Power Evidence Is Now Part Of AI Sovereignty For UK Workloads

The Sovereign Cloud

2 September 2026 | By Ashley Marshall

Sovereign AI is no longer just a question of where the data sits. UK buyers now need evidence that the infrastructure can actually run when power, planning and capacity become constrained.

Sovereignty now includes power, not just location

For UK leaders, AI sovereignty used to sound like a hosting decision: keep prompts, documents and model outputs in a chosen jurisdiction, then ask the supplier for a data residency statement. That is no longer enough. The serious constraint is whether the infrastructure can actually run when demand, policy and grid capacity collide. GOV.UK's AI Growth Zones paper says onshore data centre capability is important for sensitive data, adoption benefits and resilience from global shocks, but it also names slow planning and delayed access to power as the biggest barriers. That changes the buying question. A sovereign AI proposal is weak if it can prove where data sits but cannot prove how power is secured, how capacity is reserved, and what happens when grid connection dates slip.

The practical implication is simple: treat electricity evidence as part of the AI assurance pack. Procurement teams should ask for the supplier's site-level power assumptions, grid connection status, backup architecture, demand response approach and workload portability plan before signing a long-term commitment. This does not mean every business needs to become an energy expert. It means the AI steering group should stop approving critical workloads on the basis of cloud region names alone. A UK-hosted inference service that is queued behind speculative connection demand is not necessarily more resilient than a well-designed multi-region service with tested exit routes. The sovereignty conversation has moved from geography to operational proof.

There is a useful counterargument: buyers cannot realistically audit national energy infrastructure. True, but they can audit supplier claims. If a vendor sells resilience, local control or protected capacity, the buyer can require evidence. That includes named facilities, realistic time-to-power assumptions, contracted renewable supply where relevant, failover tests and a clear statement of which workloads may still be serviced offshore. GOV.UK itself says the UK approach is pragmatic, not isolationist. That is the right framing for business buyers too.

The numbers have become too material for procurement to ignore

The scale of demand is no longer a background infrastructure story. In its AI Growth Zones policy, the government says the package could reduce time to power by up to 5 years, save a 500 MW data centre up to GBP80 million annually in electricity bills, unlock up to GBP100 billion of additional investment and create more than ten thousand jobs. Those figures matter because they show power is not an operational footnote. It is now central to whether UK AI capacity can be delivered at the price, pace and resilience level promised to customers.

Other public reporting points to the same pressure from a different angle. The Guardian reported that DSIT's compute roadmap forecast at least 6 GW of AI-capable data centre capacity by 2030, while another departmental projection for the wider commercial services sector appeared much lower. Carbon Brief, citing government and Ofgem material, noted that Great Britain had roughly 1.8 GW of data centre capacity in 2024 and that financially committed new data centre projects could require around 20 GW of electricity if built. These are not numbers a mid-market buyer needs to model in detail, but they should change the level of scepticism applied to supplier roadmaps.

What this means in practice is that procurement due diligence should include capacity realism. If a supplier's UK AI roadmap assumes a specific data centre, ask whether that site has a live grid connection agreement, whether it is inside an AI Growth Zone, whether power pricing support is assumed, and whether those assumptions are contractual or aspirational. If the supplier cannot answer, the risk is not just sustainability. It is service availability, unit cost, latency, contract flexibility and the ability to move workloads when capacity is rationed.

Energy evidence belongs beside data protection evidence

Most AI procurement packs now ask sensible questions about data protection, information security, retention, training use, subprocessors and audit logs. Energy and infrastructure evidence should sit beside that pack because it answers the next resilience question: can this service keep running on the terms being sold? The shift is especially important for sensitive UK workloads in healthcare, financial services, legal services, local government and regulated operations. Those teams may have good reasons to prefer UK-hosted or UK-controlled services, but the business case should show more than a map pin.

A practical energy evidence pack does not need to be long. It should include the facility location and operator, region and availability design, current and expected power capacity, grid connection status, backup power approach, cooling assumptions, energy source claims, contractual commitments, known constraints, and the supplier's plan if capacity or pricing changes. It should also include a workload placement statement. Some AI workloads can run offshore without unacceptable risk; others should stay in a UK-controlled environment; some can use smaller local models; and some need rapid failover more than local hosting. That classification is where procurement, technology, legal and operations teams need to work together.

The misconception is that this creates unnecessary friction. In reality, it reduces later friction. Buyers already ask vendors to evidence ISO 27001, SOC 2, DPIAs, penetration testing and subprocessors because vague trust claims are not enough. Power evidence is the same type of discipline. It is not an environmental gesture bolted onto the deal. It is a way to turn sovereignty from a slogan into an operating model that finance, risk and engineering can test.

AI Growth Zones change the supplier conversation

AI Growth Zones are designed to make strategically important compute projects easier to build by dealing with grid connections, planning barriers and regional energy economics. GOV.UK says strategic demand projects and AI Growth Zones will be prioritised for available energy network capacity, and that new mechanisms will reallocate freed capacity and reserve future capacity for strategically important projects. It also describes possible developer self-build routes for high-voltage lines and substations, plus a Connections Accelerator Service for important customers. For buyers, this means a supplier's UK infrastructure story may increasingly depend on whether it is inside the right policy and energy pathway.

That does not make AI Growth Zones a magic stamp of resilience. It gives buyers a sharper set of questions. Is the supplier using an AI Growth Zone directly, indirectly through a cloud provider, or only referencing the policy context? Is the capacity confirmed or expected? Is the service priced on the assumption of future electricity discounts? Does the contract protect the customer if those discounts do not appear? GOV.UK proposes targeted electricity cost reductions from April 2027 of up to GBP24/MWh in Scotland, GBP16/MWh in Cumbria and GBP14/MWh in the North East for eligible projects, subject to consultation and legislation. Those are material numbers, but they are not automatically available to every AI product sold as UK-hosted.

What this means in practice: add an infrastructure claims schedule to material AI contracts. It should state which claims are binding, which are roadmap expectations, and which are general market commentary. If the sales deck says the service benefits from UK sovereign compute, ask for the evidence line that makes that claim true. The best suppliers will welcome the question because it separates credible operational design from loose positioning.

The risk is not only carbon, it is continuity

The energy debate often gets reduced to sustainability, but the board risk is broader. Carbon Brief argued that emissions from new UK data centres could be far higher than early government estimates if only a small amount of the electricity is generated by gas, and reported uncertainty around the number of centres built, how clean their power is and when they come online. That matters for net zero plans, but it also matters for continuity planning. If demand grows faster than clean power, grid capacity or planning approvals, then compute may become more expensive, more constrained or more dependent on interim arrangements.

For a business running AI in production, those constraints can show up as higher inference costs, changed service tiers, slower provisioning, stricter rate limits, regional routing changes or supplier pressure to move workloads. None of these outcomes is theoretical. They are the ordinary symptoms of capacity-constrained technology markets. The difference with AI is that the dependency chain now includes power, cooling, chips, model providers, cloud regions, connectors, identity layers and governance controls. A resilience plan that stops at backup internet and cloud availability zones is too narrow.

The sensible counterargument is that large cloud providers are better placed than individual buyers to manage this complexity. They are, but the buyer still owns business continuity. Ask for the supplier's tested recovery objectives for AI-dependent workflows, not just platform uptime. Ask whether the service can degrade gracefully to a smaller model, a cached knowledge base, a human queue or a non-AI workflow. Ask how prompts, retrieval indexes, evaluations and audit logs move if the workload is relocated. Continuity is where sovereignty becomes practical.

Build a power evidence gate before approving critical AI workloads

The useful response is not to pause AI projects until the national infrastructure picture becomes perfect. That would be a different kind of risk. The useful response is to create a power evidence gate for workloads where AI availability, data location, latency or cost materially affects customers, staff or regulated obligations. The gate should be proportionate: lightweight for low-risk internal assistants, formal for customer-facing automation, regulated operations, decision support, high-volume inference and workflows that would be expensive to unwind.

A good gate has five parts. First, classify the workload by criticality and data sensitivity. Second, identify its hosting, model and retrieval dependencies. Third, ask suppliers for evidence behind local hosting, reserved capacity, energy claims and failover. Fourth, test portability before go-live by moving a representative workflow, not just exporting a document. Fifth, set trigger points for review, such as supplier model changes, region changes, price increases, capacity warnings, regulatory changes or data centre migration notices. This is not heavy governance. It is the same operational thinking UK businesses already apply to payments, payroll, customer data and production systems.

The businesses that handle this well will not be the ones with the longest AI policy document. They will be the ones that can answer three questions clearly: which AI workloads must be local, which can move, and what evidence proves the chosen supplier can run them reliably. If the answer is based on a sales claim rather than a tested assumption, the sovereignty strategy is unfinished.

Frequently Asked Questions

What is power evidence in AI procurement?

Power evidence is documentation that supports a supplier claim about infrastructure resilience, local hosting, grid capacity, backup power, energy sourcing and continuity. It helps buyers test whether an AI service can run reliably, not just where data is stored.

Does UK data residency prove AI sovereignty?

No. Data residency is one part of sovereignty, but it does not prove workload resilience, capacity availability, cost stability or operational control. Buyers should combine data protection evidence with infrastructure and exit evidence.

Should every AI supplier provide detailed energy data?

Not for every low-risk tool. The depth should match the workload. A customer-facing, regulated or high-volume AI workflow deserves stronger evidence than an experimental internal assistant.

What should we ask a UK-hosted AI vendor?

Ask where the service runs, which facilities or cloud regions support it, whether capacity is confirmed, what happens if the region is constrained, whether workloads can move, and which claims are contractual.

Are AI Growth Zones enough to remove infrastructure risk?

No. They may improve planning, grid access and investment conditions for eligible projects, but buyers still need to verify whether their supplier directly benefits and whether those benefits are binding.

How does this affect AI ROI?

Power and capacity affect unit cost, availability and scaling assumptions. If a business case assumes cheap, always-available inference without evidence, the ROI model may be too optimistic.

What is the best first step for a mid-market business?

Create a short evidence checklist for critical AI workloads. Add power, capacity and portability questions to the same procurement pack that already covers data protection and security.

Can smaller or local models reduce this risk?

Sometimes. Smaller models, local inference and workload routing can reduce dependency on constrained frontier capacity, but they still need testing against quality, security and operating requirements.