The Power Constraint: Electricity as the Next Limiter for AI
Electricity is becoming the next physical constraint on AI deployment. A single NVIDIA B200 GPU rack consumes 120kW — a modern AI data center with 100,000 GPUs needs over 1 gigawatt of power, equivalent to a nuclear power plant. Hyperscalers are running out of available power capacity faster than new generation can come online. This creates investment opportunities in power generation (nuclear SMRs, natural gas), power delivery (transformers, switchgear), and power efficiency (liquid cooling, thermal management). The energy constraint is tracked through the Energy layer of Closelook's 6-Layer Model.
The Scale of the Problem
Global data center electricity consumption is projected to more than double by 2028, driven almost entirely by AI workloads. The challenge: building power generation and transmission infrastructure takes 3-7 years, while AI demand is growing quarterly. This mismatch creates a structural power deficit that constrains where and how fast AI infrastructure can be deployed.
Several hyperscalers have already reported delaying data center deployments due to power availability. Microsoft, Google, and Amazon are all investing in long-term power contracts, nuclear partnerships, and on-site generation — signaling that they view power as the binding constraint on their AI ambitions.
Investment Implications
Power generation: Nuclear (Constellation Energy, NuScale, Oklo), natural gas (Calpine, Vistra), and utility-scale solar/battery storage all benefit from hyperscaler demand for reliable baseload power.
Power delivery: Transformers and switchgear are bottlenecked — Eaton, Schneider Electric, and niche players see extended order backlogs.
Power efficiency: Liquid cooling reduces power wasted as heat. Every watt saved on cooling is a watt available for compute. Vertiv and Modine are the primary beneficiaries.