Abstract

The environmental footprint of large-scale AI training is real, measurable, and systematically underreported. Labs publish headline compute figures in FLOPs but rarely translate these into the energy consumption, carbon emissions, and water withdrawal numbers that would allow meaningful comparison with other industrial activities or with alternative architectural choices. The gap between what is technically knowable and what is disclosed has begun to attract regulatory attention and investor scrutiny, but current voluntary reporting frameworks are not producing the data quality needed for accountability.

What the Numbers Actually Look Like

A large training run at frontier scale - the kind that produces a GPT-4 or Gemini Ultra class model - consumes energy in the range of tens of gigawatt-hours over weeks to months of cluster time. The carbon intensity of that energy depends on the grid mix of the data center region and time of day of operation; a run on a coal-heavy grid can have carbon emissions an order of magnitude higher than the same computation on a hydro-powered cluster. Water consumption for cooling is the less-discussed variable: air-cooled data centers draw significant water for evaporative cooling, and water-cooled facilities consume it more directly. Microsoft disclosed in its 2023 sustainability report that its global water consumption increased 34 percent year-over-year, a figure it attributed in part to AI infrastructure growth. Equivalent disclosures from other major labs are not available.

Why Voluntary Reporting Is Insufficient

The ML community has made genuine progress on measurement tooling: CodeCarbon, the Green Algorithms calculator, and MLflow’s energy tracking integration allow practitioners to log emissions per training run. What does not exist is a consistent reporting requirement. The EU AI Act’s general-purpose AI provisions include energy transparency requirements for models above the training compute threshold, but implementation guidance is still being developed. In the US, the SEC climate disclosure rule, after years of litigation, applies to large public companies and covers Scope 1 and 2 emissions but not product lifecycle emissions from inference at scale. Labs that are private companies or that structure their most energy-intensive work through cloud providers face minimal disclosure pressure.

Toward Credible Accounting

Credible environmental accounting for AI would require three elements that are currently absent from standard practice: reporting granularity at the model level rather than corporate aggregate, consistent methodology for attributing cloud compute emissions rather than passing that attribution to cloud providers, and inference footprint reporting alongside training footprint. Training is a one-time cost; inference at billions of queries per day compounds continuously. The latter is almost never reported. Research from the Hugging Face sustainability team and from Carnegie Mellon’s Software Engineering Institute has begun constructing inference footprint estimates from public API traffic data, but estimates are not disclosures, and the policy lever that converts one into the other remains unused.