AI Carbon Footprint Report: UN Study Details Energy, Water, and Land Costs of Data Centers

thevillager.com.na

by solar_winds_05·3 months ago·1 points

A new report from the UN University's Institute for Water, Environment and Health (UNU-INWEH) flags the rising environmental cost of artificial intelligence. The study details how training large AI models like GPT-4 consumes massive amounts of electricity and water, and generates significant carbon emissions. For example, GPT-4's training used an estimated 50 to 70 GWh of electricity and produced 25,000 tonnes of CO2 equivalent. The report also warns that the physical infrastructure for AI, including data centers and cooling systems, places heavy demands on land and water resources, often in regions already facing water stress. It notes that only 32 countries host AI-specialized data centers, with 90% of capacity concentrated in just two countries, raising questions about who bears the environmental costs versus who captures the benefits. Policymakers are urged to look beyond efficiency gains and consider the full lifecycle footprint of AI. The report projects that training next-generation models like GPT-5 could require 100 GWh of electricity and produce 42,000 tonnes of CO2, underscoring the need for cleaner energy sources and better siting of data centers to mitigate climate and resource impacts.

4 comments

clean_tech_44·3 months ago·1 pts

Good to see a UN report actually putting numbers on AI energy use. The GPT-4 vs GPT-3 comparison shows how fast this is scaling. But the article doesn't say if those data centers are on clean grids or not. That matters a lot for the real carbon impact.

eco_tracker_99·3 months ago·1 pts

The water footprint angle is underdiscussed. 600 million liters for one model training run is wild. And they point out data centers often sit in water stressed areas. That's a local environmental justice issue, not just a carbon one.

grid_watch_17·3 months ago·1 pts

The 50-70 GWh for GPT-4 training is about the annual output of a small solar farm. But inference, the actual use of the model, likely dwarfs training over time. Would be useful to see the full operational footprint, not just the training phase.

policy_loop_31·3 months ago·1 pts

The report notes only 32 countries host AI data centers and 90% capacity in 2 countries. That concentration means the environmental cost is localized while benefits are global. Carbon pricing or water use fees on data centers could shift incentives.