AI Environmental Impact Statistics 2026: Energy, Water & Carbon

Every article about AI's environmental impact quotes a different number, and half of them are from 2023. So we built the page we wished existed: the key statistics on AI's energy, water, and carbon footprint — every figure traced to a primary source (the IEA, Google's and Microsoft's own disclosures, peer-reviewed research), current as of July 2026, with the outdated myths flagged. Journalists and researchers: cite freely, sources are linked next to every number.

The 12 headline statistics

1. Data centers consumed 415 TWh of electricity in 2024 — about 1.5% of global demand (IEA)
2. That figure is projected to reach ~945 TWh by 2030 — roughly 3% of global electricity (IEA)
3. AI-specific servers used an estimated 53–76 TWh in 2024, heading for 165–326 TWh by 2028 (Carbon Brief)
4. AI's share of data-center power: 5–15% today → 35–50% by 2030 (Carbon Brief / IEA)
5. A median Gemini prompt uses 0.24 Wh — one second of microwave time (Google, Aug 2025)
6. An average ChatGPT query uses ~0.34 Wh (OpenAI / Sam Altman)
7. Median Gemini prompt efficiency improved 33× in one year (May 2024 → May 2025) (Google)
8. AI data centers used ~264 billion gallons of water in 2025 (Axis Intelligence)
9. Every 20–50 ChatGPT queries ≈ 500 ml of water, counting cooling + generation (UC Riverside)
10. 2030 projection: 9.3 trillion liters of annual AI water use if AI hits 40% of DC electricity (UN-cited)
11. Google's greenhouse emissions rose 48% from 2019 to 2024 (Alphabet environmental report)
12. Microsoft's emissions jumped 25% in 2025 alone, to 20 million tonnes CO₂e (Bloomberg, Jul 2026)

Energy: the doubling curve

The IEA's "Energy and AI" analysis is the reference dataset here. Data-center consumption has grown ~12% a year since 2017 — four times faster than total electricity demand — and at ~945 TWh in 2030, data centers would out-consume Japan. The nuance most coverage misses: AI is still the minority of data-center load today (5–15%). The projection that matters is the share shift — to 35–50% by 2030 — which is why every hyperscaler is suddenly signing nuclear power deals.

Per-prompt reality check: your query is not the problem

In August 2025, Google did something unprecedented: it published measured per-prompt figures for Gemini — 0.24 Wh, 0.03 g CO₂e, 0.26 ml of water for the median text prompt. That's the energy of running a microwave for one second, and it made the widely-quoted "3 Wh per ChatGPT query" claim from 2023 look an order of magnitude stale. OpenAI's Sam Altman followed with a ~0.34 Wh average for ChatGPT.

The honest framing: individual prompts are trivially cheap; the aggregate is not. Billions of daily prompts, model training runs, and — the part nobody talks about — idle reserved capacity are where the terawatt-hours go. Personal guilt about asking a chatbot a question is misplaced; scrutiny of infrastructure buildout is not.

Water: the 264-billion-gallon bill

AI data centers consumed roughly 264 billion gallons of water in 2025 — approaching a trillion liters in the US — and regional concentration makes it worse: Texas data centers alone are on track from 49 billion gallons (2025) toward a projected 399 billion by 2030. The counter-trend is real engineering progress: Microsoft's fleet water-use effectiveness hit 0.30 L/kWh in FY2025 (39% better than 2021) with zero-evaporation cooling saving 125+ million liters per new datacenter per year, Amazon reports 0.12 L/kWh, and Google replenished 64% of its freshwater use in 2024. Whether efficiency outruns growth is the open question — the 9.3-trillion-liter 2030 scenario assumes it doesn't.

Carbon: big tech's climate math is going backwards

The cleanest signal of AI's real footprint isn't per-prompt math — it's what the companies report. Alphabet's emissions grew 48% between 2019 and 2024, explicitly attributed to data centers and AI demand. Microsoft's rose 25% in 2025 alone — 20 million tonnes CO₂e, up from 16 million — driven by datacenter construction (concrete and steel are carbon-intensive before a single GPU powers on) and a pause in renewable-credit purchases that exposed the underlying number. Both companies still hold 2030 carbon-negative/net-zero pledges; both are currently moving away from them, not toward them.

For context on where all this compute is going, our AI agent stack overview maps the workloads driving the buildout, and the QuitGPT movement piece covers the consumer backlash — environmental concerns included.

Sources & how to cite this page

• IEA — Energy and AI (data-center consumption, 2030 projections)
• Google Cloud — Measuring the environmental impact of AI inference (per-prompt energy/water/CO₂)
• Carbon Brief — AI data-centre energy in context (AI server estimates, share projections)
• MIT Technology Review — Google's per-prompt energy data
• EESI — Data Centers and Water Consumption
• Bloomberg (Jul 2026) — Microsoft 2025 emissions report coverage; Alphabet environmental reports (2024–25)

Frequently Asked Questions