The AI Bottleneck Cascade - From GPUs to the Grid

The AI buildout is not constrained by a single bottleneck. It is a cascade -- each resolved constraint reveals the next.

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

The Bottleneck Cascade

The AI infrastructure buildout follows a predictable pattern: GPU supply constraints resolve, only to reveal memory constraints, then packaging constraints, then power constraints, then grid constraints, then fuel constraints. Each layer looks like the last bottleneck until the next one emerges.

This is Law I (Bottleneck Migration) in operation. The investment question is not which layer is tight today -- it is which layer becomes the constraint next, and whether the market has already priced that migration.

Compute Layer

Three Signals the Market Is Missing on NVDA - T-5 earnings preview covering H200 China clearance, Culper short thesis, and photonics supply chain commitments.

Memory Layer

AI Hidden Tollbooth: HBM Memory - HBM sold out through 2026 at all three suppliers. Every AI GPU needs it. No substitute exists.

Optical / Interconnect Layer

Why NVIDIA Spent $500M on a Glass Company - Corning three new US optical factories, 10x capacity. The fiber optic layer of the bottleneck stack.

Power and Grid Layer

The Hidden Bottleneck: Switchgear - Transformer lead times of 80-128 weeks. ABB and Schneider control the electrical infrastructure.

Fuel/Energy Layer

The Nuclear SMR Bottleneck - HALEU fuel enriched to 5-20%. Centrus Energy holds the sole US DOE contract.

Subscribe

Get this analysis delivered weekly: Subscribe to The Durability Curve on Substack - free weekly research notes on the bottleneck cascade.

Subscribe: free weekly analysis on Substack | Products: AI Infrastructure Stack Bundle

Follow on Mastodon | Independent research. Not investment advice.

Part of the AI power grid bottleneck series. See the full AI power grid analysis hub.