In our article on PUE we called it ‘the thermometer, not the diagnosis’. Today we take the next step, because there is a paradox worth facing head-on: the better your PUE, the less PUE tells you. In a newly built data centre, with a PUE between 1.1 and 1.3, 70–90% of the energy is absorbed by the IT systems — servers, storage, network. And that is exactly the part PUE does not measure: you can have the most efficient facility in the world, full of servers running at 10%.
From consumption to useful work
The right question is not just ‘how much energy goes in?’, but ‘how much useful work comes out?’. To answer it, you need to put a number on work capacity — the working capacity of the installed IT — and for years the industry had no shared method for calculating it.
For traditional servers the method now exists: The Green Grid — the same consortium behind PUE and the DCMM — has published a methodology based on PerfCPU, a representative capacity value derived from standard SERT tests and defined in an ETSI standard. The interesting part is how practical it is: the value is obtained from the processor code (or, failing that, from the number of physical cores) via public tables. No benchmarks to run in the server room: you just need to know what is installed, multiply and add up.
What about AI servers? And storage?
For accelerated servers — GPUs and other accelerators for training and inference — Uptime Intelligence proposes using 32-bit TFLOPS (the sum of the vector and matrix values declared for each accelerator). It is not a perfect measure, and purists object that real-world performance depends on memory, interconnects and software: true. But it is a standardised, obtainable and comparable value year on year — and in an industry that went without any metric for far too long, ‘representative’ beats ‘perfect’.
For storage the choice is even more pragmatic: the installed terabytes in dedicated systems. And here the report offers the figure that makes you stop and think: a storage system draws 80–90% of its maximum power even when it is not handling data, because in the meantime it is running integrity checks. Unused capacity means wasted energy: pushing system utilisation above 80% — instead of adding shelves — is one of the most concrete efficiency levers there is.
In Europe it is no longer (just) a good idea
These metrics are making their way into regulation: the European Energy Efficiency Directive (EED), through Delegated Regulation 2024/1364, requires data centres with an installed IT power of 500 kW and above to report ICT capacity indicators — server work capacity and storage capacity — in the European data centre database.
The problem? The data. Calculating capacity requires a component-level inventory: how many processors, of which model, in which data centre. According to Uptime Institute surveys, today fewer than one operator in two has catalogued its CPU codes, and only a third those of its accelerators. Anyone running infrastructure in Europe has a regulatory reason to put that right quickly; everyone else has an economic one.
Work per energy: the metric that closes the loop
Put together, the three ingredients — installed capacity, average utilisation and energy consumed — make it possible to calculate work-per-energy: how much work each megawatt-hour produces. It is the productivity metric that ISO 50001 requires for energy management, and it is also the most honest way to measure a consolidation project: before and after, with a single number — from which energy savings and return on investment follow.
The industry has been too slow to adopt a work metric, letting perfection stand in the way of the good.
That is the observation the report closes with, and we share it: the proposed metrics are not perfect, but they are calculable today, with data that a well-run organisation already has — or should have.
Where to start, in practice
- Component-level inventory: CPU and accelerator codes, raw storage capacity, linked to each site — this is the job of a CMDB done properly;
- Measuring utilisation and consumption: CPU and storage utilisation data, plus power at rack or PDU level, collected continuously through monitoring;
- A first snapshot of work per energy: even an approximate one gives you the baseline against which to measure consolidations and refreshes — and it slots into the DCMM assessment grid, where PUE remains the thermometer and these metrics begin to make the diagnosis.
The principle is the same one we apply to our green data centre: measure, design on the data, improve step by step. If you want to understand how much work every megawatt-hour of your infrastructure really produces, let's talk.
Sources and references: Uptime Institute Intelligence, Calculating work capacity for server and storage products (Briefing Report 174, May 2025); The Green Grid, IT work capacity metric V1 — a methodology (white paper #94) and TGG's position on the metric to calculate ICT server capacity (#92); ETSI EN 303 740; Delegated Regulation (EU) 2024/1364 (ICT indicators under the EED directive).
