On 12 June, Washington ordered Anthropic to suspend access to its most powerful models for everyone who is not a US person. The directive did not name countries. It named a category, and Britain falls inside it. A close ally, a Five Eyes partner and a self-styled third pole in the AI race learned that none of those memberships came with an exemption. The lesson is not that the models in question matter. It is that a frontier system any UK bank, hospital or defence contractor depends on can now be switched off by an American administrative decision, with no notice, no appeal and no British institution with standing to contest it.
Two documents, one uncomfortable mirror
Within days of the cut-off, two pieces of writing did more to frame the problem than any government statement. The first, Europe 2031, is a speculative scenario published on 11 June by the Brussels-based Arq Foundation. It pairs a fictional five-year slide into dependence with real data, and the headline number is the one that travelled: Europe holds roughly 5 percent of the world’s AI computing capacity against about 80 percent for the United States. The document went viral during the G7 talks in France, was read by members of the European Parliament, and, its authors say, surfaced in unofficial British-German discussions.
The second, a column for CEPR’s VoxEU by researchers working on European defence autonomy, refuses to treat the export order as an AI story at all. It calls it a sovereignty story, and argues the precedent is the point: unlike energy or semiconductors, AI dependency has no gradual adjustment mechanism. It can be switched off globally in hours, as 12 June demonstrated.
Read together, the two documents describe the same trap from opposite ends. One measures how little compute the European periphery owns. The other measures how little control that periphery has over the compute it does use. Britain is not mentioned prominently in either. It belongs in both.
Strategic Reality: Britain has spent two years positioning itself as the pragmatic alternative to a slow-regulating Europe. On the two metrics that decide who actually leads, capacity and control, it sits in the same exposed position as the continent it has been distancing itself from.
The numbers that frame the choice
| Measure | Where it stands | Source |
|---|---|---|
| Europe’s share of global AI compute | ~5% vs ~80% for the US | Europe 2031 |
| Largest US AI supercomputer vs Europe’s largest | 1,250 MW vs 83 MW | Europe 2031 |
| US firms’ share of the European cloud market | 65% (three companies) | CEPR / Kiel |
| EU cloud providers’ share of their home market | ~29% (2017) to ~15% (2022) | Europe 2031 |
| Cost to close Europe’s military AI gaps | €500bn over a decade | Kiel Institute, Sparta 2.0 |
| Time to reach peer capability | 10 to 15 years | CEPR |
These are European figures. Britain does not publish a directly comparable compute share, but nothing about its position moves it out of the 5 percent neighbourhood. The country that hosts the gap also lives inside it.
The remedy Britain has already chosen
The most interesting part of Europe 2031 is not the diagnosis but the prescription, because it breaks with the reflex that followed the export order. The instinct in Brussels was to pour money into Mistral, Europe’s only frontier-model developer, and build a sovereign champion. The report’s authors argue the opposite. Building a homegrown rival, they say, underestimates how hard the frontier is. The better move is to build far more computing capacity and use it to draw American operators in, gaining some bargaining power over their decisions. They call for capital at a scale Europe has not attempted in peacetime, plus deregulated zones for power and permitting.
A British reader should recognise that prescription, because it is more or less government policy already. The AI Growth Zones announced in the AI Opportunities Action Plan are, in all but name, the deregulated zones for power and permitting that the report recommends. The pitch to international operators, come and build here, we will clear the planning and connect the power, is the build-to-attract strategy in action. On the compute-gap framing alone, Britain looks ahead of the curve rather than behind it.
Critical Context: The UK did not wait for Europe 2031 to choose its strategy. The AI Growth Zones already bet on attracting compute rather than funding a national champion. The question the export order raises is whether that bet answers the right problem.
The catch the second document exposes
It does not, or at least not on its own. This is where the CEPR column lands its hardest point, and where Britain’s strategy meets its limit. Hosting compute is not the same as controlling it. A data centre on British soil, running American models under American licence terms, is physically present and legally foreign. The export directive proved the distinction is not theoretical. The infrastructure stays put. The access to what runs on it can be revoked from Washington in an afternoon.
The defence dimension makes this sharper still. In January, the US Department of Defense dropped the requirement that military AI remain under human control, replacing it with a standard of any lawful use free from usage-policy constraints. European institutions had no instrument to govern any of this, because in 2024 the EU deliberately excluded military and national security AI from its AI Act, at France’s insistence. Britain, outside the Act entirely, has even less standing. When Washington makes decisions about AI systems that operate on or affect British networks, there is no UK body positioned to contest or even delay them.
So the build-to-attract strategy delivers economic activity and jobs, which are real and worth having. What it does not deliver is sovereignty. A nation can be the landlord of world-class compute and still be a tenant of the models that give it value.
Hidden Cost: AI Growth Zones are measured in gigawatts and gross value added. The thing they cannot put on a balance sheet is control. Attracting operators raises Britain’s economic stake in compute while leaving the off-switch in another government’s hands.
Who this lands on
| Stakeholder | What the compute-and-control gap means in practice |
|---|---|
| UK financial services | Critical workflows built on frontier models exposed to revocation with no contractual remedy against a sovereign act |
| Defence and intelligence | Run on infrastructure they cannot independently audit; no British standing over US military AI doctrine |
| NHS and public sector | Procurement decisions now carry a continuity risk that conventional vendor due diligence does not capture |
| AI startups | Benefit from cheaper hosted compute, inherit the dependence that comes with it |
| Government | Holds market access as leverage, but a single market of 68 million, not the EU’s 450 million |
What Britain should do whilst the gap stays open
The honest position is the one the CEPR authors take about Europe: for the next decade, the gap cannot be closed by political means alone. That does not make the next decade passive. It makes the near-term moves about buying time and reducing exposure whilst the slower, more expensive work of capacity proceeds.
For government and policy. Treat the export directive as a recorded change in the operating environment, not a one-off. The EU’s first recommended step is to make a formal political assessment of US military AI changes; Britain’s equivalent is to name, in policy, that frontier-model access is now a revocable dependency and to build that assumption into procurement, defence planning and critical-infrastructure resilience. Market access is the only real leverage a country this size holds over American providers, and it is worth far more deployed inside a governance negotiation than spent quietly on unrelated priorities.
For boards and executives. Map where frontier-model access sits on your critical path, then ask the question the export order forces: what happens to this process if access is withdrawn by an act of state, not a commercial dispute. A second-source model, an open-weights fallback or a degraded-mode plan is no longer a procurement nicety. It is continuity planning for a risk that has already materialised once.
Take Action: Run a dependency audit on every business-critical workflow that calls a frontier model. Classify each by what breaks if access is cut and how quickly you could switch. The exercise is cheap. Discovering the answer during the next directive is not.
For investors and builders. The build-to-attract strategy is genuinely good for the UK economy and worth backing. Just price the control gap honestly. A British data-centre play hosting American models is an infrastructure bet, not a sovereignty bet, and the two carry different risks. The sovereignty premium accrues to whoever owns weights, tooling or an open-model stack that no foreign directive can switch off.
Four challenges that do not show up in the headline
The leverage Britain holds is smaller than the EU’s, and it is acting alone. The CEPR remedy leans on the European single market as the bargaining chip and on collective instruments, ReArm Europe’s €800bn and a forthcoming Cloud and AI Development Act, to fund the alternative. Britain is outside all of it. Its market is a fraction of the size and it has no equivalent pooled-financing vehicle to reach for. Sovereignty by going it alone is the most expensive route there is.
The remedy and the problem pull in opposite directions. Attracting more American operators deepens the economic relationship that the export order just weaponised. Each new hyperscale facility raises both Britain’s compute capacity and its stake in keeping Washington friendly. Capacity and independence are not the same axis, and progress on one can quietly erode the other.
Energy is the real ceiling. Deregulated zones clear planning, but the constraint underneath is power. Britain carries some of the highest industrial electricity costs in the G7, and compute at frontier scale is, in the end, a bet on cheap, abundant, reliable energy. The compute gap is downstream of an energy gap that no permitting reform alone resolves.
The window is shorter than the build time. Reaching peer capability is a 10-to-15-year project on the European estimate, and there is no reason Britain’s is shorter. The leverage of being a friendly, well-regulated host is real now and erodes as American operators put down roots they do not need to defend. The advantage has to be spent on building leverage, not banked as if it were permanent.
The strategic takeaway
Britain’s AI strategy answers half the problem well. The build-to-attract bet, the Growth Zones, the lighter rulebook, is a credible response to the compute gap, and arguably ahead of where Brussels has got to. What 12 June exposed is the other half. Capacity without control is hospitality, not sovereignty, and a guest who can be evicted by a third party is not the one setting the terms.
The three things that decide whether Britain narrows the gap rather than just hosting it:
- Owning something that cannot be switched off. Weights, open-model infrastructure or sovereign tooling. The economic upside of hosting is real; the strategic upside lives only in what Britain controls outright.
- Spending market access as leverage, not revenue. A 68-million-person market is a modest chip. It is worth more inside a negotiation over notification and continuity than collected quietly as data-centre business rates.
- Closing the energy gap underneath the compute gap. Cheap, reliable power is the precondition for everything else. Permitting reform without an energy answer just relocates the bottleneck.
For UK leaders, the immediate checklist is shorter than the strategy:
- Audit frontier-model dependence across critical workflows and rank by what breaks if access is cut.
- Build a second-source or degraded-mode fallback for anything that cannot tolerate a sovereign-act outage.
- Treat AI continuity as a board-level resilience risk, not an IT procurement detail.
We have argued before that the UK’s regulatory advantage over Europe is real but unearned, and that Trump’s Anthropic ban exposed Britain’s dependence. This is the same argument from the infrastructure end. The rulebook decides how AI is used. The compute and the control decide whether Britain gets to use it at all.
Sources and attribution
This analysis draws on two sources. “Europe 2031” scenario pegs Europe’s AI compute at 5% of the world, published by Implicator.ai on 19 June 2026, reports the Arq Foundation scenario, its compute figures and its build-to-attract recommendation. Europe’s ungoverned space: Military AI and the autonomy that cannot be bought, published by CEPR / VoxEU on 19 June 2026, supplies the sovereignty framing, the 12 June export directive, the Kiel Institute Sparta 2.0 cost estimates and the European cloud-dependence figures.
Resultsense analysis applies these European frames to the United Kingdom’s distinct position, outside both the US tent and the EU’s collective instruments. Figures are reproduced from the cited sources; the strategic interpretation is our own.