Research note · Physical infrastructure · August 8, 2026
The data center is not a chip.
The AI conversation usually begins with the chip because it is visible, expensive, and easy to compare. But a chip only produces revenue inside a system that can power, cool, house, connect, finance, and operate it. The bottleneck can move without the demand story changing at all.
Figure 1 · The capacity stack
A usable GPU-hour is the output of six constraints clearing together.
This is why “more GPU demand” and “more near-term data-center revenue” are related but not interchangeable claims.
Where the bottleneck moves
When chips are scarce, silicon pricing gets the attention. When chips arrive faster than grid interconnection, the bottleneck shifts to power. When power is available but construction is delayed, electrical equipment and contractors matter. Every stage can create a different winner, and every stage can delay the same end demand.
| If this constraint binds | What gets delayed | What I would watch |
|---|---|---|
| Interconnection | New energized capacity | Utility queue and contracted power |
| Cooling/electrical gear | Commissioning | Lead times and vendor backlog |
| Network | Cluster performance | Switching and optical spend |
| Financing | Provider-owned buildout | Prepayments, debt terms, and interest burden |
The investment mistake is treating a stack like a line
“Demand for AI” does not flow automatically and evenly from a model release to every company adjacent to a data center. A project can be power-constrained, financing-constrained, or simply late. That does not make the long-term narrative wrong. It makes the timing and the cash conversion much more complicated.
MapFollow the bottleneck, not the slogan
The on-site Data Center Alley map organizes the supply chain by constraint. It is a better starting point for a research question than assuming every AI-infrastructure company is exposed to the same driver.