Technology
Once the trays, switches, shelves and manifolds exist, somebody has to build them into a rack, connect several thousand terminations, fill the coolant loop, prove it does not leak and run the whole thing at full power for days. That step is a manufacturing business of its own, and it has been the visible constraint more than once.
In one sentence
Rack integration is the final assembly stage in which compute trays, switches, power shelves, busbars, cabling and a liquid cooling loop are built into a rack, wired, filled, leak-tested and burned in as a working machine before shipment.
The industry describes assembly in levels: boards, then chassis, then the fully populated and tested rack. As the unit of design moved up to the rack, the last of those levels grew from a packing operation into the step where the machine is actually made. It is where the cooling loop is closed, where every link is trained and verified, and where a fault is far cheaper to find than it is on a customer's floor.
It is also a physical business with unglamorous limits. Integration floor space has to be near enough to the customer to ship a filled two-tonne rack, the factory needs power and cooling of the same order as a small data centre to burn in what it builds, and the labour is trained rather than generic. Those constraints do not scale on the same clock as chip supply, and they sit with a small number of contract manufacturers.
A rack of this kind contains thousands of terminated conductors, dozens of blind-mate fluid connections and a coolant loop that must be filled, purged and proved leak-free with electronics already installed. Each of those is a process with an inspection step behind it, and the cost of getting one wrong is measured in the value of the hardware around it.
Burning in a rack means running it at close to full power long enough to expose infant failures, which means the integration site must be able to supply and reject that heat for every rack on the floor simultaneously. Building integration capacity is therefore closer to building a small data centre than to fitting out an assembly hall, and it is why capacity cannot be added quickly.
Filled racks are heavy, fragile and awkward to ship, so integration tends to migrate towards the regions being built out. Trade policy and customer preference push in the same direction. The result is that the same companies operate the same process in several countries rather than concentrating it in one.
4 of these are marked as a chokepoint: a handful of qualified suppliers, a multi-year lead time, or a single geography.
The populated boards arrive as an input to this step, and their availability is usually what sets the schedule.
Compute traysThe power path is installed and commissioned as part of integration, and a shelf shortage stops a rack as effectively as a missing accelerator.
Rack power deliveryCold plates, manifolds, quick-disconnects and distribution units are fitted, filled and tested here. They have repeatedly been the part the schedule waited on.
Direct-to-chip liquidThousands of high-speed and power terminations are made at this step, and harness lead times have appeared on the critical path for deliveries.
Cables and connectorsVerifying that every link in the rack trains at rate is part of the test, and a rack that passes electrically but not at link rate is not shippable.
Scale-up fabricsSpace near the customer, enough power and cooling to run everything on the floor at once, and people trained on fluid and high-speed assembly. None of it can be added at the pace demand has moved.
What each company supplies at this step, and — where a public figure exists — its share of this specific market — with what that share measures, the period it covers and who published it. Some rows also show the company’s own reported revenue for the segment covering this step, which is a different thing: it says how much this business matters to that company, not how much of the market it holds. Not a ranking and not a recommendation.
Integrates more AI racks than anyone else, largely under other companies' names.
Builds and tests rack-scale systems for hyperscale operators.
Integrates rack-scale systems for operators designing their own hardware.
Integrates and tests its own liquid-cooled rack designs at its own sites.
Integrates and validates full AI rack deployments for enterprises and operators.
Builds and commissions liquid-cooled rack and cluster deployments.
Provides rack integration and test capacity under contract.
Integrates rack-scale compute and networking systems for operators.
Integrates liquid-cooled AI racks and ships them as complete systems.
Specifies the rack rather than the module as the product, which is what moved the integration work up to this level.
Bought a rack-scale systems builder so it can ship accelerators as finished racks rather than as parts.
Integrates rack-scale systems under contract for brand vendors and operators.
Supplies the power and liquid-cooling equipment built into the rack, and increasingly commissions it on site.
Integrates most of the AI racks deployed inside China, where the Taiwanese integrators cannot ship.
Whether integration and burn-in floor space, rather than components, is named as the reason a deployment slipped.
How much integration capacity moves to the regions being built out, and what that does to unit cost.
Whether operators take integration in-house at their own sites as deployment volumes rise.
Each page explains one technology in plain language, states what it depends on, and names companies by what they supply at that step. Company roles are described qualitatively and deliberately carry no market shares, revenue figures or rankings — those change faster than an explainer can, and a stale number is worse than none. Ticker links point at company pages on this site and are provided for reference only.
Nothing here is investment advice, a recommendation, or a forecast. A company named on a page about a technology is not thereby a good investment, and the chokepoints described are structural facts about supply chains rather than predictions about prices. Technology moves; where a page describes something as unresolved or in development, that was true when it was written.
Plutux is not an investment adviser. Market data and AI-generated analysis are for information and education only, not investment advice. Disclaimer