AI Growth Zone Evidence Should Shape UK Workload Placement

The Sovereign Cloud

15 September 2026 | By Ashley Marshall

UK AI infrastructure is now a power, water and resilience question. That should change how leaders decide where business AI workloads run.

Workload placement now needs infrastructure evidence

UK businesses are used to asking where data is stored. For AI systems, that question is now too narrow. The more useful question is whether the place where the workload runs can prove enough power, water, connectivity, resilience and operational capacity for the job it is being given. That matters because the UK is deliberately turning AI infrastructure into a national growth project, not treating it as ordinary server estate. In its policy paper on delivering AI Growth Zones, DSIT says onshore data centre capability is essential for protecting sensitive data, maximising adoption benefits and resilience from global shocks.

The practical shift for business leaders is simple. Sovereignty is no longer just a legal or procurement claim. It has to be evidenced at workload level. A board reporting assistant can tolerate a different placement decision from a customer decisioning workflow, a retrieval system over regulated records, or an agent that can update operational systems. Each one has different latency, availability, assurance and exit requirements. If the evidence stops at the phrase UK hosted, the risk has not been understood.

What this means in practice is that workload placement should have a short evidence pack before procurement signs off. That pack should name the provider region, the fallback region, the data classes involved, the model or inference tier used, the capacity assumptions, the operational owner and the exit route if performance, cost or compliance changes. This does not need to become a 60-page architecture paper. It does need to be specific enough for finance, legal, security and operations to make the same decision from the same facts.

The government has made power and water part of the AI equation

The AI Growth Zone criteria show how much the conversation has moved. The GOV.UK guidance for AI Growth Zone applications says sites must demonstrate access to at least 500MW of power capacity by 2030. It also says sites must demonstrate sufficient access to water to support at least 500MW of AI infrastructure, with written confirmation from the relevant local water supplier covering volumes required and available, infrastructure constraints and delivery timelines.

That is not a small footnote. It is a signal that compute capacity depends on real infrastructure, not just cloud account availability. A supplier can show a polished product demo and still be exposed to location risk if the underlying hosting path depends on scarce grid capacity, uncertain cooling arrangements or a future site that has not cleared planning and utility constraints. For a buyer, the lesson is not to become a civil engineer. The lesson is to ask whether the provider can explain its capacity assumptions and whether those assumptions are appropriate for the workload.

The counterargument is that most SMEs will never buy directly from an AI Growth Zone or negotiate power contracts with hyperscalers. That is true, but it misses the point. Smaller businesses still inherit these constraints through SaaS platforms, AI assistants, retrieval products and managed service providers. If a vendor changes region, queues a migration, raises inference prices or shifts a workload offshore because capacity is constrained, the business feels the impact. Evidence requests are a way to find those weak points before the workflow becomes business critical.

Capacity claims need to be separated from availability promises

Recent infrastructure reporting makes the buyer risk more concrete. New Civil Engineer, drawing on House of Commons Library analysis, reported that data centres currently consume 2.5% of the UK electricity supply and that consumption is expected to rise by a factor of four by 2030. The same report cited UK data centre demand estimates ranging from 3.3GW to 6.3GW, while warning that even the higher figure may still not meet demand for computing power. It also cited JLL research saying new 50MW data centres in London face an average grid connection lead time of seven years.

Those figures should change how leaders read supplier promises. There is a difference between a vendor saying a service is available today and proving that the placement model can support growth, incidents, seasonal demand and regulatory scrutiny. A pilot can run beautifully on borrowed capacity. A production workflow has to survive a budget cycle, vendor roadmap changes, model upgrades, customer peaks and outages. This is especially true when the AI system is not just generating text, but routing work, summarising evidence, preparing client outputs or acting across connected tools.

What this means in practice is that procurement questions should separate capacity from service-level language. Ask which region the workload normally runs in, which region takes over if that fails, whether failover changes data protection or latency assumptions, and whether the supplier has notified customers of material infrastructure constraints in the past. Ask what happens if a model tier is throttled, retired or moved. A good supplier should not need to reveal every commercial secret to answer these questions. They should be able to provide enough evidence for a proportionate placement decision.

The right placement decision depends on the job, not the slogan

The useful placement decision starts with the workload. Training a model, running low-risk drafting, searching internal documents and triggering operational actions are not the same thing. New Civil Engineer quoted Arup data centre leader Gareth Williams explaining that AI training workloads can be located further from users, including in places with economical green power and cooler climates, while inference services used by customers often need to sit closer to users. That distinction is exactly the kind of nuance business leaders need in their own AI estate.

For a UK firm, this creates a practical segmentation exercise. A public marketing ideation assistant may be acceptable on a standard global SaaS path with contractual safeguards. A support assistant reading customer records may require UK or EU processing, stronger logging, tighter retention and a tested human escalation path. A claims, finance or HR workflow may need extra approval gates and clear evidence that automated decisions are not being made without proper oversight. A browser agent touching a CRM or payment portal may need a dedicated control layer before placement is even the main question.

The common misconception is that sovereignty always means choosing the most local option. That can become expensive and brittle if applied without judgement. The better approach is risk-weighted placement. Put the strongest evidence where the business impact is highest. Accept lighter controls for low-risk tasks. Keep an exit path for anything that becomes critical. This is how sovereignty becomes an operating discipline rather than a procurement label.

A simple evidence pack can make procurement sharper

Most organisations do not need a new committee to handle this. They need a small evidence pack attached to each AI workflow that matters. The first page should identify the workload, owner, users, data classes, tools connected, supplier, model family, hosting region, fallback region, expected volume and business process affected. The second page should record the decision: approved placement, conditions, review date, known risks and the person who can stop or move the workload. That is enough to turn a vague discussion about AI sovereignty into a decision that can be audited later.

The AI Growth Zone application criteria provide a useful template for buyer thinking, even outside infrastructure projects. They ask for evidence of power availability, water availability and discharge, land, planning, connectivity, local impact and low-carbon energy solutions. A business buyer can translate that into vendor questions: what infrastructure assumptions sit behind the service, what constraints could affect delivery, what regions are used, what support is available during migration, what environmental or resilience claims are evidenced, and what changes would trigger customer notice.

This also makes finance conversations cleaner. If a workload needs a premium region, private deployment or specific resilience model, the cost should be justified by risk and business value. If it does not, the business can avoid over-engineering. Either way, the decision becomes explicit. That matters because hidden placement assumptions tend to surface at the worst time: during an outage, supplier price change, compliance review or urgent migration.

Leaders should ask for proof before workflows become dependent

The safest moment to ask for placement evidence is before the workflow becomes embedded. Once staff depend on a tool, customers expect the output and integrations have been wired into daily work, the business loses negotiating power. Supplier answers that sounded acceptable in a pilot can become awkward when legal asks where data is processed, security asks how failover works, or finance asks why inference costs have moved. The work is much easier if the evidence is gathered during selection and refreshed before scale.

There is also a strategic upside. The UK is trying to build domestic AI infrastructure quickly, with DSIT saying its AI Growth Zone package could reduce time to power by up to five years, save a 500MW data centre up to £80 million annually in electricity bills and unlock up to £100 billion of additional investment into the programme. Buyers who understand the evidence behind those claims will make better decisions about which workloads deserve UK placement, which can use allied regions, and which should stay portable because the market is still moving.

The practical next step is to add three questions to every meaningful AI procurement or rollout. Where will this workload run under normal conditions. Where will it run under failure, constraint or migration. What evidence proves that this placement is suitable for the data, users, volume and business process involved. If the answer is vague, do not block progress by default. Mark the workload as conditional, limit its scope and require better evidence before it becomes production critical.

Frequently Asked Questions

What is AI workload placement?

AI workload placement is the decision about where an AI task runs, which supplier or region handles it, what data moves there, and what happens if that location changes or fails.

Does every UK business need UK-hosted AI?

No. Low-risk drafting and research tasks may be suitable for standard SaaS terms. Sensitive, regulated or operational workflows need stronger evidence about location, controls, resilience and exit options.

Why do AI Growth Zones matter to buyers who are not building data centres?

They show the infrastructure evidence behind AI capacity. Even if you buy SaaS, your supplier depends on power, water, connectivity and regional capacity choices that can affect cost, resilience and compliance.

What evidence should a supplier provide?

Ask for normal and fallback processing regions, data handling controls, service-level assumptions, customer notice triggers, migration support, model retirement policy and any location-specific constraints.

Is data residency the same as sovereignty?

No. Data residency tells you where data is stored or processed. Sovereignty also covers control, resilience, portability, audit evidence, supplier dependency and the ability to move or stop a workload.

How often should placement evidence be reviewed?

Review it before production rollout, before major model or supplier changes, after incidents, and at least annually for workflows that touch sensitive data or core operations.

What should SMEs do first?

List the AI workflows that handle customer, staff, finance or regulated data. For each one, record where it runs, who owns it, what fallback exists and what evidence the supplier has provided.

Can a workload use offshore AI infrastructure safely?

Yes, if the data, contracts, controls and business impact fit that choice. The point is to make the decision deliberately and document why the placement is acceptable.