Eighteen companies cross four layers of this stack. 984 never leave one.

A data center is usually described as a stack: silicon at the top, then the memory beside it, the network between the racks, the rack, the water, the air, the switchgear, the generation behind the switchgear, and the building that holds all of it. Eleven layers, in the order this catalog uses. It is tempting to assume the companies work the same way — that a supplier picks a level and lives there.

Most of them do. That is the finding, and it is more lopsided than the framing suggests.

84% of suppliers appear on exactly one layer

Of the 1,163 companies in this catalog carrying at least one part placement, 984 — 84.6% — appear in exactly one of the eleven layers. Another 133 appear in two, 28 in three. Eighteen companies reach four or more.

The specialists are not a fringe. They hold three-fifths of every placement on the site. And they are more specialised than “one layer” makes them sound: 882 of the 1,163 hold exactly one part placement, full stop. The median company in this catalog makes one thing.

Where those single-layer companies sit is the second half of the picture. The two layers holding the most of them are the two that get built rather than shipped: 277 of them appear only in building shell, site and construction, and 198 only in power generation. Memory and storage, at the other end, holds 27. A layer made of poured concrete, regional trades and interconnection agreements is made of local firms who do that and nothing else. A layer made of DRAM is made of a handful of companies who do it everywhere.

The eighteen

Counted as distinct layers containing at least one part placement for that company.

Two of the eighteen are private, and the widest span on the list is one of them. Huawei appears on eleven parts across seven layers — storage, switch silicon, switch systems, coherent optics, CDUs, rear-door heat exchangers, CRAH units, UPS, inverters, its own cloud, and prefabricated modules. Nothing on this site prices that, because there is no share to price. Every market-cap and concentration view here is a view of the listed subset, and the company with the broadest reach across the stack is missing from all of them.

The pairs that occupy the same squares

Breadth alone is not the interesting part. Three pairs on that list sit on nearly or exactly the same parts, which is a sharper statement about competition than any table of revenue.

Dell and HPE hold the identical five placements — AI storage systems, switch systems and white-box ODMs, rack-scale systems, containerized AI factories and cluster management and GPU orchestration. Not similar. The same five rows, in the same five layers. Supermicro sits on all five plus GPU baseboards and compute trays. What the enterprise server vendors have converged on is not a product; it is a shape — take the accelerator someone else designed, wrap it in a rack, a container, storage and an orchestration layer, and sell the assembly.

Nexans and Prysmian also hold the identical five — fiber, connectors and structured cabling, LV cable, wire, tray and raceway, HVDC, advanced conductors and refining and products. Furukawa holds three of the five. These are the only companies on the list whose span crosses the data/power boundary in both directions, and the reason is that the boundary is a category we impose, not one the factory observes. Drawn copper is drawn copper whether it ends up carrying bits or amps, which is also why these three reach down into the materials layer — specifically copper refining — which nobody else on the list touches.

Schneider Electric and Vertiv are the near-miss, and the miss is the point. Both hold 22 placements across the same six layers. Fifteen of those placements are the same part. The seven each holds alone is where the two companies actually are:

• Schneider alone: HV switchgear, GIS, breakers and protection relays, MV/LV transformers, LV switchgear and switchboards, circuit protection, fuses, surge and grounding, plus relays, PLCs and in-rack busbar.
• Vertiv alone: direct-to-chip cold plates, rack manifolds, coolant distribution units, rear-door heat exchangers, plus dry coolers, power shelves and supercapacitor-based power smoothing.

Two companies that look interchangeable at the layer level diverge exactly along the line between protecting a building’s electrical supply and taking heat off a die. Schneider’s exclusive seven are almost entirely the electrical-protection stack. Vertiv’s are almost entirely the liquid cooling loop.

Breadth is not the same as generality

ABB reaches four layers, but thirteen of its sixteen placements are in power distribution and generation; the other three are fans, building controls and prefabricated power skids. It is a power company that steps sideways, not a generalist. Eaton is the same shape from the other end — 21 placements, thirteen of them in the electrical layer, with the rest reaching into rack enclosures and the cooling loop’s fittings and filtration.

Nvidia is the counter-example worth naming, because it is the company most often described as spanning everything. In this catalog it holds fourteen placements across four layers: silicon, fabric, the rack, and the software that manages the rack. It does not appear on a single part in memory and storage, liquid cooling, facility cooling, electrical distribution, power generation, the building, or materials — seven of the eleven layers, including the six that decide whether a site gets built. The company that defines the unit of compute stops at the rack door.

What this measures

Two limits, and they are load-bearing.

The first: this counts placements in a catalog, not revenue, contracts or share. A company reaches four layers here if the taxonomy behind this site put it on parts in four layers. A conglomerate with a poorly documented division will look narrower than it is, and none of these counts are weighted by how much of a part a company actually supplies.

The second is the same caveat that belongs on every table here. Of the 1,863 placements in this catalog, 1,847 are tagged inferred and 16 are tagged reported. None is disclosed. An inferred edge means the company operates in that category on the evidence of its own description of itself — not that it has been named as a supplier to a specific data center. So “Vertiv is on 22 parts” is a statement about the breadth of what Vertiv sells, not about what it has shipped to anyone. See sources and method for how the tags are assigned.

Read that way, the shape still holds, and it is a useful one. The AI data center is assembled from roughly a thousand companies that each make one thing, and knitted together by about eighteen that do not. When a name shows up on four layers at once, it is worth asking which of the four it would defend.