AI data centers consume vast energy and water. This guide explains PUE, WUE, grid strain, and the growing push for transparency and mandatory reporting.
Artificial intelligence has a physical footprint — and it is growing. Every query answered by a large language model, every image generated, every recommendation served at scale requires chips consuming real electricity inside real buildings that draw real water to stay cool. For years, the data center industry operated with relatively limited public disclosure requirements. That is changing. Governments, investors, and communities near major data center campuses are demanding answers, and the metrics they are reaching for — PUE, WUE, carbon-adjusted power consumption — are becoming the vocabulary of a new era of accountability.
This guide explains what the numbers mean, why AI is accelerating scrutiny, and what mandatory reporting could look like.
Traditional data center workloads — websites, databases, file storage — are computationally modest compared to training and running frontier AI models. A single large-scale AI training run can require thousands of specialized accelerators operating continuously for weeks. Inference — the process of actually running a trained model to answer a user query — is less intensive per request but happens at a scale that can dwarf training in aggregate energy terms. As AI services multiply and user adoption grows, the cumulative demand compounds.
This dynamic is explored in depth in our guide to AI inference costs and why running AI is expensive. The short version: AI workloads are among the most power-dense tasks data centers have ever been asked to handle, and the industry is building capacity at a pace that is straining local grids and regional water supplies.
Before discussing regulation, it helps to understand the metrics that already exist. Two have become industry benchmarks.
PUE measures how efficiently a data center uses the electricity it draws. The formula is simple:
PUE = Total Facility Energy ÷ IT Equipment Energy
A perfect PUE of 1.0 would mean every watt coming into the building goes directly to compute. In practice, cooling systems, lighting, power conversion losses, and other overhead push PUE above 1.0. Older facilities often post PUE values above 1.5; modern hyperscale campuses from operators such as Google, Microsoft, and Amazon have published figures closer to 1.2 or lower for some sites. However, PUE only captures electrical efficiency — it says nothing about where that electricity came from or how much water the cooling system consumed.
WUE fills part of that gap. It measures water consumption per unit of IT energy:
WUE = Annual Site Water Usage (liters) ÷ IT Equipment Energy (kWh)
Evaporative cooling towers — common in warm or arid climates — can require substantial water volumes. The figure varies enormously depending on local climate, cooling technology, and whether operators count only on-site water or also include water embedded in electricity generation (power plants also consume water). This distinction matters enormously for communities in water-stressed regions such as the American Southwest or parts of Europe experiencing longer drought cycles.
Neither PUE nor WUE captures everything. Carbon intensity of the grid, the source of cooling water, and local ecological effects are absent from both. That incompleteness is part of what is fueling demands for richer disclosure.
Data centers are not abstract cloud infrastructure — they are large industrial buildings that plug into local grids and municipal water systems. In regions with significant data center concentrations — Northern Virginia, Central Texas, the Netherlands, Singapore, and parts of Ireland — the cumulative load is measurable at the utility level.
Grid operators in several jurisdictions have publicly flagged that the pace of data center interconnection requests is outpacing grid expansion plans. When a large campus comes online rapidly, it can shift local energy prices, require new transmission investment, and affect reliability margins. Renewable energy procurement agreements — Power Purchase Agreements, or PPAs — are one tool operators use to address carbon concerns, but physical grid capacity is a separate problem that PPAs alone do not solve.
Water impacts are similarly localized. A data center in a water-rich region operates in a fundamentally different context than one drawing from an over-allocated river basin. The absence of standardized public reporting makes it difficult for regulators, communities, or investors to compare facilities or assess cumulative watershed stress.
Understanding these infrastructure tensions is also central to the AI cost crisis, where capital expenditure on data center buildout is reshaping the economics of the entire AI industry.
Despite voluntary reporting by some large operators, disclosure remains inconsistent. A company may publish a global aggregate figure for renewable energy coverage without disclosing site-level water consumption. Another may report PUE for a subset of facilities. Independent verification of self-reported figures is rare.
This opacity creates several problems:
The gap is not unique to AI — it predates the current AI buildout — but AI’s acceleration of data center construction has made the stakes considerably higher.
The European Union’s revised Energy Efficiency Directive (EED), which entered into force in October 2023, includes specific provisions for data centers. Operators of data centers with installed IT power demand of 500 kW or more in EU member states are required to report a standardized set of metrics — including PUE, WUE, renewable energy use, and server utilization rates — to national authorities, with data aggregated into a publicly accessible EU-wide register. The first reports, covering 2023 data, were due in September 2024; from 2025 onwards, operators submit annual reports by 15 May. This is the most comprehensive mandatory data center reporting framework currently in operation anywhere in the world.
The EU’s broader approach — layering sectoral rules into existing environmental and energy legislation — reflects its general preference for prescriptive disclosure requirements. The EU AI Act extends this logic further into AI-specific obligations, though environmental reporting for AI systems specifically is still evolving at the EU level.
The United States does not yet have a federal equivalent to the EED’s data center provisions. The Department of Energy has maintained voluntary efficiency programs and published research on data center energy use, but mandatory disclosure has largely been a state-level initiative. Several states with significant data center footprints have introduced or enacted reporting or permitting requirements that include energy and water data.
At the federal level, the Securities and Exchange Commission adopted climate disclosure rules in March 2024, but stayed them almost immediately to allow litigation to proceed. By March 2025, the Commission voted to stop defending the rules in court. As of May 2026, the SEC has proposed to rescind the rules entirely, stating they exceed the agency’s statutory authority — meaning they never went into force and are unlikely to take effect in their original form. Investor-facing climate disclosure at the federal level remains, for now, voluntary. The AI Regulation Tracker 2026 provides a continuously updated view of these and other relevant legislative developments.
The influence of major technology companies on how environmental rules are ultimately written is significant — a dynamic covered in detail in our guide to Big Tech’s influence on AI regulation and policy.
Several major cloud and AI infrastructure operators publish annual sustainability or environmental reports. These typically include:
The limitations are real. Aggregate global figures obscure wide variation across sites. Commitments to future neutrality are not the same as current performance. And the methodologies used — particularly for Scope 2 (purchased electricity) and Scope 3 (supply chain) emissions — differ across companies, making direct comparison unreliable.
Some operators have begun providing more granular, site-level data in response to local government requests or community pressure. This trend toward more localized disclosure reflects the reality that the impacts of data centers are inherently local, even if the compute they provide is global.
If transparency requirements expand — whether through the EU’s model, US state action, or bilateral frameworks — the practical effects would likely include:
For AI developers and cloud providers, more rigorous reporting also has strategic implications. Operators with genuinely efficient infrastructure would gain a verifiable advantage. Those relying on older, less efficient capacity would face pressure to invest in upgrades — costs that would flow through to AI service pricing, as explored in our guide to AI inference costs.
Last updated: June 2026