The further you get from the chip, the less of the supply chain is listed

Almost every attempt to get exposure to AI data centers starts from the same place: the accelerator, then the memory stacked next to it, then the rack the two arrive in. That is where the listed names are, and it is not an accident of attention. It is what the supply chain actually looks like when you sort it by whether you can buy a share of it.

This catalog holds 1,164 companies that sit on at least one part. 652 of them — 56% — have a primary stock-market listing. The other 512 are private, subsidiaries of something else, or state-owned and cooperative bodies with no equity to buy. That average hides the interesting part, which is how unevenly the 56% is spread across the eleven layers.

The listed share, layer by layer

Counted as distinct companies carrying at least one part edge in that layer. A company that appears in several layers is counted in each.

The ordering is close to a physical one. The five layers above 70% are the ones that ship in a crate: silicon, the memory beside it, the optics between racks, the rack itself, and the refined materials feeding all of them. Those are global product businesses with long public histories, and they are the layers a screen of AI hardware names will already contain.

The four layers at the bottom are the ones that get poured, bolted and welded on site. Building shell and construction is the largest layer in the catalog by supplier count — 323 companies, more than silicon and memory together — and it is the least investable, at 41%. Facility cooling sits at 48%. Power distribution between the grid and the rack sits at 52%.

Where the money is not

Four parts with five or more suppliers have no listed supplier at all:

The first three are real findings. Startup accelerators is a tautology worth stating plainly — Cerebras, Groq, Etched, Tenstorrent and the rest are in the catalog precisely because they are the non-listed alternative to the merchant GPU, and the day one of them lists it stops being that. Fusion is the same shape at a longer horizon. Immersion cooling systems is the one that should give a reader pause: nine independents — LiquidStack, Submer, Iceotope, GRC, Asperitas and others — supplying a cooling approach that hyperscalers keep piloting, with no public-market way to express a view on whether it wins.

The fourth is not a finding, it is a gap in this catalog, and it is more useful to say so than to publish the number. The part is named “transmission, interconnection and grid equipment”, but the nine rows on it are all system operators — PJM, ERCOT, CAISO, MISO, NYISO, ISO-NE, SPP, AESO, IESO. Those are the bodies that run the interconnection queue, and they are correctly there. The equipment half of that part — the conductor, the HVDC converter, the interconnection hardware — has no suppliers attached yet. A zero on that row means the catalog has not filled it in, not that nobody makes the equipment. It is on the open list.

The part that reads as a warning

The three construction and trade parts are worth reading together, because they are where the listed share collapses without the category being young or pre-revenue:

• MEP & electrical contractors: 5 listed of 25
• Architecture, engineering & MEP design: 8 listed of 30
• General contractors & design-build: 20 listed of 60

These are old, large, profitable businesses. Rosendin, Faith Technologies, Cupertino Electric and Bergelectric are among the firms that physically wire data centers, and they are private, employee- owned or held inside a larger group. The listed way to touch that work is a short list — EMCOR, Comfort Systems, IES Holdings, Limbach, Southland — which is why those names get bid on data center headlines far out of proportion to their share of the actual work. The supply of listed exposure to electrical contracting is much smaller than the supply of electrical contracting.

What this is and is not

Two limits, both of which matter more than the table.

The first is that ownership type is a fact this catalog holds for all 1,199 companies, and it is the reason this particular cut is publishable. Headquarters country is populated for only 97 of them, which is why there is no map in this post. A layer-by-country version of this table would look authoritative and describe 8% of the data.

The second is that a part edge is not a purchase order. Almost every supplier placement in this catalog is tagged inferred — the company operates in that category, on the evidence of its own description of itself. Sixteen edges are tagged reported. None yet carries a source URL to a contract. So this table says where listed companies are in the supply chain, not what share of any part’s revenue they hold. Read it as a map of where public-market exposure exists and where it does not, which is a different and more modest claim than a market-share table.

The share of a layer that is listed is not a measure of how important the layer is. It is closer to the opposite: the least listed layer in this catalog is the one with the most companies in it, doing the work that takes the longest and is hardest to hurry.