The planet: measuring and governing AI’s footprint
Lesson 4 of 5 in AI, Elections, Labor, Competition, and the Planet.
Every environmental-governance conversation about AI starts with a measurement problem, so start there. An AI system’s footprint has four components that get conflated constantly: embodied impacts (mining, chip fabrication, construction), training energy (large, one-time, headline-friendly), inference energy (small per query, colossal at deployment scale — for successful systems it dwarfs training over the model’s life), and water (evaporative cooling, plus the water embedded in electricity generation). Each number depends radically on where and when the compute ran: a training run on a hydro-heavy grid at night and the identical run on a coal-heavy grid emit wildly different carbon. Any single-number claim — per-query, per-model, per-company — that does not state its boundaries is marketing.
The scale that forced regulators to care: data centres consumed roughly 1.5% of global electricity in 2024, with the IEA projecting a doubling toward ~945 TWh by 2030, AI the main driver; US data-centre demand was on track for up to 12% of national electricity by 2028 on government lab projections. Those load curves — not ethics papers — are what put AI on energy ministries’ desks.
Where the footprint hides — and who must disclose what
- Embodied: chips, fabs, construction
Semiconductor fabrication is energy- and water-intensive and geographically concentrated. Almost no AI-specific disclosure regime reaches this layer; it hides in suppliers’ Scope 3.
- Training runs
The layer regulation actually touched first: EU AI Act Annex XI requires GPAI providers to document known or estimated training energy consumption.
- Inference at scale
Per-query costs are small; multiplied by billions of queries they dominate lifetime footprint. The least-regulated, worst-measured layer.
- Data-centre operations
Where energy law bites: EU Energy Efficiency Directive reporting for centres ≥500 kW, German PUE mandates, local permitting and water fights.
- Grid & water systems
Carbon intensity varies by location and hour; 24/7 matching vs annual offsets is the fight behind every “100% renewable” claim. Water draw concentrates in specific — often drought-prone — basins.
- Corporate disclosure (CSRD / ESG)
AI compute lands in Scope 2 and 3 of corporate sustainability reporting — where boards and investors, not AI regulators, apply the pressure.
Now map the actual law, which is thinner than the discourse suggests and lives mostly outside AI statutes. The EU AI Act touches energy in three places: Annex XI makes systemic-scale GPAI providers document the known or estimated energy consumption of training; the standardisation machinery is instructed to pursue resource-performance improvements; and Article 95 invites voluntary codes of conduct on environmental sustainability. That is a documentation-and-encouragement regime — no cap, no efficiency floor. The harder instruments are energy law: the recast Energy Efficiency Directive requires data centres above 500 kW to report energy, water, and efficiency metrics into an EU database (first reports 2024), and Germany’s Energy Efficiency Act imposes actual PUE limits and waste-heat-reuse duties on new data centres. Corporate sustainability reporting (CSRD) then folds AI compute into Scope 2/3 disclosures — though verify its post-omnibus scope, which was cut substantially in 2025–26. The US has no federal AI-environment mandate; the proposed AI Environmental Impacts Act never passed, leaving state utility regulation and local permitting as the venue. On standards, watch ISO/IEC 20226 on the environmental sustainability of AI and the Software Carbon Intensity specification — check publication status before citing either as settled.
Myth: “Training is the environmental problem”
Training is the measurable problem — one run, one bill, one disclosure line — which is why Annex XI targets it. For any widely deployed model, inference dominates lifetime energy: billions of queries outweigh one training run within months. The regulatory attention is inversely proportional to the impact, purely because training is easier to count.
Myth: “Our cloud is 100% renewable, so our AI is carbon-free”
Most such claims rest on annual offsets and certificates: buying enough renewable credits over a year to match consumption, while the actual megawatts at 3 a.m. came from gas. The stricter standard — 24/7 carbon-free matching, hour by hour, on the local grid — is what a diligence-grade review checks for. Ask which accounting the claim uses before repeating it in a sustainability report.
Myth: “Water use is negligible”
Evaporative cooling consumes water in the basin where the data centre sits — and siting economics push centres toward cheap land and power that is often drought-prone (Arizona, parts of Spain, Chile). Aggregate percentages look small; the governance issue is local concentration, which is why water became the flashpoint in permitting fights from The Dalles to Zaragoza.
Fact with a caveat: “AI also serves the climate”
True — grid optimisation, materials discovery, climate modeling — and it is the standard industry counterweight to footprint criticism. The expert move is to refuse the netting: benefits accrue to society if they materialise; the footprint accrues now and is measurable. Report both, offset neither against the other without evidence.
Key terms: carbon footprint, energy disclosure, pue, csrd, water usage
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.