AI Electricity Scale

A single ChatGPT query consumes roughly 10x the electricity of a Google search. Multiply by billions of queries, add training runs that consume megawatts for months, and AI becomes the largest new source of US electricity demand since industrialization.

US data centers consumed about 176 TWh of electricity in 2023, roughly 20 GW of average demand. DOE and Lawrence Berkeley National Laboratory project that could reach 325-580 TWh (37-66 GW average) by 2028, driven almost entirely by AI workloads. For scale: 20 GW is roughly New York State's average electricity demand. That growth within five years means building the equivalent of dozens of large power plants, plus the transmission to connect them.

Size the training demand. Training a frontier AI model like GPT-4 required an estimated 50-100 GWh of electricity over several months. Next-generation models may require 10x more. A single training cluster might draw 100-300 MW continuously, comparable to a small city.

Can renewable energy alone power this growth?

The reliability requirement. Data centers require 99.999% uptime ("five nines"). Solar delivers an average of about 25% of its rated capacity across a year; wind about 35%. AI workloads run 24/7 and cannot tolerate interruption. This means data centers need firm, dispatchable power: gas, nuclear, or renewables with massive storage. Several tech companies have signed nuclear power agreements precisely because nuclear matches the 24/7 load profile.

AI electricity demand is not a future problem. It is reshaping utility planning, generation investment, and grid expansion decisions today.

---