The energy and water footprint of generative AI has gone from a curiosity in 2023 to a board-level concern in 2026. With ChatGPT, Claude, Gemini and Copilot embedded in everyday business workflows, every UK business is now indirectly an AI energy consumer. This guide breaks down what we actually know about ChatGPT’s energy and water use and what it means for UK business energy demand.
How much energy does a ChatGPT query use?
Estimates from peer-reviewed studies and OpenAI’s own disclosures put a single ChatGPT query at 0.2-0.5 Wh depending on:
- Model size (GPT-4o vs smaller GPT-4o-mini).
- Query complexity (simple text vs reasoning vs image generation).
- Token count (input + output).
For comparison:
- Google search: ~0.03 Wh.
- YouTube minute (HD): ~0.5 Wh.
- ChatGPT query (text): ~0.3 Wh.
- ChatGPT image generation: 2-5 Wh.
- 1 hour Netflix HD: 100-200 Wh.
How much water does ChatGPT use?
AI data centres use water for evaporative cooling. Recent academic work estimates roughly 0.32ml per ChatGPT query on the data-centre side, plus another 0.5-2ml on the upstream electricity generation side (cooling water for thermal power plants).
For 1 billion daily queries that totals around 120 megalitres/year of direct cooling water and 200-700 megalitres/year of upstream water — roughly the annual water use of 8,000 UK households.
What about training a model like GPT-4?
Training the GPT-4 family is estimated to have used around 50 GWh of electricity over several months — roughly the annual consumption of 14,000 UK homes. Inference (serving queries to users) is now the dominant energy cost, exceeding training within ~6 months of model deployment.
Why AI is the biggest driver of UK electricity demand growth in 2026
The UK National Grid ESO 2024 Future Energy Scenarios projected data-centre electricity demand to roughly quadruple by 2030, driven primarily by AI workloads. UK data-centre demand was ~3% of national electricity in 2023; it is forecast to reach 8-12% by 2030.
This is reshaping UK grid investment plans, accelerating both renewable build-out and gas-fired backup, and putting pressure on connection queues for new data-centre developments in West London and the M4 corridor.
What can a business do to reduce its AI energy footprint?
- Use smaller models for simple queries (GPT-4o-mini vs GPT-4o where appropriate).
- Cache and reuse responses for repetitive queries.
- Avoid generating images and video unless necessary — image and video generation are 10-100x more energy-intensive than text.
- Choose providers powering data centres with renewables (Google, Microsoft and AWS all market significant renewables in their AI services in 2026).
- Track your team’s API spend as a proxy for energy footprint.
Frequently Asked Questions
About 0.3 watt-hours per text query on GPT-4o, roughly 10x a Google search. Image generation queries use 2-5 Wh and reasoning-heavy queries use 0.5-2 Wh. The numbers vary by model size and query complexity.
Estimates put direct data-centre cooling at 0.32ml per query, plus 0.5-2ml of upstream water at the electricity generation stage. At 1 billion daily queries that totals 320-700 megalitres/year of total water footprint.
Estimated around 50 GWh of electricity over several months — roughly the annual consumption of 14,000 UK homes. Inference (serving queries) now exceeds training as the dominant energy cost.
The UK National Grid ESO projects data-centre demand to quadruple by 2030 driven by AI. This is a meaningful driver of UK demand growth but by itself is unlikely to drive a sharp price rise — renewable build-out is roughly tracking the new demand. Bigger price drivers remain wholesale gas and policy costs.
Use smaller models where possible, cache repeated queries, avoid unnecessary image and video generation, and prefer providers with strong renewable energy backing for their data centres.
Want a benchmark of your business’s total energy and AI footprint? Get a free 60-second business energy review — we can include AI / data-centre exposure in the analysis.
Related Reading
More from the Connection Technologies blog.