Microsoft Study: AI Energy and Water Use Lower Than Expected

Microsoft published a blog post citing a study in Joule that shows AI's energy and water use in large-scale deployment is much lower than previously thought. The study says the focus should be on the 'per query' cost for typical user calls, not just a single model's training or inference size.

A typical AI call uses 0.16 to 0.60 watt-hours of electricity, equal to a 40-watt computer running for 15 to 60 seconds or a 1000-watt microwave working for 0.6 to 2 seconds. This is 1/4 to 1/20 of earlier reports. For water, a typical query uses 0.0 to 0.067 milliliters for cooling, about 1/100 teaspoon, less than a drop.

Microsoft notes that larger systems are more efficient per query, handling more requests at once and optimizing across models, scheduling, and hardware. In a scenario with 1 billion daily queries, baseline electricity use is about 0.7 GWh, which can drop to 0.3 GWh with optimization. Even if 10% of requests are long tasks like code generation, total energy can still be cut by more than half.