The five AI data center parts this catalog flags as tight or constrained

This catalog carries 130 top-level part categories across 11 layers. Exactly five of them carry a supply flag — two marked tight, three marked constrained. Everything else is blank, which in this catalog means no sourced constraint is on file, not that supply is comfortable.

The interesting part is where the five sit. Only one is silicon. The other four are power, metal and turbines — the parts of an AI data center that have nothing to do with the rack.

The one in the rack

HBM is the only flagged part inside the machine, and it is the one flagged hardest. The catalog’s note on it is one sentence: the most supply-constrained component in the stack. The reason it stays that way is structural rather than cyclical — each accelerator generation asks for more stacks and more capacity per stack as models get larger and context windows get longer, so demand moves with model architecture, not with the ordinary memory cycle.

Passives — the MLCCs and inductors scattered across every board — is the second in-rack entry, flagged tight. It is a volume problem, not a physics one: an AI board carries roughly 10–15× the passive count of a standard server, and high-spec MLCC lead times were running around 24 weeks in mid-2026 trade press. Twelve manufacturers sit on that part here, all of them publicly listed, which is a much wider bench than HBM has.

The four outside the building

The other three flagged parts, plus the fourth flag, form one chain — and it is a chain about electricity, not compute.

Large power transformers is flagged constrained with the bluntest number in the whole catalog: lead times of 2–4 years, the single biggest schedule risk for GW campuses. That is longer than it takes to build the hall the transformer feeds. Eighteen manufacturers are on file for it — a wide bench that still has not translated into short lead times, because the constraint sits downstream of the nameplate on the transformer.

Downstream means grain-oriented electrical steel and amorphous cores, flagged tight, with the catalog’s own explanation of the link: GOES supply is the input constraint behind transformer lead times. You cannot order your way out of a transformer queue by adding transformer factories if the laminated core steel is the thing in short supply. Nine suppliers are on file, and they are steelmakers — ArcelorMittal, Nippon Steel, POSCO, Cleveland-Cliffs, thyssenkrupp — not electrical equipment firms.

Which leaves gas turbines and HRSGs, flagged constrained. These are what a campus reaches for when the grid interconnection queue is longer than the build schedule, and the catalog notes the obvious consequence: new-build turbine slots are sold out years ahead; refurbished and aeroderivative units command premiums. The escape hatch from the transformer queue has its own queue.

What this actually tells you

Read the five together and the shape is hard to miss. The scarce thing in an AI data center in 2026 is not, on this evidence, the accelerator. It is the ability to get power to the accelerator: a transformer that takes two to four years, core steel that gates the transformer, and onsite generation that is sold out as the alternative. HBM is the one piece of silicon in that company.

Two honest caveats, because this site runs on them.

First, absence of a flag is not evidence of comfort. 229 of 234 parts carry no supply status, and that means only that nobody has filed a sourced constraint against them here. It is a record of what has been checked, not a clean bill of health.

Second, every supplier placement named above is tagged inferred. In this catalog that is the weakest of three tags — it means the placement comes from the taxonomy without a citation attached yet, so it is a lead to verify rather than a disclosure to quote. The supply flags and the lead-time figures come from the taxonomy’s sourced notes; the lists of who makes each part are still leads. Those are different claims with different weight, and the tag on every row tells you which you are reading. See sources & method for how the three tags are assigned, and bottlenecks for the live version of this list — it changes when the underlying data does, and this post does not.