AI environmental impact: the honest numbers.

Training a model used 1300 MWh once. Inference for a billion users uses far more. The order-of-magnitude shift in AI footprint is happening now, not five years ago.

AI's environmental footprint is two separate things: training (one-time, huge) and inference (per-query, multiplied by billions). The interesting story in 2026 is inference.

Training: large but bounded

Training a frontier-class model uses 50-300 GWh depending on size. That's roughly the annual electricity of 5,000-30,000 US homes. Significant. But it's a one-time cost amortized across every query the model serves. Also concentrated in a small number of labs.

Inference: the new dominant load

Per-query inference is cheap: ~0.0003 kWh per LLM call. But Google reports >1B daily AI queries. ChatGPT and Claude each see hundreds of millions. Sum across global production AI and inference is now larger than training annually.

The data-center electricity story

Data centers consume ~2% of US electricity in 2026, up from ~1.2% in 2022. AI is responsible for ~40% of that increase. Microsoft, Google, Amazon are signing 20-year power-purchase agreements for nuclear (Three Mile Island restart) and large-scale renewables. The hyperscalers' marginal electricity is increasingly carbon-free.

What this means for buyers

Procuring AI services? Ask your vendor: 1. What's the marginal carbon intensity of your inference electricity? Hyperscaler-hosted: increasingly green. Self-hosted in coal grid: not. 2. Do you offer carbon disclosure for AI usage? Azure and GCP both report this. Smaller AI providers usually don't. 3. What's the lifecycle plan for retired GPU hardware? 3-5 year useful life, then e-waste unless your vendor has takeback.

The honest framing

AI does have environmental cost. Not in the same league as steel, cement, or aviation. Comparable to streaming video. The marginal user-question is far less than driving to a meeting it replaced.

Risk is not AI specifically. It's growth. If AI inference grows 10x by 2030 (plausible), data center load doubles. That's the conversation worth having with utilities and policymakers.

What we tell agency leaders: prefer hyperscaler-hosted inference (cleaner electricity), require carbon disclosure in the contract, don't fall for marketing on either side (it's neither cure nor villain). It's an electricity load. Treat it like one.