The exit option: portable workloads and a sovereign British cloud

British government and much of the country's critical infrastructure run on three American cloud providers. The arrangement works in normal conditions, which is the condition the case for it rests on.

The risk is widely misunderstood as a question of where data sits. Four parts of a cloud platform matter here: the control plane that orchestrates deployment, the data plane where workloads run, the licensing and authentication systems that decide who may use what, and the update mechanisms that push changes. For the three large providers, compute can run in a British data centre while the control plane, identity federation, licensing, and patching all sit under the parent company's jurisdiction. Hosting in Britain does not make a system sovereign. The off switch sits in the control plane, the licensing system, and the identity provider.^[1]

What that means in practice was demonstrated in 2025. After the United States sanctioned officials of the International Criminal Court under Executive Order 14203, and the Office of Foreign Assets Control issued designations against named individuals, the Court began moving some 1,800 workstations off Microsoft 365 to openDesk, the open-source suite built by the German state-owned ZenDiS.^[2] Individual sanctioned judges lost access to consumer services including Apple iCloud and PayPal while sitting in The Hague. Dutch reporting has it that Microsoft told the Court it would end services organisation-wide unless the prosecutor's access was cut. Microsoft denies suspending any service to the Court, and in Feb 2026 apologised to the House of Commons Business and Trade Committee and asked it to correct the record of evidence given on 10 Feb.^[3]

No contract prevented any of it. The same instrument, in the same form, is available against any organisation a United States administration designates, and against any platform serving one.

The current implicit British policy is to accept this and manage it with contracts, data classification, and a thin layer of sovereign infrastructure for the most sensitive workloads. That judgement is renewed with every framework contract, and nobody has ever taken it in the open.

Portable by design

The remedy has to reach all three switches. Public cloud deployments should be built on open-source orchestration so that workloads can move between providers, identity should be anchored somewhere the provider does not control, and licensing terms should survive a change of provider. Every major government deployment should hold one live node, sized to carry critical services on its own, on infrastructure outside the reach of a single foreign jurisdiction. For ordinary application workloads, an exit programme measured in months rather than years then becomes credible, and the live node carries the days while the migration takes the months. The Court's migration to openDesk shows the productivity and identity layer can be moved too.

That changes what a denial of service can achieve. Cutting off a customer that has made no arrangement to move is decisive. The Court is moving, but under duress, at emergency cost and on somebody else's timetable rather than its own. Cutting off a customer that has already made the arrangement is a transfer. The workload relocates on its own schedule, and the sanction achieves nothing. A shutoff at this level ends the relationship for critical workloads. The trust that deep platform dependence requires is rarely rebuilt afterwards. That is why the arrangement has to exist before the order, not after it.

A supplier also prices and behaves differently when the customer can leave.

Portability has limits. Open orchestration covers most ordinary government workloads well. It covers managed databases only partly, and only if you are willing to run them in-house. It barely reaches frontier artificial intelligence workloads, because proprietary accelerator stacks tie them to their providers. The leverage is therefore strong where most of government computing sits, and weak at the frontier until sovereign compute catches up.

Somewhere to move to

Portability only helps if there is somewhere to move to. Britain has no provider at hyperscaler scale. Civo and iomart are real businesses serving government frameworks, and neither is at that scale.^[4] The capability has to be built, through a joint venture drawing together the domestic operators that exist, a European partner with real scale, and a state investment vehicle, anchored by a founding-client commitment from government and the regulated sectors.

The venture should be built to a published standard rather than to its own definition of sovereignty. The European Commission's Cloud Sovereignty Framework sets five ascending assurance levels, SEAL-0 to SEAL-4, across eight objectives, and assesses control plane, key custody and identity anchoring as criteria within them.^[5] Buyers can procure against it today, and it gives them something specific to hold a supplier to.

Four terms that prevent a protected champion

European industrial policy has attempted versions of this several times, and the recurring failure is the protected champion that never has to get good. Four terms in the founding documents prevent it.

• The founding-client commitment carries a sunset. Government anchors the venture for five years and re-tenders at year five, so the venture must be competitive against the hyperscalers on the open market by then or lose its anchor revenue.
• The board is independent, with no ministerial appointments, and its mandate is explicitly time-limited.
• The chief executive is a founder-quality operator with equity at stake, recruited from the market, because seconding a senior official is how that starts.
• The anchor is a guaranteed baseline volume, and everything above it is contested from day one. Government keeps using the American providers for the bulk of workloads through the transition, and the venture wins the rest on merit.

Where this has been done

The nearest precedent is four months old. In Apr 2026 the European Commission awarded a €180m sovereign cloud framework, running six years, to four consortia, led by Post Telecom with OVHcloud and CleverCloud, by STACKIT, by Scaleway, and by Proximus NXT.^[6] Eligibility required SEAL-2, the data sovereignty level. Three of the four reached SEAL-3, digital resilience. An anchor customer brought sovereign capacity to market on a published standard. That proves the procurement can be run and that suppliers can clear the levels. It does not prove durability, and four months is not long enough to claim otherwise. It was done through procurement rather than through a new subsidy programme.

The supply side moved with it. Euro-Office reached general availability in Jun 2026 as an open-licensed alternative to the incumbent office suite, and sovereign disaster-recovery tooling built explicitly against the denial-of-service scenario is now sold commercially.^[7] For the ordinary workloads this paper is concerned with, what now limits European sovereign capacity is grid connection, not capability.

What it costs

A hyperscaler-class region with artificial intelligence capability is a build measured in the billions, and a ten-year programme of several regions with serious compute sits in the tens of billions.^[8] The money would come from several places: the Treasury, the national investment institutions, departmental budgets, and the savings a credible exit option wins back in renegotiated contracts.

Power is the harder constraint. The grid cannot connect new hundred-megawatt loads on demand in the south east, so the venture sites its capacity where headroom exists, in the north and the devolved nations, and the regional economies get the jobs the constraint creates.

Objections

The hyperscalers now sell sovereign clouds of their own. They do, and they help. Amazon's European Sovereign Cloud launched in Jan 2026 through a legally separate European subsidiary, Microsoft offers an on-premises control plane with disconnected operation, and Google sells sovereign controls with European key management.^[9] Each reduces data-residency exposure and routine regulatory exposure. None severs parent-entity jurisdiction. The intellectual property, the software, and the operational knowledge remain group assets, and licensing, updates, and identity federation still anchor to systems the parent controls. A sanctions instrument aimed at the parent is still effective. None of the three has been independently assessed against the Commission's assurance levels, which is itself the point: the framework exists so that sovereignty is demonstrated to a published standard rather than asserted by the seller.

This would never be done to Britain. The Court is a small international body with no state protector, and Britain is a treaty ally with intelligence-sharing running both ways. The argument does not need a cut-off to be probable. It needs the standing possibility to be priced, and the first ask here is cheap enough to clear at a low probability. It is also worth being clear about who the ally is: the administration in question sanctioned an institution that its own allies fund and sit on.

Multi-sourcing across the three American providers is cheaper and does most of the job. It is cheaper, and it delivers real resilience against commercial failure and outage. It delivers none against the risk this paper describes, because all three sit under the same jurisdiction and one instrument reaches all of them at once. A few days of denial of service while a court decides whether the order was lawful would be catastrophic for a department or a British company, and the remedy arriving later does not undo it. How much sovereignty matters when things go wrong is a question of how much risk ministers are willing to carry. They should settle the question deliberately, not let a position accumulate out of procurement defaults.

Europe already tried this and failed. The failures are real. They came from governance with no paying anchor customer, and protection with no competitive pressure. The Apr 2026 framework award is the counter-example, and the four rules above are addressed to exactly that record.

It is unaffordable in this fiscal climate. The headline number ignores the other side of the ledger. Renegotiated pricing under a credible exit option offsets part of the cost, adoption is deliberately partial, and the alternative is full-price dependence renewed indefinitely.

Government should not pick winners. The state's contribution here is demand certainty and patient capital, both of which it already gives the incumbents, unpriced, with every framework renewal. The venture is exposed to the market with a five-year clock running, and government keeps buying from the American providers throughout.

Where this lands

Portability by design is a condition that can be written into the next call-off, not a programme that needs a decade. Its cost is operational rather than capital, it is reversible, and it stands on its own whether or not the venture is ever built.

The venture is the larger decision, and it does not have to be taken all at once. A founding-client commitment sized to the Commission's framework rather than to a national champion would test whether British and European suppliers can meet a published standard, at a cost measured in tens of millions, and it would establish the assurance level government is willing to buy at. The tens of millions buys the assurance-level decision. The tens of billions is the programme that decision would later authorise. The frameworks sit with the Government Commercial Agency, which Crown Commercial Service became in Apr 2026, so the portability condition and the procurement that carries it are now in the same hands.