The AI boom has an energy problem. Data centers training and running large language models are consuming electricity at a scale that's beginning to stress power grids, raise utility costs for ordinary consumers, and complicate corporate sustainability commitments. The numbers, when laid out plainly, are sobering.
The Scale of the Problem
Training a single large frontier model like GPT-4 consumed an estimated several gigawatt-hours of electricity — roughly the annual consumption of thousands of households. Inference is even more significant in aggregate: millions of ChatGPT queries run daily, each requiring meaningful compute. The International Energy Agency projects that data center electricity consumption could double by 2026 compared to 2022 levels, driven substantially by AI workloads. Some hyperscale facilities are now negotiating directly with power utilities for dedicated capacity in the hundreds of megawatts.
Nuclear Power Enters the Equation
The most striking response to AI's energy appetite has been a renaissance of interest in nuclear power. Microsoft signed a deal to restart a unit at Three Mile Island specifically to power its data centers. Google and Amazon have made similar commitments to advanced nuclear technologies. The appeal is clear: nuclear provides dense, reliable, carbon-free power that can be sited near data centers. Small modular reactors (SMRs), still in development, promise even more flexible deployment options for future hyperscale facilities.
The Sustainability Contradiction
This creates an uncomfortable tension for AI companies that have made public climate commitments. Google, Microsoft, and others have pledged carbon neutrality, but their Scope 2 emissions from purchased electricity have risen sharply as AI workloads scale. Renewable energy purchases and carbon offsets help, but they don't change the physical reality that AI is a major and growing driver of energy demand. The honest answer is that the industry is betting that nuclear and renewables will scale fast enough to keep up — a bet that isn't guaranteed to pay off.
Efficiency as the Partial Answer
Model efficiency has improved dramatically. Today's frontier models achieve the same or better performance with far less compute than models from two years ago, thanks to better training techniques, architecture improvements, and hardware optimizations. Smaller, more efficient models like Gemini Flash-Lite and Llama 3.3 8B handle many tasks at a fraction of the energy cost of their larger siblings. But efficiency gains have historically been absorbed by increased usage rather than reducing total consumption — a pattern known as Jevons' paradox that AI is unlikely to escape.