Choosing a UK Data Centre for Cloud Repatriation in 2026

If you are moving workloads out of public cloud and into a UK colocation facility, the decision that matters most is no longer which provider has the nicest tour route. It is whether the site can give you the power you need now, hold capacity for what you will need in three years, and let you leave on terms you can live with. The UK market makes that harder than it sounds: national data centre vacancy sat at 8 per cent in the first quarter of 2026, according to Savills, and 91 per cent of the country's live capacity is concentrated in and around London, where grid connections now queue for years. This guide sets out what is actually available, what the power situation really looks like, and the questions to put to any provider before you sign.

How real is UK cloud repatriation?

Cloud repatriation is real, measurable and mostly partial. The strongest evidence comes from an unexpected witness. In its evidence to the Competition and Markets Authority's cloud market investigation, in a hearing summary published on gov.uk in July 2024, AWS itself argued that the perception that customers who move to the cloud never return is not correct, and gave examples of customers moving workloads back to on-premises IT, for reasons including cost reallocation and greater ownership of resources, data and security. When the largest cloud provider tells a regulator that customers leave, the trend has passed the point of vendor folklore.

The survey evidence points the same way, though most of it is vendor-commissioned and should be read with that in mind. Broadcom's Private Cloud Outlook 2025, a survey of 1,800 senior IT leaders conducted by Illuminas and published on 29 May 2025, found 69 per cent of respondents considering repatriating workloads from public to private cloud, with one-third having already done so, and 53 per cent naming private cloud their top priority for new workloads over the next three years. Broadcom owns VMware and sells the private cloud stack the report recommends, so treat the framing as sold and the numbers as directional. IDC, drawing on its Cloud Pulse survey for the fourth quarter of 2023, reported in October 2024 that close to half of cloud buyers spent more on cloud than they expected in 2023, with 59 per cent anticipating similar overruns in 2024, and noted that production data and backup and disaster recovery are among the most commonly repatriated elements.

Two facts keep the honest boundaries on this trend. First, full exits are rare. The CMA's provisional findings of January 2025 found that fewer than one per cent of businesses migrate between cloud providers in a given year, and the dominant real pattern is selective — steady-state workloads with predictable resource needs come back, while burst capacity, SaaS and anything genuinely elastic stays put. Second, on-premises never went away. The Uptime Institute's Global Data Center Survey 2025, published in July 2025, found 45 per cent of IT workloads still reside in corporate facilities. Repatriation is less a reversal than a rebalancing, and in the UK it carries a sovereignty flavour: two-thirds of UK IT leaders have said they would switch cloud provider to regain sovereignty.

The regulatory ground has also shifted in the leaver's favour. On 31 March 2026 the CMA announced a package under which Microsoft and Amazon committed to actions on cloud egress fees and interoperability for UK customers, alongside a strategic market status investigation into Microsoft's business software ecosystem that began in May 2026. Under the commitments described in the CMA's decision documents, both firms extend free data transfer windows for switching customers to 180 days — Microsoft from 60 days, AWS from 90 — with Microsoft also covering partial exits, so a single workload can leave Azure without the whole estate following. The CMA Board reviews progress after six months, and the commitments are voluntary rather than binding, so verify the current position on the CMA's cloud services market investigation case page at gov.uk before building a business case on them. For a repatriation project, the practical effect is that the egress bill, long the tax on leaving, is now substantially reduced for UK customers who follow each provider's documented switching process. The Broadcom repricing of VMware licensing has pushed many of the same organisations to re-examine where those newly expensive workloads should physically run, which is how a licensing dispute became a data centre question.

What colocation capacity does the UK actually have?

The UK is Europe's largest data centre market, and it is overwhelmingly a London market. Savills reported in June 2026 that London accounts for 1,637MW of the UK's 1,803MW of live capacity — 91 per cent, up from 81 per cent a decade earlier — leaving just 166MW of live capacity outside the capital's orbit. CBRE's UK Real Estate Market Outlook 2026 draws the same map from the supply side: the London area (Docklands and Essex to the east, Slough and Berkshire to the west, Oxfordshire and Hertfordshire to the north) holds over 80 per cent of national supply, the Newport and Cardiff cluster in South Wales accounts for 9 per cent, and Manchester, the next largest cluster, just 2 per cent.

Availability is the tight variable. Savills puts UK vacancy at 8 per cent in Q1 2026, compressed from 27 per cent in 2016, with London tighter still at 7 per cent. A record 217MW was taken up across the UK in 2025 (212MW in 2024), and CBRE projects London vacancy will fall to 5.9 per cent by the end of 2026, a record low, as take-up exceeds new supply for a fifth consecutive year. Pre-letting is now normal — capacity is committed before buildings complete — so an enterprise buyer wanting a few hundred kilowatts with room to grow is competing for space in a market being absorbed by hyperscalers and GPU cloud operators. The trade body warnings about the UK's data centre infrastructure gap are about exactly this squeeze.

The operators below are the significant names a UK buyer will meet. Every entry was taken from the operator's own website on 6 August 2026. The list is alphabetical, the certifications are the operator's own claims, and inclusion is not a recommendation — the right provider depends on your requirements, not on scale.

One tier of the market is not for you: wholesale and build-to-suit developers such as Yondr, which announced a partnership with GLIL Infrastructure on a UK hyperscale campus in July 2026, build single-tenant capacity for hyperscalers rather than retail colocation. If a name on a market map will not quote for ten racks, this is usually why.

Why is power the constraint that decides everything?

Power, not space, now decides what gets built and where you can grow. Ofgem's call for input on demand connections reform, published on 13 February 2026, reported that Great Britain's demand connection queue grew from 41GW in November 2024 to 125GW by June 2025 — against a GB peak demand of roughly 45GW. Around 50GW of that queue is data centre projects. The regulator's own framing is that the queue contains a significant number of likely non-viable projects, that well-progressed projects cannot connect quickly enough, and that no mechanism exists to prioritise strategically important demand. Its proposed reforms — structured around curating the queue, planning capacity and connecting viable projects faster — drew 120 responses, summarised in June 2026, with consultations continuing. Reform will eventually improve queue discipline, but it does not create grid capacity, and in the meantime Savills reports developers in the most constrained West London locations planning for multi-year waits for firm power. That queue is why UK data centre vacancy has fallen to 8 per cent while London tightens its 91 per cent grip on capacity, and why new supply keeps landing pre-let.

For a buyer, the queue is not an abstraction. It determines whether the hall you sign for can ever give you more than the kilowatts on the order form. A site with grid headroom already secured can double your allocation with a purchase order; a site waiting on reinforcement cannot, whatever the sales deck implies. This is the single area where the power market's dysfunction becomes your contractual problem, and it is worth more due diligence than every certification on the wall.

What should you actually assess when selecting a provider?

Assess a colocation provider on seven things: power, connectivity, certifications, efficiency, physical risk, contract shape and exit terms. For each, there is a specific question that separates a marketing answer from a real one.

Power availability and growth headroom

The question is not "can you support 10kW per rack" but "what is your secured grid capacity, how much is contractually committed to existing customers, and what written commitment will you give me on additional capacity at years three and five?" Ask whether the site's connection agreement is firm or interruptible, whether headroom claims depend on a pending grid application (and if so, where it sits in that 125GW queue), and what happens to your expansion rights if the site fills. Densities are rising — the Uptime Institute's 2025 survey found most operators' highest-density racks still peak below 30kW, but the direction is up — so a hall that cannot feed and cool tomorrow's servers is a five-year lease on a dead end.

Connectivity and carrier neutrality

Repatriation almost never means leaving cloud entirely, so the facility must connect you back to it cheaply and privately. Ask three things: how many carriers are on-net (and whether the operator is genuinely carrier-neutral or steers you to an affiliate), what cloud on-ramps are available for private connectivity into AWS, Azure and Google Cloud, and what a cross-connect costs per month — because cross-connect fees are a classic place where a cheap rack becomes an expensive relationship. A hybrid estate lives or dies on this plumbing, and it is a big part of why London's carrier-dense sites command their premium. If your architecture spans multiple suppliers with no gaps between them, the data centre is the hub that makes that workable.

Certifications, and what they do not prove

Certifications are useful primarily as evidence that an operator submits to external audit, and each one proves less than the logo suggests. An Uptime Institute Tier certificate comes in flavours — design documents, constructed facility, operational sustainability — and "designed to Tier III standards" is a self-description, not a certification, so ask which certificate the site actually holds and for which building. ISO 27001 certifies an information security management system within a defined scope; ask for the scope statement, because a certificate covering head-office processes says little about the hall your racks sit in. SOC 2 comes as Type 1 (controls exist at a point in time) and Type 2 (they operated over a period); ask for the report, not the badge. None of these tells you whether the generators start on a wet Tuesday — ask instead for the site's tested runtime, refuelling contracts and the date of its last full black-building test.

PUE, and how it is gamed

Power usage effectiveness is total facility energy divided by IT energy, and the industry average has barely moved for six years — 1.54 in the Uptime Institute's 2025 survey. A modern site should beat that comfortably, but the number is easy to dress up: design PUE at full load is quoted as though it were operating reality, annualised figures hide summer peaks, partially filled halls flatter the ratio, and exclusions (offices, mixed-use loads) vary by operator. Ask for measured, annualised PUE for the specific building at its current load, and how it is metered. Efficiency matters to your bill because UK commercial electricity is the most expensive in Europe, as CBRE notes in its 2026 outlook, and power pass-through is most of what you will pay — the mechanics are covered in our guide to green cloud computing and net-zero data centres.

Physical risk, starting with flooding

Before any site visit, put the postcode into the Environment Agency's flood map for planning on gov.uk — it is free, takes two minutes, and shows whether the site sits in flood zone 2 or 3 for river and sea flooding, plus surface-water risk. A surprising amount of UK data centre stock sits on floodplain-adjacent industrial land, and an operator's mitigation story (raised halls, flood barriers, pumping) is worth hearing only after you know what it is mitigating. Ask separately about the site's power resilience topology (N+1 versus 2N), fuel storage, and proximity hazards — flight paths, fuel depots, protest-attracting neighbours — that insurers will ask you about later.

Contract shape and the price of small print

Colocation pricing comes per kW or per rack, and the honest comparison is always per kW including power pass-through, because a cheap rack with a low power allowance is not cheap. Ask how power is charged (metered versus allocated), what the pass-through formula is and how energy price movements reach you, what remote hands cost per hour and what response time that buys, and — again — the cross-connect tariff. Setup fees, minimum terms, indexation clauses and auto-renewal windows all belong on the comparison spreadsheet. The failure modes are the same ones we catalogued in red flags in UK IT support contracts — hidden costs, weak SLAs and exit clauses, and they translate directly to colo.

Exit terms, because that is why you are here

You are reading this because exit terms bit you once already, so write the exit before you sign the entry. Ask what notice is required, what decommissioning and "make good" charges apply, who owns the cross-connects at termination, how long you have to remove equipment, and what happens to pricing at renewal — the standard trap is a keen first term repricing sharply once your hardware is racked and moving it costs real money. Ask for renewal caps or index-linked increases in writing. A provider confident in its service will agree reasonable exit terms; one that resists has told you something useful.

Should you choose London or a regional site?

Choose London when your architecture genuinely needs it; choose regional when it does not, and most workloads do not. The physics is friendlier than intuition suggests: light in fibre covers roughly 100km of route for about one millisecond of round-trip latency, so Manchester to London costs in the order of 2-4 milliseconds and Edinburgh to London under 10 milliseconds on sensible routes. For synchronous database replication with a cloud region, ultra-low-latency trading or real-time media, that matters, and the Docklands-Slough axis keeps its purpose. For the workloads that actually lead repatriation lists — backup and disaster recovery, steady-state line-of-business applications, file services, private AI inference — it does not.

What the regions offer is availability and power. Savills forecasts growth in the North East, citing renewable infrastructure, grid connectivity and a more receptive planning environment; Stellium's Newcastle campus sits in a government-designated AI Growth Zone with subsea routes to North America and the Nordics. Wales has the largest single regional concentration: the Newport and Cardiff cluster is 9 per cent of UK supply on CBRE's figures, anchored by Vantage's 148MW campus, roughly 100 miles from London. Scotland has DataVita's Tier III-certified sites and three of Pulsant's Edinburgh facilities. Northern Ireland is a thinner market: Prescient DC describes itself as Northern Ireland's first commercial carrier-neutral data centre, sited near a transatlantic cable landing station with a claimed 56 millisecond round trip to North America, and many NI organisations weigh Belfast options against Dublin's much larger market in a neighbouring jurisdiction.

Regional operators generally price below the London market — power, land and business rates all run cheaper — though like-for-like public price lists are rare and any differential is quote-specific, so get both quotes rather than assuming. The regional cost that is real but underrated runs the other way: your people. Repatriated kit needs hands — installs, swaps, upgrades — and a site 40 minutes from your engineers beats a marginally cheaper one three hours away, remote hands fees notwithstanding. Put staff travel time into the cost model next to the rack price.

How do you decide, workload by workload?

Start from the workloads, not the destination. The candidates that suit colocation are the ones AWS described to the CMA and IDC's data confirms: steady-state, predictable workloads where cloud elasticity is paid for but unused — databases with flat utilisation, backup and disaster recovery estates, file and media storage, virtualisation platforms hit by licensing repricing, and regulated data with location preferences. The workloads that should stay are the ones that earn their cloud premium: genuinely elastic demand, managed services you could not staff, SaaS, and anything whose roadmap depends on provider-native services. If your footprint is small enough, the honest alternative to colo is a server room of your own — our small office server guide covers when that stops being sensible.

Then build the true cost comparison, which has four lines on the colo side against one cloud bill: the colocation contract (per kW, with power pass-through modelled at current UK prices), the hardware (bought or leased — and note that a separate Compare the Cloud guide covers the chip and memory shortage currently inflating server prices and stretching lead times, which belongs in any 2026 business case), the connectivity (circuits, cross-connects, cloud on-ramps), and the people (engineering time, travel, out-of-hours cover). Run it over five years, because colo economics improve with time in a way monthly cloud bills do not, and include the one-off migration cost — which the CMA's 180-day free egress windows have made materially smaller than it was.

Finally, take this checklist to every shortlisted provider: secured grid capacity and written growth headroom; firm or interruptible connection; carrier count and cross-connect tariff; cloud on-ramp options; certification scopes, not badges; measured annualised PUE for your building; flood zone status and resilience topology; per-kW pricing with the power pass-through formula; remote hands rates; and exit terms with renewal caps. A provider who answers all ten in writing is a partner. A provider who answers eight is a negotiation. Anything less is your next repatriation project.

Sources

Market figures in this piece are drawn from Savills' UK data centre research published in June 2026 and CBRE's UK Real Estate Market Outlook 2026; power and grid figures from Ofgem's Call for Input on Demand Connections Reform of 13 February 2026 and its June 2026 summary of responses; regulatory positions from the Competition and Markets Authority's cloud services market investigation, its provisional findings of January 2025, the AWS hearing summary published in July 2024 and the CMA's package of actions announced on 31 March 2026, all on gov.uk. Repatriation survey data comes from Broadcom's Private Cloud Outlook 2025 (conducted by Illuminas), IDC's Cloud Pulse research as reported by IDC in October 2024, and the Uptime Institute's Global Data Center Survey 2025. All operator locations, capacities and certifications were taken from the named operators' own websites on 6 August 2026.