The fallacy of sustainable Generative AI: limitations in EU environmental regulation of data centres and paths forward
arXiv preprint · Daria Onitiu, Sandra Wachter et al.
This arXiv preprint is a legal and policy analysis of facility-level PUE and WUE reporting under the recast EU Energy Efficiency Directive (Art. 12), and describes an 'efficiency paradox'.
On this page
Overview
The work is a 2026 arXiv preprint by Daria Onitiu, Sandra Wachter and Brent Mittelstadt (arXiv:2607.22604). It is a legal and policy analysis of the facility-level PUE and WUE reporting required by the recast EU Energy Efficiency Directive (EED, Art. 12) and its delegated acts for the common data-centre rating scheme.
The authors describe an 'efficiency paradox'. Because PUE and WUE are ratios, operators can keep or even lower their reported scores while expanding into ever larger AI facilities; gains from technical measures such as virtualisation, accelerators and batching diminish while rebound effects push absolute energy and water use up. For example, cutting IT load alone (through virtualisation) does not show up as an improvement and can even raise PUE unless the infrastructure is also optimized.
Within the regulation, EED Article 12 requires facilities with 500 kW or more of power demand to report key metrics, including PUE and WUE, to a European database, using the CEN/CENELEC (EN 50600-4-x) methodology under Delegated Regulation (EU) 2024/1364. Illustrative EU averages from the first EED technical report are PUE 1.36 and WUE 0.58 L/kWh for mid- and large-scale data centres, with a lower average PUE of 1.21 for sites above 10 MW. The second technical report suggests targets of PUE 1.3 for new data centres and WUE below 0.4 m3/MWh.
The paper notes that WUE under the CEN/CENELEC standard lets operators include or exclude potable and non-potable sources and does not require linking water sources to local water stress. The cooling choice trades water against energy, so WUE must be read together with PUE. The authors propose three EU measures: a delegated regulation documenting PUE/WUE trade-offs, minimum performance standards that act as a comparative baseline including water stress, and a mechanism to track diminishing returns and rebound effects over time.
Key points
- EED Article 12 requires facilities with 500 kW or more of power demand to report key metrics, including PUE and WUE, to a European database using the CEN/CENELEC (EN 50600-4-x) methodology under Delegated Regulation (EU) 2024/1364.
- PUE is annual total data-centre energy (kWh) divided by annual IT energy (kWh); WUE is annual water use (m3) divided by annual IT energy (MWh).
- Illustrative EU averages from the first EED technical report: PUE 1.36 and WUE 0.58 L/kWh for mid- and large-scale data centres; average PUE 1.21 above 10 MW.
- The second EU technical report suggests targets of PUE 1.3 for new data centres and WUE below 0.4 m3/MWh.
- WUE under the CEN/CENELEC standard lets operators include or exclude potable and non-potable sources and does not require linking water sources to local water stress.
- The cooling choice trades water against energy: evaporative towers in hot or dry climates raise WUE, while air-side economisation uses little water but more energy.
- The authors propose three EU measures: trade-off documentation, minimum performance standards and a mechanism to track diminishing returns and rebound effects.
- Because PUE and WUE are ratios, a growing facility can keep low scores while absolute use rises.
Using it in design
- For EU projects, add an EED compliance module that flags facilities of 500 kW and above as reportable.
- Compute annual PUE and WUE per the EN 50600-4-2 and EN 50600-4-9 definitions with a source breakdown (potable, freshwater, reclaimed, seawater), and compare against the suggested benchmarks for new builds (PUE 1.3, WUE below 0.4 m3/MWh).
- When users compare cooling options, always show PUE and WUE side by side as a trade-off, and record a written rationale for choosing a water-intensive system.
- Report absolute annual MWh and m3 next to the ratios, so a facility that grows while keeping a low PUE/WUE still shows its rising absolute use.
- Cite the related standards (EN 50600-4-2, EN 50600-4-9, ISO/IEC 30134-4) in reporting.
Monthly newsletter · free
Keep your reading list current
Once a month: new library sources, changes to Turkish regulations and upcoming data centre events.
SoSo CAD
Compare the cost of design decisions
Evaluate equipment costs, energy expenses and total cost of ownership (TCO) using your inputs and assumptions in SoSo CAD.
Request a demo
Who should read it
Engineers, policy specialists and compliance leads designing or operating data centers in the EU who deal with EED reporting and sustainability policy.
Related standards
- EN 50600-4-2
- EN 50600-4-9
- ISO/IEC 30134-4
- EU Energy Efficiency Directive (EU) 2023/1791
- Commission Delegated Regulation (EU) 2024/1364
- EU Code of Conduct for Data Centres
- EU Taxonomy Climate Delegated Act (EU) 2021/2139
Publication details
- Authors
- Daria Onitiu; Sandra Wachter; Brent Mittelstadt
- Publisher
- arXiv preprint
- Year
- 2026
- Document number
- arXiv:2607.22604
- Document language
- English
- Source type
- Scientific
- Access
- Free to read
- Licence
- arXiv distribution permission — reuse conditions remain with the rights holder
Original licence note
arXiv default non-exclusive licence; not redistributable.
Citation
Daria Onitiu; Sandra Wachter; Brent Mittelstadt (2026). The fallacy of sustainable Generative AI: limitations in EU environmental regulation of data centres and paths forward. arXiv preprint. arXiv:2607.22604. https://doi.org/10.48550/arXiv.2607.22604
Events on this subject
Browse the events calendarAll content and rights belong to the respective authors and publishers. SoSo CAD provides only the citation, its own summary and a link to the source, and is not responsible for the sources' content.