NVIDIA reports Q1 FY2027 earnings tomorrow. The consensus sits at roughly $79B in revenue with a 7% options move priced in. Every analyst note and headline will be about whether the number clears the bar.

None of that matters for the three bottlenecks that will define compute infrastructure investment for the next three years.

These constraints are structurally decoupled from NVIDIA’s quarterly variance. They exist whether the beat is 3% or 8%. They exist because of a first-principles property of large-scale systems: value migrates upward as lower layers commoditise.

The GPU layer is the lower layer. These three sit above it.

1. The Grid Transformer: 128 Weeks and No Substitute

The electrical transformer that steps voltage from transmission lines to data center distribution levels has a procurement lead time of 80 to 128 weeks globally. This is a structural ceiling on how fast AI infrastructure can physically be built.

This constraint exists regardless of GPU supply. You can have every Blackwell GPU allocated and paid for. If the transformer cabinet is not bolted to a concrete pad with an energized feed from the grid, those GPUs are room-temperature silicon.

The market is slowly noticing. ABB, Siemens Energy, and Schneider Electric have rerated upward. But the real asymmetry sits two layers deeper.

2. The Optical Link: Data Movement, Not Compute

Inside every large AI training cluster, data moves between GPUs at terabit speeds. As clusters scale to 100,000+ GPUs, the distance between nodes forces a fundamental migration from electrical to optical interconnects.

NVIDIA spent approximately $4B securing supply from Lumentum and Coherent. The photonics supply chain is tight through at least 2027.

3. The Evaluation Layer: Trust as Infrastructure

As AI agents move from demos into production workflows, the binding constraint shifts from “can the model do this?” to “can we prove it did it correctly?”

2-3% of AI-generated code passes tests but contains subtle errors. Catching that minority requires evaluation infrastructure that most teams do not have.

This market barely exists today. No dominant platform for AI evaluation exists. That absence is itself the pattern: the layer below is commoditising, and the layer above is where value forms next.

Why These Three

Each bottleneck sits at a layer above the GPU in the infrastructure stack. The GPU is being efficiently priced by the market. The constraints above it are where the market’s attention has not yet reached.

NVIDIA will beat or miss tomorrow. By Thursday the financial media will have moved on. These three bottlenecks will still be tightening.

36-page NVIDIA earnings research report: Gumroad

Free weekly analysis: The Durability Curve