How much energy does a single AI query use?

A typical text query to Google's Gemini model uses about 0.24 watt-hours — roughly the energy of watching TV for 9 seconds (Google Environmental Report, 2024). Per-query energy has fallen roughly 33× over the past year as models, hardware, and serving infrastructure have become more efficient.

Is AI energy efficiency improving or getting worse?

Energy per AI task is dropping dramatically even as total AI demand rises. AI accelerator performance per watt has improved roughly 30× in recent generations, and software optimizations have cut per-query energy by another large factor. Aggregate demand is growing because more queries are being served, not because each query is less efficient.

Can AI itself reduce data center energy use?

Yes. DeepMind's AI-driven cooling control cut energy used for cooling at Google data centers by about 40% (DeepMind, 2016 and follow-ups). Similar machine-learning controls are now used by other hyperscalers for workload scheduling and HVAC optimization, with reported savings of 10–40% at deployed sites.

What are hyperscalers doing about water and clean power?

Meta has supported 40+ water restoration projects, returning over 1.59 billion gallons to stressed watersheds in 2024. AWS has improved Water Usage Effectiveness 40% since 2021 and aims to be water-positive by 2030. xAI's Colossus supercomputer in Memphis uses treated wastewater rather than potable municipal water. Amazon, Google, and Microsoft have all signed nuclear power purchase agreements (including SMR investments) to match growing demand with firm clean energy.

Topic · Efficiency

Efficiency & industry action

While total demand rises, energy per AI task is dropping dramatically — and hyperscalers are scaling new cooling, water stewardship, and clean power.

03 — Efficiency

AI energy efficiency: rapid progress

While total demand is rising, the energy required per AI task continues to drop dramatically thanks to better models, hardware, and software optimization.

0.24

Wh per query

Energy for a typical text query (e.g. Gemini) — roughly 9 seconds of TV watching.

Source · Google Environmental Report, 2024

30×

hardware gains

Performance-per-watt improvement across recent generations of AI accelerators (e.g. Google TPU).

Source · Google Cloud / MLPerf

33×

lower in 1 year

Drop in median prompt energy reported by Google over a single year of model and serving optimization.

Source · Google, 2024

Figure

Energy per AI query vs. everyday tasks

Source: Google Environmental Report, 2024. Typical estimates.

Figure

Hardware efficiency gains: AI accelerator generations

Performance-per-watt improvement relative to TPU v2 baseline. Source: Google Cloud / MLPerf.

Key takeaway

Energy per AI task is falling faster than overall demand is rising in many workloads — efficiency is doing real work alongside new capacity.

03b — Industry Action

What Tech Companies Are Doing

Major tech and AI companies are actively developing and deploying new technologies to reduce the water and energy demands of their data centers.

Microsoft

Zero-Water Cooling

In 2024, Microsoft introduced a new data center design using closed-loop, direct-to-chip cooling that consumes zero water for evaporative cooling. New facilities in Arizona and Wisconsin will pilot this approach starting in 2026.[1][2]

Google

Climate-Conscious Cooling

Google uses a science-based approach to choose between air cooling and water cooling depending on local conditions. The company aims to replenish more freshwater than it consumes by 2030 and has supported over 100 water stewardship projects across 68 watersheds.[1][2]

Amazon (AWS)

Reclaimed Water

AWS has improved its Water Usage Effectiveness by 40% since 2021. The company uses reclaimed wastewater at many sites and is working toward being water positive by 2030. They have also launched initiatives to use AI to help solve broader water challenges.[1][2]

xAI

Wastewater Cooling

xAI's Colossus supercomputer in Memphis uses treated wastewater instead of drinking water for cooling, avoiding the use of millions of gallons of municipal potable water per day.[1][2]

Energy Efficiency Improvements

Companies are rapidly adopting direct-to-chip and immersion cooling, which are far more energy-efficient than traditional air cooling.[1]
Newer AI chips and optimized software are delivering significantly better performance per watt.[1][2]
Hyperscalers are investing heavily in renewable energy and exploring nuclear power deals to meet growing demand sustainably.[1][2]
AI itself is being used to optimize cooling systems and workload scheduling, often reducing energy use by 10–40% in some facilities.[1]