The finance and sustainability teams inside most UK enterprises still treat the question of where their computing runs as a facilities matter, settled once and rarely reopened. AI has ended that. A single GPU training cluster can draw more power in a quarter than a mid-sized office estate uses in a year, and that load is landing on corporate inventories faster than the emissions attached to it can be counted. Uptime Institute’s 2025 global survey found that the share of operators tracking Scope 1 to 3 emissions fell over the year, at the moment AI made those numbers harder to pin down. The reporting is thinning as the load gets heavier.
The uncomfortable part is that the same cluster, doing the same work, produces materially different emissions depending on the building it sits in. Grid mix, power usage effectiveness and cooling architecture between them can move the reported figure by a wide margin. For any organisation with a science-based target or a CDP disclosure to defend, where you place high-density compute has become one of the larger levers available, and one of the least examined.
The number depends on who is counting
Start with the accounting, because this is where most comparisons fall apart before they begin. For a company renting rack or cage space under a typical colocation agreement, the electricity its servers consume usually falls in Scope 3 rather than Scope 2, because the tenant neither owns nor operates the facility. Under the GHG Protocol, it most often lands in Category 8, upstream leased assets, though some organisations book colocation as a purchased service and report it under Category 1 instead. The same physical kilowatt-hour can therefore appear in different parts of two companies’ inventories: Scope 2 for the operator that buys the power, Scope 3 for the tenant that uses it.
This is not pedantry. It means a like-for-like comparison between running your own server room and moving into colocation requires you to follow the same energy across a scope boundary. It also means two competitors reporting the same workload can show very different carbon numbers depending on consolidation approach and contract structure. Anyone signing against a provider’s headline emissions figure without knowing which method produced it is signing against a number they cannot audit.
Location-based, market-based, and the gap between them
The second fracture is methodology. The GHG Protocol permits two ways to report the emissions of purchased electricity. The location-based method applies the average carbon intensity of the local grid. The market-based method applies the emissions of whatever the buyer has contracted for: a power purchase agreement, a green tariff, or renewable certificates, which in the UK are REGOs.
The two can diverge sharply. Britain’s grid ran at about 126 grams of CO2 per kilowatt-hour on a generation basis across 2025, roughly flat on the previous year’s record low, according to Carbon Brief’s analysis of NESO and government data. The emission factor that UK companies apply for statutory reporting is a separate, lagged figure, which fell about 14.5% for 2025 to roughly 0.177 kilograms per kilowatt-hour. Two defensible bases, two different answers, before any renewable claim is layered on top.
Where in the country the facility sits matters as much as the certificate attached to it. Regional intensity varies widely: analysis of NESO data by Hoare Lea found a London load can carry roughly a quarter less carbon than the national average would suggest, while parts of northern Scotland ran zero-carbon for more than 60% of the year. A workload placed on a cleaner regional grid starts from a lower number before anyone buys a single instrument.
That is where the real scrutiny now falls. A facility can report near-zero market-based emissions by buying unbundled REGOs, certificates detached from the physical electricity, often from plants built years ago that would generate regardless. These are compliant under the current GHG Protocol. They are also the weakest form of the claim. A 2022 paper in Nature Climate Change concluded that renewable certificates let companies overstate their emission reductions, and the RE100 technical criteria have tightened to exclude unbundled certificates in some markets. The GHG Protocol is now in its first major revision in a decade, with time-matching of supply and demand under active discussion. A provider leaning on annual, unbundled certificates today may find the claim does not survive the next version of the standard.
There is a counterintuitive trap here that catches operators early in their renewable journey. If most of a facility’s supply is residual grid power and only a fraction is covered by certificates, the market-based figure can come out higher than the location-based one, because residual mix carries a heavier factor than the grid average. A provider that reports only the flattering number, or only one of the two methods, is not giving you enough to work with. SECR already requires qualifying UK organisations to disclose both.
Where cooling changes the maths
Now the physical layer, where the differences become large rather than presentational. Power usage effectiveness measures how much of a facility’s total power reaches the IT equipment rather than being spent on cooling and overhead. The industry has been stuck for years. Uptime Institute put the global weighted average at 1.54 in 2025, essentially unchanged for six consecutive years, held back by legacy air-cooled halls. A PUE of 1.54 means roughly a third of the power drawn never does any computing.
AI density breaks air cooling well before it breaks the grid connection. A traditional hot-aisle and cold-aisle hall typically runs a PUE between 1.4 and 1.8. Direct-to-chip cooling, where coolant is brought to a cold plate on the processor, and immersion cooling, where the hardware sits in dielectric fluid, both cut the cooling overhead sharply. Reported partial PUE for these approaches commonly falls to around 1.02 to 1.08. On a facility drawing tens of megawatts, the difference between 1.5 and 1.1 is not a rounding error. It is a large block of emissions that either exists or does not, for identical compute.
Cooling architecture also decides whether the waste heat is worth anything. Air-cooled exhaust is low-grade and hard to reuse. Direct-to-chip systems return water warm enough, in the range of 35 to 50 degrees, to feed district heating or an industrial process, which is where a facility’s effective efficiency can drop below the theoretical floor of a standalone building. For an organisation whose target reaches past its own four walls, that recovered heat is a line item, not a footnote.
This is the practical reason the emissions case increasingly favours moving AI and HPC workloads out of self-run rooms and legacy halls and into purpose-built high-density colocation. The efficiency gain is real and measurable. What matters for a defensible inventory is whether the buyer can see the underlying numbers rather than take a marketing figure on trust.
What an audit-defensible comparison actually needs
Put the three layers together, and the shape of a credible assessment is clear. It needs the facility’s measured PUE, verified in operation rather than quoted from a design document, and ideally broken out for the specific hall the workload will occupy rather than a site-wide average that flatters. It needs both the location-based and market-based emissions, with the contractual instruments behind the market-based figure named, so their additionality can be judged. It needs clarity on the cooling architecture and whether any heat recovery is real or aspirational. And it needs the scope treatment stated plainly, so the number lands in the right category in your own inventory rather than being quietly double-counted or dropped.
None of this is exotic. It is the difference between an emissions figure that holds up in front of an auditor and one that reads well in a brochure.
The honest counterweight
Colocation is not automatically greener, and it would be misleading to imply otherwise. A poorly run facility on a dirty regional grid can be worse than a well-managed enterprise room on a clean one. Efficiency gains can be swallowed by the sheer growth in compute they enable, a rebound effect no PUE figure captures. And market-based instruments, used well, on new-build power purchase agreements rather than recycled certificates, are a legitimate and important tool, not a dodge.
The point is not that colocation wins by default. It is that the decision now carries an emissions weight it did not a few years ago, and the organisations still treating it as a flat line item are the ones most likely to be holding an indefensible number when someone finally checks. The compute is coming either way. The only question is whether the emissions attached to it are measured before the contract is signed, or discovered afterwards.
