EU Label for Data Centres: How Europe Wants to Measure AI’s Hunger for Energy and Water

Set Trending Topics as a preferred source on Google.

The energy consumption of AI data centres is about to get a common yardstick in the EU. The European Commission has put forward a draft for a mandatory sustainability rating that would make the energy and water use of data centres comparable, much like the familiar energy labels on household appliances. In parallel, research institutions such as Silicon Austria Labs (SAL) are working on the hardware and software side of the problem, namely on delivering the same computing performance with less energy.

What the draft says

The planned rating is based on Article 33 of the EU Energy Efficiency Directive and builds on the existing European data centre database, into which operators already have to report. It would cover facilities with an installed IT power demand of at least 500 kilowatts; smaller sites may participate voluntarily.

A bundle of metrics is to be assessed: energy efficiency (Power Usage Effectiveness, or PUE), water efficiency (Water Usage Effectiveness, WUE) and the share of renewable energy, broken down into guarantees of origin, power purchase agreements (PPAs) and on-site generation. Notable are the tighter rules for green power claims: guarantees of origin would have to match actual consumption in time and come from the same bidding zone, and the underlying generation assets should be no more than ten years old. That makes it harder to make a site look green through purchased certificates alone.

The labels are to be generated automatically from the reported data, updated annually and published, with the first round foreseen for 2027. The legal act still has to pass scrutiny by the member states and the European Parliament. Beyond transparency, the rating is likely to carry weight because it is meant to feed into access to sustainable finance and into public procurement, and to play a role in the planned Cloud and AI Development Act.

The scale of the problem

The International Energy Agency expects electricity consumption by data centres worldwide to more than double by 2030, to around 945 terawatt hours, slightly more than Japan consumes in total today. How steep the curve is can be seen at a single company: Google’s data centres doubled their electricity consumption within four years, from 14.4 to 30.8 million megawatt hours. On top of that comes the water needed for cooling, which the planned label assesses as a second dimension.

Gas as the fast bridge

Because new grid connections and renewable capacity take years, operators, above all in the US, reach for what is quickly available: gas. The xAI supercomputer Colossus in Memphis runs to a considerable extent on gas turbines, and environmental groups accuse the company of exceeding air pollutant limits in doing so. Elon Musk has also bought an operator of fossil power plants to supply data centres with electricity. That the build-out is not without consequences shows up in retail prices too: in several US regions, AI data centres are pushing up household electricity prices.

In Europe this route is harder. Emissions trading, climate targets and expensive natural gas make fossil self-supply unattractive, while grids are the bottleneck in many regions. For sites such as the planned Google data centre in Kronstorf in Upper Austria, access to transmission capacity is as central an issue as the price of power.

Nuclear power, with a delay

The industry’s second answer is nuclear. Hyperscalers are securing electricity from existing reactors and investing in small modular reactors (SMRs); Google, for instance, has bought capacity from the startup Kairos Power. The appeal lies in baseload: a data centre runs around the clock, wind and solar do not.

Nuclear power, however, is not a short-term option. Most SMR projects will be operational towards the end of the decade at the earliest, and expectations swing widely: the SMR sector recently lost around 30 billion dollars in market value in a short space of time. Even in Austria, which rejects nuclear power politically, developers such as the Graz-based company Emerald Horizon are working on such concepts. Within the EU the question remains contested; an overview of the debate can be found in the AI Talk on nuclear power for AI.

The planned label does not rate the power source directly, but it does rate the share of renewable energy. Gas-fired and nuclear electricity therefore show up differently: fossil supply lowers the reported renewable share, while nuclear counts as low-carbon but not as renewable. How strongly that will influence site selection depends on how tightly financing and procurement end up being tied to the rating.

The lever at the root: more efficient chips

Alongside the question of where the electricity comes from stands the question of how much is needed in the first place. This is where the research of Silicon Austria Labs comes in, particularly in the fields of Embedded Systems and Intelligent Wireless Systems. SAL develops hardware accelerators, specialised chip components that carry out AI computations with far less energy than conventional processors. In the European project ISOLDE, SAL works within the Chips Joint Undertaking on such accelerators based on the open chip architecture RISC-V.

“Our strength lies in making AI efficient where it actually computes, in the chip and in the datacenter,” says Christina Hirschl, CEO of Silicon Austria Labs. “We develop hardware and methods that deliver the same computing performance with significantly less energy.”

A second approach moves computing work out of the data centre altogether: SAL optimises AI models so that they run directly on devices, in household appliances, sensors or satellites, instead of sending data to large data centres in an energy-intensive detour. In federated learning, a method that trains AI without pooling sensitive data, SAL is likewise examining particularly energy-saving approaches.

Politically, the work is framed as a question of competitiveness. “Our industrial strategy is built on the idea that ecological and economic goals belong together,” says Innovation and Infrastructure Minister Peter Hanke. “The fact that artificial intelligence is being researched here in Austria that is both powerful and resource-efficient shows that this path works.” Hirschl argues along similar lines: “Whoever masters the most economical technology becomes more independent and connects digitalisation with climate targets.”

What remains open is whether efficiency gains actually dampen consumption or whether cheaper computing simply leads to more computing. Which is why the decisive question about the planned label is likely to be less the letter scale itself than what is attached to it: funding, contracts and permits for new sites.