Will more efficient AI chips fix AI’s energy problem?
The claim: “More efficient AI chips will fix AI’s energy problem.”
Not on their own. Efficiency is improving fast: the IEA says energy per AI task has fallen at least tenfold a year recently, and Google says its median Gemini text prompt used 33x less energy over 12 months. But use grows faster. Google processed 50 times more tokens a month than a year before, and data centres’ electricity consumption grew 17% in 2025. Even in the IEA’s High Efficiency case, data centres use around 970 TWh in 2035, about double 2025.
The bigger pictureThe world’s data centres used about 485 TWh in 2025, about as much electricity as France uses in a year. The IEA expects that to roughly double to 950 TWh by 2030.
- 485 TWh 2025, estimated: IEA, Key Questions on Energy and AI (2026)
- 700 TWh 2035, Headwinds case: lowest case: slower AI uptake plus efficiency
- 970 TWh 2035, High Efficiency case: faster efficiency gains, 20% below the base case
Unit: TWh a year.
Why the claim misleads
Cheaper answers get used more, and heavier uses such as video, reasoning and agents can take hundreds or thousands of times more energy per query than plain text (IEA). In the IEA’s forecasts, the only case that levels off, at around 700 TWh, pairs efficiency with slower AI uptake.
What nobody knows yet
- Nobody outside Google has checked its 33x figure, and Google doesn’t publish its total AI electricity, so the fall per prompt can’t be set against the 50 times rise in tokens (prompts and tokens are different units).
- The 2035 figures are IEA forecasts. How fast people take up AI, and how heavy new uses get, matter as much as chip efficiency, and nobody knows them yet.
Sources (12)
- IEA, Key Questions on Energy and AI (2026) Estimated
“Software and hardware advances have resulted in the energy use per AI task dropping by at least an order of magnitude annually in recent years.”
Quote checked on the page, 25 Oct 2026 - Google Cloud (median Gemini Apps text prompt; Google’s own figure) Company says so
“over a recent 12 month period, the energy and total carbon footprint of the median Gemini Apps text prompt dropped by 33x and 44x, respectively”
Quote checked on the page, 25 Oct 2026 - Google (Sundar Pichai, I/O 2025 keynote; tokens across Google’s products and APIs) Company says so
“This time last year, we were processing 9.7 trillion tokens a month across our products and APIs. Now, we’re processing over 480 trillion — that’s 50 times more.”
Quote checked on the page, 25 Oct 2026 - IEA, Key Questions on Energy and AI (2026) Estimated
“The global electricity demand of data centres – the critical infrastructure for training and running AI models – grew by 17% in 2025, in line with IEA projections.”
Quote checked on the page, 25 Oct 2026 - IEA, Key Questions on Energy and AI (2026) Estimated
“However, new energy-intensive AI applications are increasingly being launched and used, such as those for video generation, reasoning and agentic tasks. These kinds of tasks can consume hundreds or thousands of times more energy per query than simple text generation.”
Quote checked on the page, 25 Oct 2026 - IEA, Key Questions on Energy and AI (2026) Estimated
“The energy demand of AI is therefore the result of three rapidly evolving and uncertain trends: improvements in efficiency, surging uptake, and changing model capabilities”
Quote checked on the page, 25 Oct 2026 - IEA, Energy and AI (2025), High Efficiency Case Forecast
“This unlocks energy savings of more than 15%, with global electricity demand from data centres reaching around 970 TWh by 2035.”
Quote checked on the page, 25 Oct 2026 - IEA, Energy and AI (2025), High Efficiency Case Forecast
“In this case, electricity demand from data centres is 20% lower in 2035 than in the Base Case.”
Quote checked on the page, 25 Oct 2026 - IEA, Energy and AI (2025), Headwinds Case (slower AI uptake) Forecast
“Similar to trends seen in the early 2010s, the improvements in efficiency are expected to offset most of the impact of increased IT stock utilisation. This leads to a plateau in energy demand at around 700 TWh, limiting the growth of the data centre share of global electricity demand to less than 2% in 2035.”
Quote checked on the page, 25 Oct 2026 - Luccioni, Strubell and Crawford, From Efficiency Gains to Rebound Effects (arXiv, 2025) Estimated
“Rebound effects undermine the assumption that improved technical efficiency alone will ensure net reductions in environmental harm.”
Quote checked on the page, 25 Oct 2026 - IEA, Key Questions on Energy and AI (2026) Estimated
“electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030”
Quote checked on the page, 3 Oct 2026 - Ember, yearly electricity data (CSV, CC BY 4.0) Data row
Data row: France, 2025, electricity demand 480.6 TWh. Demand includes grid losses, so it is close to, not the same as, consumption.
Row checked against the published file, 6 Oct 2026
Hum’s sums
- 970 TWh in 2035 ÷ 485 TWh in 2025 = 2, about double
Cite this
IEA, Key Questions on Energy and AI (2026), as quoted in Hum, “Will more efficient AI chips fix AI’s energy problem?”, https://humearth.org/q/efficiency-rebound (quote checked 25 Oct 2026).
Cite the original source first; Hum is where you found it.
What do Wh, kWh and TWh mean?
Wh is a watt-hour: a 1-watt light left on for an hour. 1 kWh = 1,000 Wh, and 1 TWh = 1 billion kWh. Hum compares each number with something everyday so you don’t need the units.
Send this answer
More efficient chips won’t fix AI’s energy use on their own: energy per AI task is falling fast, yet data-centre electricity grew 17% in 2025, and even the IEA’s High Efficiency case reaches about 970 TWh by 2035.