The Guardian reports George Osborne linking domestic data centres to AI sovereignty. Guardian
For a public-service buyer, that claim raises a practical question: what decisions become possible when the computers are nearby? Hosting AI in the UK can support national objectives, but the useful measure of control is what an institution can actually do with the service. A local address does not tell a buyer whether it can reserve capacity, choose the model or move the work elsewhere.
The UK’s own compute policy offers a more specific starting point. It links sovereignty to the ability to act independently where needed while continuing to work with international suppliers. That leaves room for several kinds of infrastructure, and requires a clearer account of what each provides. UK Compute Roadmap
The big change
- What changed: The UK’s compute roadmap sets an ambition to direct resources toward national priorities and support independent action where needed.
- Why it matters: A service’s location leaves important choices about access, administration and replacement unresolved. Those choices affect what public bodies and businesses can deliver.
- What to watch: Whether proposals specify who can allocate capacity, authorize changes and move workloads, with evidence that those rights can be exercised.
The government’s ambition goes beyond a building
The Compute Roadmap was presented in July 2025; its current page was updated on April 23, 2026. It commits up to £2 billion through 2030 for public compute, including an intended twentyfold expansion of the AI Research Resource. It envisages both dedicated supercomputers and cloud capacity, with public access and strategic control in partnerships. These are policy commitments and delivery plans, not evidence that every promised resource is available today. Roadmap
The distinction matters because construction and access solve different problems. A company can add processing capacity to a country while selling it under terms that leave a particular research group unable to use it. A smaller resource with a dependable allocation could be more useful to that group than a much larger facility for which it has no budget or entitlement.
For national policy, the question is therefore which work gains a feasible route to completion. That might mean access for an experiment that commercial customers would otherwise displace, or a service whose operator must maintain agreed capacity for public use. Such arrangements would need to identify the beneficiary and the conditions of access. Announced spending alone cannot answer those questions.
Government could retain particular decisions while contracting out day-to-day operation. The buyer needs to know which decisions it retains, which it delegates and how that division can change over the life of the arrangement.
A local site can have several sources of authority
A research paper revised in May 2026 examined 775 non-US data-centre projects, using information collected through the third quarter of 2024. It found substantial participation by foreign operators, making operator nationality relevant to possible jurisdictional claims. The sample combines planned and existing projects, including clouds and supercomputers; it is no current census. Operators were unidentified for many entries. Its investment-based estimates are rough proxies for computing capacity, not direct measurements. Research paper
The authors also distinguish a government’s possible jurisdictional claim from its ability to exercise control. Their findings cannot establish that a particular UK service will be interrupted or that any foreign operator is unsuitable. They show why examining the operator adds information that geography alone misses.
Ownership is a useful starting question. Buyers also need to identify who runs the equipment and administers the cloud service, along with who supplies the model and can change its terms of use. These roles need not belong to the same organization. A domestic brand on the customer contract may describe only one layer of the arrangement.
The National Cyber Security Centre’s asset-protection guidance makes the inquiry concrete. It asks buyers to understand where data is stored, processed and managed, which jurisdictions may apply, and the provider’s rights of access. It includes derived material such as logs and models. The guidance also distinguishes supplier assertions from independently supported assurance. NCSC Principle 2
Buyers need to examine the exact service being offered. A statement about the location of its main processing system does not explain every administrative or supporting service. That requires an account of the relevant operations and evidence appropriate to their sensitivity.
Two UK services, different choices
Consider a hypothetical group of councils buying an AI search service for their public archives. Both offers process the archive in UK data centres. Both meet the same agreed test for finding documents and answering with accurate references. Neither system makes decisions about residents.
In the first offer, the vendor provides a managed application. It selects the model and schedules model changes. The councils buy requests to the service and can export their original documents, but they cannot take the model or the application to another operator. The vendor handles maintenance, which suits a small council team that wants a working search service without employing specialists to run it.
In the second offer, the councils obtain a reserved allocation of machines and a model licence that permits operation with another provider. They control when a tested model version changes. Their contract supplies the configuration and documentation another qualified operator would need to run the application. They also pay for staff or a contractor to maintain it.
Both arrangements provide local processing. The second gives the councils more operational choices, at the cost of additional responsibility. The first might still be the better purchase if its quality, price and contractual protections fit the task. Calling either offer sovereign without describing those differences would conceal the decision the buyer actually faces.
Now suppose the councils need to add a newly digitized archive during a busy period. In the first arrangement, additional use depends on the service’s terms and available capacity. In the second, the councils can reprioritize work within their reserved allocation. That is a meaningful form of control, although it does not create more machines or guarantee that the job will finish quickly.
Suppose instead that they want to replace the operator. Possessing documents is only part of the move. The receiving service needs a compatible application, usable model rights and enough capacity; it must pass the councils’ quality checks. The hypothetical second offer makes that move more feasible, but its portability claim still needs a demonstration. The ability to leave is worth more when the buyer has verified the work involved.
Independence carries work of its own
NCSC’s supply-chain principle asks buyers to understand third-party access and how the provider manages its suppliers’ security. It warns that a managed service provider and its dependencies also form part of the customer’s supply chain. For sensitive applications, assessment needs to consider the underlying services. NCSC Principle 8
Our council example illustrates why a contract cannot confer every kind of independence at once. Control over a model version leaves somebody responsible for evaluating updates. The right to change operators leaves somebody responsible for maintaining the export and deployment process. If nobody funds that work, an impressive set of rights may be difficult to exercise.
Building this competence could improve a buyer’s negotiating position. Understanding the application makes it easier to specify changes or commission alternatives when a supplier’s proposal does not meet its needs. Shared expertise across institutions could make those options affordable for organizations too small to maintain them separately. Those are possible benefits, not measured outcomes from the two invented offers.
The pessimistic case is that an institution pays for dedicated infrastructure and still remains dependent on a service it cannot replace. It might retain the hardware while lacking the model rights or specialist knowledge needed to use it differently. Avoiding that outcome requires deciding which dependence matters for the task, rather than demanding a vague promise of total independence.
Ask what the investment lets people do
Residents assessing a proposed facility can reasonably ask which public benefit requires that facility and how access to that benefit will be secured. A claim about national capacity should identify who can use it. A claim about control should identify the decisions someone gains the authority and practical means to make.
For buyers, a useful comparison would place the two offers against one real task and one plausible change: adding demand, changing a model or replacing an operator. Ask each provider to demonstrate the relevant action and specify who bears the work and cost. That produces evidence more useful than a broad sovereignty label.
