When Matt Clifford handed his AI Opportunities Action Plan to the UK government in January 2025, the question was whether Whitehall would treat it as a glossy think-piece or a serious blueprint. Sixteen months on, the answer is in: Keir Starmer's government has accepted all 50 recommendations and is now executing against them, with the first AI Growth Zone announced in Culham, Oxfordshire, a National Data Library in the works, and a 20-fold expansion of UK sovereign AI compute targeted for 2030.
For a country that has spent the better part of a decade watching American hyperscalers and continental regulators define the terms of the AI economy, this is the most decisive industrial-policy posture Britain has adopted on technology since the early 2000s. And — crucially — it is the right one.
The plan in plain English
The Clifford report is unusually concrete for a UK strategy document. Its core argument is that AI is a general-purpose technology, that the UK already has structural advantages (top-tier research, the English-speaking talent market, a flexible regulatory tradition), and that the binding constraints are compute, data access, planning, and procurement — not novel cross-sector rules.
The headline commitments now formally adopted include:
- A 20x increase in public AI compute by 2030, anchored by an expanded AI Research Resource and a sovereign frontier compute capability.
- AI Growth Zones — planning-fast-tracked sites with grid priority for data centres and training clusters. The first, at Culham, sits next to the UK Atomic Energy Authority's site and has both legacy grid capacity and an obvious clean-energy story.
- A National Data Library to make high-value public datasets — health, geospatial, transport, education — usable by researchers and startups under structured access.
- A pro-innovation, regulator-led approach: existing regulators (Ofcom, ICO, MHRA, FCA, CMA) handle AI in their contexts, with the AI Safety Institute (now the AI Security Institute) focused on frontier model evaluation. Explicitly rejected: an EU-style horizontal AI Act.
Why the EU divergence is the headline
The most consequential line in the entire programme is not about gigawatts or GPUs. It is the decision not to copy the EU AI Act.
Brussels' horizontal framework — passed in 2024 and now phasing in — classifies AI systems by risk tier and layers obligations on general-purpose model providers regardless of downstream use. It is coherent on paper. In practice, it has produced a compliance economy that European startups privately concede is gating their ability to ship. The European Commission itself has signalled a willingness to revisit elements of the implementation timeline, and Mario Draghi's competitiveness report explicitly named AI regulatory burden as a drag on EU productivity.
The UK's bet is the opposite: regulate AI through the harm it actually causes in a specific domain, not through what a model might theoretically enable. A medical diagnostic model gets the MHRA. A credit-scoring model gets the FCA. A recommender system gets Ofcom under the Online Safety Act. There is no separate "AI lawyer" tax on every product that happens to use a transformer.
This is the same regulatory tradition that produced sectoral, outcomes-based financial conduct rules and a flexible data protection regime under the post-Brexit Data (Use and Access) Act. It is also the approach the US has converged on, post-executive-order-revocation. Britain sits comfortably between the two largest AI markets in the world — and that is a far better place to be than as a junior partner to either.
The compute story matters more than people think
The 20x compute target is the part of the plan that critics dismiss as industrial-policy theatre. They are wrong. Compute is the single most binding constraint on national AI capability. Without sovereign training capacity, the UK is permanently downstream of US export controls and corporate roadmap decisions made in Santa Clara and Redmond.
AI Growth Zones are the planning-system unlock that makes this possible. Britain's data centre build-out has been throttled less by capital than by grid connection queues and local planning friction — the new town of Culham is being explicitly designated to bypass both. Whether the model scales beyond one site is the real test.
Where the plan needs sharpening
None of this means the Clifford agenda is above critique. Three issues deserve serious follow-through:
- Civil liberties guardrails on data sharing. The EFF's recent submission on the UK's digital ID consultation is a useful reminder that ambitious data-sharing infrastructure needs hard limits on secondary use, function creep, and law-enforcement access. The National Data Library should be built with structured access, not centralised pooling.
- Online safety scope creep. A May 2026 joint letter from EFF and 18 other organisations warned UK policymakers against expanding intermediary liability further. The regulator-led approach only works if Ofcom resists the temptation to retrofit AI-specific duties into the Online Safety Act through guidance.
- Procurement discipline. Government is the UK's largest single AI customer. The plan's procurement reforms must avoid locking the public sector into single-vendor stacks before the market matures.
The bottom line
Britain has chosen growth, compute, and contextual regulation over a horizontal rulebook. It is a credible posture, it is differentiated from both Brussels and Washington, and — provided the civil-liberties and competition guardrails are honoured — it is the right one. Execution risk is now the only story that matters.