With the surging demand of AI, data centres create a significant environmental burden on electricity grids and fresh water resources. Under the recast Energy Efficiency Directive, big tech companies and data centre operators are required to quantify and report on the energy and water impact of their facilities. 

While this may appear to be a step into the right direction towards environmental sustainability, this reporting framework can be gamed to give a false sense of ecological gains. This is why a new EU strategy is urgently needed.  

A single click on your smartphone to send a picture triggers a transmission from your device to the nearest cell tower, through a network hub, and ultimately to a data centre server. A data centre is a physical facility powering almost all digital services, from social media posts to complex AI models.  Data centres for cloud-based and Large Generative AI applications are increasingly owned and scaled across Europe by major big tech companies. 

For the EU Commission, data centre expansion is an important building block to accelerate competitiveness, sovereignty and AI innovation. At the same time, investing in larger facilities comes with a significant price tag: facilities are expanding in water-stressed areas in Europe and are straining electricity grids in West London and Denmark.  

With surging demand for resource-intensive AI services and products looking unlikely to diminish any time soon, a reliable strategy is needed to assess and balance the expansion of data centres with environmental sustainability. 

Focusing on the facility-level reporting of the energy and water footprint in the recast Energy Efficiency Directive (recast EED), a new pre-print argues that the EU Commission is not currently succeeding in resolving these tensions. The recast EED produces specific obligations for data centre operators to report key indicators, such as on water and energy consumption, and using benchmarks in Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE). However, operators can adopt these recast EED endorsed PUE and WUE benchmarks to present “efficient” low scores.  

The core of the problem is an efficiency paradox where data centre operators can expand their facilities endlessly but maintain (or even reduce) their average scores across their facilities. This reporting framework risks obfuscating or ignoring the resulting strain of larger data centres on local and global energy and water resources.   

This paradox operates on two levels. First, operators can make their data centres more energy efficient by employing a range of “technical optimisation measures,” such as using existing servers more efficiently.

However, second-order effects, such as “rebounds,” can quickly eliminate these gains when AI demand rises. Second, data centre operators must scale aggressively to maintain their lower PUE and WUE scores, for example by increasing the number of servers or retrofitting cooling infrastructure, both of which pose additional environmental costs. 

These aggregate and individual effects are not captured in PUE and WUE metrics but point to a broader problem: data centre operators can achieve low scores while ignoring the second-order effects of scaling and retrofitting. Big tech can effectively follow its own market-incentives to expand and sustain larger AI workloads, ultimately driving further data centre expansion at the expense of local communities. 

Addressing this mismatch between reported and actual efficiency requires targeted revisions to the recast EED framework, focusing on a new Delegated Regulation that incorporates efficiency improvements, additional reporting PUE and WUE trade-offs and bespoke mechanisms to track endpoints over time. 

Download the full paper, “The fallacy of sustainable Generative AI: limitations in EU environmental regulation of data centres and paths forward”, by Daria Onitiu, Sandra Wachter and Brent Mittelstadt, currently available as a pre-print.  

Find out more about the work of Dr Daria Onitiu, Professor Sandra Wachter and Professor Brent Mittelstadt.