AI's Power Problem: Three Grid Bottlenecks That GPU Earnings Won't Fix

The Durability Curve

Transformer lead times of 80-128 weeks. HALEU fuel production at 5% of projected 2030 demand. A grid interconnection queue growing faster than anyone can build. These are the real constraints on AI infrastructure — and none of them are about GPU supply.

Subscribe to The Durability Curve on Substack — free weekly analysis on AI infrastructure, verification, and durable systems.

The AI buildout story is usually told through chip names: H100, B200, Blackwell, Vera Rubin. But those chips need power to run, and the infrastructure that delivers that power is running into constraints that predate AI and will outlast any single hardware generation.

This is Bottleneck Migration (Law I) in action. The constraint has moved from GPU compute to memory, then to optical interconnects, and is now arriving at the grid and electrical equipment layer.

The Three Grid Bottlenecks

1. Electrical equipment (switchgear, transformers). Lead times for large power transformers have stretched to 80-128 weeks. Switchgear is constrained by GOES electrical steel supply. ABB and Schneider Electric control the market and are not adding capacity fast enough. Every AI data center needs this equipment. None of them can get it quickly.

2. Grid interconnection. The queue of power generation and storage projects waiting to connect to US transmission grids has grown to over 2,000 GW — more than the entire existing US generation fleet. Data centers must wait 3-5 years for interconnection studies. Some utilities have stopped accepting new interconnection requests entirely.

3. Fuel supply for next-gen power. Small modular reactors (SMRs) are the most-promised solution for data center power. But the fuel they require — HALEU — is not being produced at scale. Current US capacity covers roughly 5% of projected 2030 demand. No commercial HALEU production facility exists outside Russia.

Our Analysis

Each of these articles covers one dimension of the power bottleneck:

The Hidden Bottleneck: Switchgear — 51 views. The most constrained component in AI data centers is not a GPU. It is electrical switchgear, and two companies control it.

The Nuclear SMR Bottleneck: HALEU Fuel — 42 views. SMRs are the promised solution for data center power, but the fuel supply chain does not exist yet.

The AI Bottleneck Cascade — The structural framework for understanding how each resolved constraint reveals the next.

Falsification Framework

The grid bottleneck thesis is intact if: transformer lead times remain above 40 weeks, GOES electrical steel supply stays constrained, grid interconnection queues continue growing, and HALEU fuel production shows no commercial-scale commitments.

The thesis is breaking if: transformer manufacturing capacity doubles within 12 months, a substitute for GOES electrical steel emerges at scale, grid interconnection reform reduces timelines to under 18 months, or SMR developers switch to low-enriched uranium fuel that is already commercially available.

Falsification trigger: If ABB or Schneider Electric announce a major transformer production expansion (50%+ capacity increase), the tightest bottleneck in the electrical layer begins to ease. Track this through earnings calls and capacity announcements.