The Energy Bottleneck: Why Power Is the Real Constraint on AI Growth
AI data centers need 10x more power than traditional compute. Nuclear and natural gas are the only solutions at scale.
The AI industry's dirtiest secret: there isn't enough electricity to power the future of artificial intelligence. Not even close.
The Scale of the Problem
A single NVIDIA H100 GPU consumes 700W. A training cluster with 10,000 GPUs needs 7MW just for the chips — add cooling and networking and you're at 15-20MW per cluster. A hyperscale AI data center needs 100-500MW.
For context, 500MW is enough to power a city of 400,000 people.
Why Renewables Aren't Enough
Solar and wind are intermittent. AI training runs 24/7 for weeks or months. You can't pause a $100M training run because the wind stopped. AI needs baseload power — consistent, reliable, always-on electricity.
The Nuclear Renaissance
Constellation Energy's deal with Microsoft to restart Three Mile Island is the canary in the coal mine. Nuclear provides carbon-free baseload power, and the AI industry is willing to pay premium prices for it.
Key beneficiaries: Constellation Energy (CEG), Vistra (VST), and NRG Energy (NRG).
Natural Gas Bridge
While new nuclear takes 10+ years to build, natural gas is the immediate solution. GE Vernova (GEV) and Quanta Services (PWR) are building the gas turbines and grid infrastructure needed today.
Investment Implications
Energy infrastructure companies are trading at a fraction of AI software valuations, despite being the fundamental bottleneck. This is the most underappreciated layer of the AI supply chain.