Britain’s Data Centre Vacancy Rate Is Measuring the Wrong Thing
The headline numbers on British data centres tell a story of scarcity meeting a building boom. Vacancy across the UK has fallen from 27% in 2016 to 8% in the first quarter of 2026, according to Savills, which also recorded a record 217MW taken up across the country during 2025. CBRE forecasts London vacancy falling to 5.9% by the end of this year, a record low, with take-up exceeding new supply for the fifth consecutive year, and new supply delivered to London across 2025 and 2026 reaching 373MW, more than double the previous two-year period.
Read as a market report, that is a straightforward supply squeeze. Read as a guide to whether a British company can find somewhere to run an AI cluster, it is close to useless.
Start with how slippery the headline figure is. For London in the first quarter of 2026, Savills put vacancy at 7%. CBRE, in its global trends report covering the same quarter, put it at 8.6%, up from 8% a year earlier. Both are serious firms using their own definitions of the market’s boundaries and of what counts as available. Neither is wrong. The gap is a useful reminder that a single national percentage is a summary of many different things, and summaries lose the detail that matters.
The detail that matters here is that the vacancy rate counts megawatts without asking what those megawatts can support.
Two numbers that do not belong in the same sentence
The Uptime Institute’s 2025 global survey reports that the average of the typical rack densities operators describe reaches almost 9kW. Adoption in the 10kW to 30kW band is growing, but few facilities exceed 30kW and extreme densities remain rare.
Now the hardware. NVIDIA’s NVL72, which packages 72 Blackwell GPUs into a single rack-scale system, is rated at 132kW in an otherwise ordinary 19-inch rack, on Uptime’s own reading of the specification. Vendor figures for this class of system vary with configuration and generation: Supermicro’s datasheet for the GB300 version of the same rack gives a range of 132kW to 140kW. Take the lower end, and it is still around fifteen times the density of the typical installed rack, in the same physical footprint.
The consequence, in Uptime’s own framing, is that a few small rows of these systems can consume as much power as a sizeable data hall filled with hundreds of conventional racks. A megawatt of AI-capable capacity and a megawatt of general colocation capacity are not the same product, and cannot be substituted for one another, but the vacancy statistics report them as though they can.
The scarcity of facilities that can take this load is documented. Uptime’s 2024 survey found that only around 1% of operators had any racks exceeding 100kW, and most of those were running traditional high-performance computing rather than AI. CBRE now names the problem directly in its 2026 global outlook, noting that the AI-readiness of existing stock is a growing concern because many legacy facilities were never designed for these workloads.
Why the obvious fix is harder than it looks
The intuitive response is that existing halls should simply be upgraded. There is a strong commercial reason to try: an existing site already holds a grid connection, and grid connections are the scarcest asset in the British market. Retrofit is genuinely the fastest route to AI-capable capacity where it works.
It does not always work, and the reasons are physical rather than financial. Uptime puts the floor loading of an NVL72 at 1.4 tonnes per rack. Data Center Knowledge, reporting on the limits operators are hitting, puts high-density AI racks at up to 1.5 tonnes and identifies floor loading in multi-storey retrofits as a binding constraint, alongside power distribution systems originally sized for lower densities.
The obstacles compound. Raised floors in older halls are frequently congested with decades of accumulated cabling and air handling, leaving nowhere to route coolant. Legacy ceiling grids are often not rated for the static load of filled manifolds plus the additional cabling a GPU cluster brings. Introducing liquid pathways into a live hall carries leak risk that has to be engineered out rather than accepted. An opinion piece published by Data Centre Dynamics, written by a supplier of the very equipment retrofits depend on and so arguing against its own commercial interest, concedes that operators have in some cases found it more cost-effective to demolish and rebuild than to work around these constraints.
The 132kW figure also understates the change in cooling architecture involved. At full load on an NVL72, cold plates handle around 100kW of thermal power while the air exhaust still exceeds 25kW. This is not a facility choosing between air and liquid. It is a facility running both, at densities where the air side alone would have been a respectable whole-rack load a few years ago.
That inertia shows up in the efficiency data. Uptime has now recorded six consecutive years of essentially flat average PUE, and attributes the lack of movement to legacy infrastructure and to region-specific climatic limits on efficient cooling. The second of those nobody can do anything about. The first is the installed base failing to modernise, and it has held the industry average still for longer than the current AI cycle has existed.
What this means for a British buyer
The practical upshot is that national vacancy figures answer a question most compute buyers are not asking. Nobody publishes UK vacancy segmented by achievable rack density, so the precise split is not knowable from public data. What the density figures above imply is that the great majority of available space sits in the single-digit to 30kW band, because that is what the installed base is. For a company deploying enterprise IT, storage or most inference workloads, that space is entirely fit for purpose, and the headline number means what it appears to mean. For a company deploying training hardware at rack scale, the relevant availability figure is a fraction of the published one, and it is not reported separately.
Geography compounds this. CBRE puts London at more than 80% of UK data centre supply, and within London the established clusters are power-constrained. In its 2025 mid-year review, CBRE noted that a key substation serving Slough, the largest single data centre sub-market in Europe, was not due for upgrade until 2029 at the earliest. The firm’s 2026 outlook expects more capacity to be built outside the established Slough and Hayes clusters precisely because power there is limited. So the concentration of British supply sits in the part of the country least able to add power to it.
For anyone specifying compute in the next couple of years, this reframes the search. The question is not whether space is available but whether a given hall can deliver sustained high density with liquid cooling in place, what its floor loading actually permits, and whether the power distribution behind the rack was designed for the load or adapted to it. Those questions are answered by a purpose-built colocation data centre in a way that a converted hall often cannot match, and the difference is not marketing. It is structural engineering, and it was decided when the building went up.
The counterweight
None of this makes the legacy estate obsolete, and it would be wrong to imply otherwise. The large majority of British computing is not AI training, and will run perfectly well at conventional densities for years. Retrofit projects do succeed, particularly where a building was overbuilt for its original purpose, and the value of an existing grid connection is high enough to justify considerable effort. Nor is extreme density always the right answer even for AI work: lower-density configurations of the same silicon exist, running in the 80kW to 90kW range, and plenty of inference deployments sit comfortably below that.
The narrower point stands. Britain’s data centre statistics currently report one market when there are two inside it, and the smaller of the two is where the AI investment is going. A buyer reading 8% vacancy and concluding that capacity is tight, but findable, has understood the first market and learned nothing about the second. Until the reporting separates them, the useful question is not how much space is left. It is how much of it can take the weight.
Gavin Lester is COO at Carbon-Z, a UK colocation provider running high-density and liquid-cooled data centres for AI and HPC workloads.