GPT-4 water consumption study raises environmental concerns
A University of California study found that generating 100 words with GPT-4 consumes 1,408 litres of water due to the energy demands of data centre cooling. The research highlights significant environmental costs of large-scale AI deployment, with potential electricity consumption equivalent to residential usage across entire cities if usage scales further.
What happened
A University of California study published by The Washington Post found that generating 100 words with GPT-4 consumes 1,408 litres of water, equivalent to three half-litre bottles.
The high water consumption results from the substantial energy demand required by data centres to process AI models and cool servers, with usage levels dependent on data centre location and cooling methods used.
Reported impact
If 10% of Americans used GPT-4 once per week, the electricity consumed would equal the residential electricity use of Washington D.C. over 20 days.
The environmental cost is magnified by the millions of daily queries ChatGPT receives, many of them trivial.
Index analysis
Processing large language models like GPT requires thousands of specialized chips, particularly Nvidia hardware, which generate substantial heat requiring efficient cooling systems that use large quantities of water.
In regions where electricity is cheaper or water is scarce, data centre operators may use electricity instead of water for cooling, making consumption levels location-dependent.
Governance controls
Representatives of OpenAI, Google, Microsoft and Meta have stated their commitment to reducing water consumption at their data centres, but have not presented specific solutions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.