britain's AI sector has no equivalent to Mistral AI at the time of launch.

Spreading 500,000 GPU-hour grants across six companies and taking a single direct equity position in Callosum acknowledges that the UK does not yet know which firm will be its anchor. The VC committee structure, with James Wise from Balderton Capital chairing, is designed precisely to retain optionality and follow-on rights rather than commit to a winner before the market has selected one.

Counter-view: Institut Montaigne's digital policy programme (Paris) argues the portfolio-bet model concedes the key sovereignty question before it is answered. GPU hours on a government supercomputer preserve no IP, create no compelled-data relationships, and build no moat; they are a subsidy for compute costs that any well-funded US AI lab can match in a single procurement cycle.

The French military's January 2026 framework with Mistral , with its infrastructure-exclusivity condition, creates a hard sovereignty floor. The UK's compute grants create a temporary speed boost for firms that remain free to raise from US investors and incorporate abroad.

Key tension: Whether public compute access bundled with procurement guarantees and visa tooling can create a sovereign AI sector with durable UK ownership, or whether it primes companies for acquisition by US labs that can outbid any government follow-on cheque.

First Reported In

Update #2 · Brussels buys, Britain backs, Google unlocks

UK DSIT· 19 Apr 2026

Read original →

Causes and effects

Caused by

UK launches £500m Sovereign AI Unit

DSIT's first Sovereign AI investees are the direct deployment phase of the £500m UK Sovereign AI Unit launched 16 April and logged as ID:2338.

Occurred 16 Apr 2026

Read story →

This Event

UK names first Sovereign AI investees

Britain's first sovereign AI portfolio tilts firmly toward defence and regulated applications, trading scale for speed and an equity claim on a Cambridge founder team.

Led to

Kendall names seven infrastructure bets for £500m Sovereign AI Unit

The investee naming in ID:2618 was advanced further by the Liz Kendall formal launch event, which confirmed all seven firms with full package details including 1 million GPU-hours rather than the 500,000 cited in prior coverage.

Occurred 16 Apr 2026

Read story →

UK picks champions as EU shares the bet

The first Sovereign AI Fund cohort (equity in Callosum, compute for six firms) is the predecessor programme the second-wave competition extends.

Occurred 5 Jun 2026

Read story →

Different Perspectives

United States (Google/Alphabet)

Alphabet lost its final Android appeal on 2 July with no further court to hear it, a result its Computer and Communications Industry Association allies frame as precedent, not deterrence, since the €4.1bn fine changed nothing about Google's Play Store terms across eight years of litigation.

UK Department for Science, Innovation and Technology

DSIT opened its £96m second Sovereign AI wave on 3 July, switching from April's equity stakes to fixed-price contracts because Britain has no domestic hyperscaler or Bpifrance-style lender to fund capacity another way. It is betting on buying outcomes it controls alone rather than joining an EU-wide framework.

German federal government

Berlin backed both German deliverables this week, Infineon's fab and Aleph Alpha's merger, but is finding one far harder to close than the other. It wants enforceable protective rights inside Cohere's cap table before the merger closes, a legal instrument the Bundeskartellamt has no filing to review yet.

European Commission

The Commission banked a clean CJEU win on the eight-year Android case on 2 July, removing Google's last comparator argument before President von der Leyen rules on the far larger DMA self-preferencing fine due 27 July. Brussels treats Infineon's early Dresden delivery as proof the Chips Act mechanism works, at the node Europe already led.

Bruegel (EU industry sceptics)

Bruegel economist Mario Mariniello argued the EU sovereignty package mimics US and Chinese strategy while EU cloud providers hold roughly 15% of their home market; using nationality as a proxy for security without fixing the underlying capital and energy gaps that drive the dependency creates €86bn of migration cost without the security benefit it is sold as delivering.

France

France published a joint sovereignty definition with Germany at VivaTech and mobilised €13bn under Tibi Phase 3, placing SAP's partnership with Mistral as the working proof that a German enterprise-software giant running a French sovereign model inside public administration is what digital sovereignty looks like in practice.