Asger Risom
Artificial intelligence is pushing data centers into a new era of growth. But the environmental cost of that surge is still hard to pin down.
Artificial intelligence is pushing data centers into a new era of growth. But the environmental cost of that surge is still hard to pin down. What’s missing most, researchers argue, is clear, comparable disclosure from the companies running the hardware.
Governments are being asked to regulate data center growth, power procurement, and water use. But policy is running ahead of the accounting.
The core question is simple: if AI demand is rising fast, what does that mean for emissions and water, and where should rules or incentives land?
Before that can be answered credibly, the study argues, reporting has to catch up.
Policy needs numbers
The study’s premise is that climate and water planning depends on knowing not just how much electricity AI uses, but where it is used, because grid pollution and water intensity vary widely by location.
Yet companies largely don’t break out AI from other computing in their reporting, making AI’s footprint hard to separate from the broader data center boom.
The paper says this gap forces outsiders to rely on approximations, even as data center impacts rise quickly enough to matter at national or city scales.
What can be inferred
On the power side, the IEA estimated AI accounted for 15% of data center electricity demand in 2024 (excluding crypto mining). Separately, “other recent research” cited in the paper suggests AI reached about 20% by the end of 2024 and could climb to 23 gigawatts of demand by end-2025.
For overall data centers, the IEA estimated 460 TWh of electricity in 2024 tied to about 182 million tons of CO2, and 560 billion liters of water consumption in 2023. The IEA did not specify the AI share of those totals, the paper notes.
Using company sustainability disclosures as a proxy and applying intensity factors, the paper’s own calculations estimate AI systems could emit 32.6 to 79.7 million tons of CO2 in 2025, depending on assumptions about where the load runs and how clean the power is.
Water is murkier
The paper stresses water accounting is tougher because it spans both onsite cooling and “embedded” water used in electricity generation. The IEA’s 2023 split attributes most data center water use to indirect consumption.
But indirect water is rarely disclosed. Meta is highlighted as an exception, and the study also cites Google’s choice not to report that metric in one AI disclosure because it “does not fully control the water consumption in electricity generation.”
From those comparisons, the paper estimates AI’s total water footprint could reach 312.5 to 764.6 billion liters in 2025, and warns common benchmarks may understate indirect water use.
A practical minimum disclosure set would be: data center electricity by site, the share used for AI, and site-level cooling water, alongside contracted power attributes. Mandating it would improve accountability, but it also raises tradeoffs around commercial sensitivity and security for facility-level details.
Sources: Patterns; A Cell Press Journal
