AI's Energy Footprint

The Perception

AI is an environmental catastrophe in progress. Data centres are drinking rivers and burning the grid so that people can generate slop.

This is no longer a fringe position. It is the majority one, and it is hardening:

• 71% of Americans oppose an AI data centre being built in their area; 48% strongly oppose. That is more opposition than nuclear power plants attract in the same survey, with parallel question wording (53% — the highest reading in 25 years of Gallup's nuclear trend). Gallup, March 2026, n=1,000
• 39% say data centres are mostly bad for the environment — a plurality, not a majority; 31% land on neither/not sure, and the base includes the quarter of adults who've heard nothing about data centres at all. Only 4% say mostly good. On home energy costs: 38% bad, 6% good. The same survey finds data centres viewed net positively on local jobs (25% good vs. 15% bad) and tax revenue (23% vs. 12%). Pew, fielded January 2026, n=8,512
• The gap widens with exposure. Among Americans who have heard “a lot” about data centres, 63% say bad-for-environment and roughly two-thirds say bad-for-energy-costs — against 48% and 42% among those who have heard “a little.”

That last finding is worth sitting with carefully: Pew reports it as an association, not a cause. People who report hearing “a lot” about data centres are also disproportionately likely to live near one, or to have gone looking for coverage because they were already alarmed. The honest reading isn't “information makes people more negative” — it's that whatever is driving both the exposure and the alarm hasn't been settled by the coverage either.

What the Data Actually Shows

One clarification before the numbers: almost every figure below is the entire data centre sector — cloud, storage, streaming, ordinary corporate computing — not AI specifically. The IEA does not publish an AI-only electricity total. Its only AI-specific claim is that AI-driven “accelerated servers” account for roughly half of the sector's growth through 2030, not half its total consumption.

1. The entire global data centre sector was ~1.5% of world electricity in 2024. 415 TWh. In the IEA's Base Case that roughly doubles to 945 TWh by 2030 — “just under 3%” of projected 2030 demand, in the IEA's own words, not the rounder “about 3%” often quoted. IEA, Energy and AI
2. A text AI query costs about nine seconds of television. Google published measured figures in August 2025: 0.24 Wh, 0.03 g CO₂e, 0.26 mL of water per median Gemini text prompt. Ten queries a day is roughly 0.03% of a British user's daily electricity use, or 0.1% for an American user.
3. The per-query trend runs against the doom narrative. Between May 2024 and May 2025, Google reports energy per prompt fell 33× and its carbon footprint fell 44× — the larger carbon figure mainly reflects clean-energy purchasing, not compute efficiency alone. Independent estimates moved the same direction: Epoch AI now puts a typical query near 0.3 Wh, roughly an order of magnitude below the ~3 Wh figure that circulated through 2023–25.
4. Data centre water is real, but the direct-only figure usually quoted badly understates it — see Figure 2 for the same-basis comparison.
5. Globally, data centres are about one-tenth of electricity demand growth to 2030. Space cooling and heat pumps account for nearly half the buildings-sector growth worldwide; EVs are over 10%. A major driver, not the driver. IEA, Electricity 2026

The Long View

Data centre emissions today total 180 Mt CO₂ — under 1.5% of total energy-sector emissions. The IEA's 2035 base case is 300 Mt; even the aggressive “Lift-Off” case tops out near 500 Mt. For scale, global energy-related CO₂ runs above 37,000 Mt. AI could triple its emissions and remain inside the noise band of the global total.

The trendline that matters is efficiency, not consumption. Every compute-intensive technology in history — mainframes, PCs, mobile, streaming — has followed the same arc: alarming early per-unit costs, a steep efficiency collapse, then a demand rebound that partially eats the gains. AI is currently mid-collapse, which is precisely why the 2023 per-query estimates were so badly wrong by 2025. Straight-lining today's 12%-a-year growth to 2040 assumes the one thing that has never held.

The Perception Gap

1. The missing denominator. “945 TWh by 2030” sounds apocalyptic. “About 3% of global electricity” sounds like a manageable industrial sector. They are the same number. The press reliably prints the first form. This is Rosling's size instinct, working exactly as described.
2. Category confusion. The entire data centre buildout gets attributed to AI chatbots. AI is a portion of a sector that also runs banking, logistics, streaming, and every corporate back office on earth.
3. Locality projected onto the planet. The real, felt harm is a substation in Loudoun County and a utility bill in Maryland. That is a legitimate local grievance being narrated as a planetary one — and the planetary framing makes the local problem harder to fix, because “kill the planet” does not route to a rate case.
4. The awareness paradox. Pew's finding that better-informed Americans hold more negative views suggests the coverage feeding public understanding is itself uncalibrated: heavy on absolute numbers, light on comparison.

Appendix: The Actual Question in the Thread

The implied claim was narrower and more interesting than the global one: does AI-assisted editing that saves 20 hours of human work cost more or less energy than the 20 hours?

Nobody has published the number needed to answer this. Google's 0.24 Wh figure covers text inference. No major lab has released measured per-call energy for video-understanding inference, which is meaningfully more expensive. So this stays honest as bounds, flagged as estimate rather than finding:

• 20 hours of editing workstation time at a typical 200–400 W draw ≈ 4–8 kWh. Estimate — assumption stated, not measured.
• An AI pass over a clip library — a few hundred multimodal calls — plausibly lands well under 1 kWh at text-comparable rates, but could reach several kWh if video understanding runs 10–100× a text prompt. Unresolvable with public data.

The one clear flip: generating video is the expensive case — Hannah Ritchie singles it out as the exception where personal AI footprint stops being negligible. Cataloguing b-roll you already shot is almost certainly a net energy save. Generating synthetic b-roll instead of shooting it is a different question with a different answer.

Key Data Points

Sources

1. Energy and AI — Executive Summary · IEA, 2025
2. Electricity 2026 — Demand · IEA, 2026
3. Measuring the environmental impact of AI inference · Google Cloud, Aug 2025
4. What's the carbon footprint of using ChatGPT or Gemini? · Hannah Ritchie, Aug 2025
5. US views of how data centers affect the environment, energy costs, jobs and more · Pew Research Center, Mar 2026 (fielded 20–26 Jan 2026)
6. Americans Oppose AI Data Centers in Their Area · Gallup, Mar 2026
7. How thirsty is AI? · Andy Masley, compiling LBNL 2024, USDA NASS 2023, EPRI Feb 2026
8. How much have data centers increased electricity prices? · PolitiFact, Jun 2026
9. Microsoft's emissions surged 25% in 2025 during data center boom · Fortune, Jul 2026
10. Google's AI boom sends emissions, power use soaring · Axios, Jun 2026