Operational notes Engineering

Data centre carbon footprint: what goes into the count and what stays out

7 min read

Star-shaped metal cookie cutters on a floured work surface, behind them the rolled sheet with the shapes already cut out and a rolling pin, black and white photograph
The sheet is all there: the number depends on where the cutter is set down, and the offcuts stay outside the count.

You have already handed somebody your data centre’s carbon footprint figure: in a tender, in a sustainability report, in an answer to a client questionnaire. A competitor with a room identical to yours can hand over one three times lower without writing a single false word. What is missing under your number is the line that says where the count ends: without it, one question takes the whole thing apart — what did you put inside.

Six stages in the definition, three in the count

Recommendation ITU-T L.1308, Guidelines for the implementation of low-carbon data centres to climate change mitigation and adaptation, edition 12/2025, first edition, was approved on 14 December 2025 by Study Group 5. The catalogue shows it In force, with a single component in force, and the English text is freely downloadable.

The life cycle, in clause 3.2.2, has six stages: site selection, design, procurement, construction, operation, decommissioning. The emissions count, in clause 6.5, uses three. The first three are left out because their emissions are “low and difficult to calculate”, and so “the carbon emissions in these three stages are not considered in this recommendation”.

What remains is construction, operation and demolition, added together in metric tonnes of CO2. Construction counts the production and transport of materials and equipment, plus site plant, vehicles and personnel; demolition counts engineering plant, vehicles and demolition personnel. Operation counts “the operation of IT equipment and services, the cooling system, power equipment, and other data centre equipment”.

Read the operation line again: it says operation of, the running of the kit. The servers you swap out every four or five years with the room live are not on that list. Their manufacture sits somewhere else. For a room that has been through two refresh cycles that is no rounding error, and where it lands decides the size of the number you publish.

Somewhere else means scope 3, and scope 3 is optional

Clause 6.5 continues: “Econstruction stage, Eoperation stage, Edemolition stage may include scope 1 emissions, scope 2 emissions and scope 3 emissions, if the data centre chooses to assess scope 3 emissions”. Chooses. And the data needed to do it cannot be demanded from anybody: at the procurement stage the Recommendation asks that “vendors should be encouraged to provide the carbon footprint reports of the facilities”. Encouraged. A recommendation cannot make a supplier produce a figure it does not hold; a tender can. Whoever fails to obtain them reports without scope 3, and reports less — the same boundary on which the allocation of shared infrastructure between host and tenant is settled.

Whoever declares the number picks the numerator of CUE

Clause 6.6 introduces the ratio metric, carbon usage effectiveness: CUE = E / PIT, with PIT the IT equipment power usage in megawatt-hours. And E? It “is the GHG emission of the data centre, which can be Etotal, Econstruction stage, Eoperation stage, Edemolition stage, as needed”.

Four admissible numerators, times two — scope 3 in or out: eight legitimate numbers for the same facility, none of them wrong. And if the renewable electricity bought in is carried at zero emissions, the spread widens again: clause 7.5.4.2 accepts that consumption purchased from outside be confirmed by renewable energy purchase vouchers or third-party authentication certificates.

The definition, in clause 3.1.6, is taken from ISO/IEC 30134-8:2022, Carbon usage effectiveness (CUE), edition 1.0, published on 24 March 2022: a paid standard, and we have not read it. In L.1308 it reads “Ratio of the data centre annual CO2 emissions to IT equipment energy demand”, and the public catalogue record adds that CUE quantifies emissions “during the use phase of the data centre life cycle”. Annual emissions, use phase. Clause 6.6 instead admits the whole life cycle as a numerator, or demolition on its own: same acronym, different meanings.

The only line about the boundary does not fix the boundary

In the text of L.1308 the word boundaries appears exactly once. It is in clause 6.5, and it is a requirement: “When comparing emissions across time periods or multiple data centres, the accounting boundaries shall be consistent”. Under the conventions in clause 5, shall indicates “a requirement that must be followed strictly and from which no deviation is permitted”. Of the eleven instances of shall in the body, this is the only one that concerns accounting. The others deal with siting and environmental capacity, cooling water, the impact on the local grid, cabling materials, and a generic high standard for PUE and REF. But it says consistent, not which one. The boundary stays a choice made by whoever declares the figure, and it becomes binding only when somebody writes it into a contract.

Appendix I — which does not form an integral part of the Recommendation — shows where that leads: a data centre in north-western China with PUE below 1.2 and REF at 100%, for which “total GHG emission is zero, CUE is zero”. Total. A building put up without concrete and without site plant does not exist: what exists is a boundary that does not count those emissions, and the appendix does not declare it.

The five lines that make a number comparable

Which stages are in the count and which are not, named one by one. Whether scope 3 is included, and which of the 15 upstream and downstream categories. Which emission factor applies to electricity, with source and year, and whether purchase certificates enter the calculation. Over which period PIT is measured, and at which metering point. Each of those is a line a supplier can answer in writing before the award, and none of them can be reconstructed from a spreadsheet afterwards. And the clause that turns the obligation in 6.5 into a contractual commitment: this year’s boundary is next year’s boundary, and any change arrives together with the recalculation of the previous year.

Thresholds are a different matter, and we have already dealt with them: the Recommendation gives them for PUE (1.3 in cool climates, 1.4 in warm ones), for REF (75%) and for the reuse and recycling of e-waste and batteries (100%). For CUE there is no number at all: only emission factors “significantly below the local grid average” and a trend towards zero.

What we do not know

  • The period of PIT. Clause 6.6 measures it in megawatt-hours but does not say over what span: with E covering the whole life cycle and PIT covering a year, the ratio means nothing.
  • The content of ISO/IEC 30134-8. A paid standard: we know its number, title, edition, date and catalogue record. Whether its measurement categories would rule out the numerators L.1308 admits, we cannot say.
  • How much the three excluded stages weigh. The Recommendation offers no order of magnitude anywhere, so there is no telling whether leaving them out moves the total by a percent or by a fifth.
  • Whether demolition covers waste treatment. The item lists plant, equipment and personnel; material recovery sits in clause 7.6, outside the formula.
  • The boundary of the case study. The appendix gives the results and does not declare what it counted.

The point

A number without a boundary defends nothing, and the boundary is not fixed by the Recommendation: it is fixed by the tender document. In the specifications and compliance work we deliver, those five lines are clauses, checked one by one at acceptance: the declared meter exists where it is declared, and the delivered value can be rebuilt from the raw data. On the other axis, meters, invoices, purchase certificates, waste registers and as-built records converge into a single map of the site, where an AI flags the meter that has stopped incrementing, the emission factor left at last year’s value and the drift that in December takes you outside the figure you declared, and a team of ours steps in before that number ends up in a tender. Where an AI is needed on those data, it runs inside the client’s perimeter, on-premise or in a dedicated cloud with data centres in Italy, together with CSIDIA, the other company in the group.

Do you have to declare a carbon footprint in a tender, or check the one the room operator handed you? Talk to an engineer: the site visit is free of charge, and the boundary is settled before the number is written.

Sources