Sub-system
An accelerator is not deployable on its own. Someone has to put eight of them in a chassis, deliver tens of kilowatts to it, take the heat back out, and wire them to each other and to the network. That integration has become a design problem in its own right rather than a packaging exercise.
In one sentence
An AI server system is the complete machine — accelerators, host processors, memory, interconnect, power delivery and cooling — assembled and validated as one unit, increasingly designed at the scale of a whole rack rather than a single chassis.
The unit of design has been moving upward. A single server, then a chassis of eight accelerators, and now a full rack wired so that dozens of accelerators behave as one large machine. Each step upward exists because models grew past what the previous unit could hold in fast memory.
That shift changes who does the hard engineering. When the unit was a server, the value was in the chip and the box was commodity. When the unit is a rack with a shared power bus, a liquid loop and a switched backplane, the integration is where a large part of the difficulty and the cost sits.
Split by what one supplier actually delivers — the boards, the power path inside the rack, or the finished machine.
Racks of this kind draw far more than conventional IT racks — enough that distributing power as direct current across a rack-wide busbar, rather than to each chassis separately, becomes the sensible engineering. That in turn changes the power shelves, the busbar, the protection and what the data centre has to bring to the rack.
Past roughly thirty to forty kilowatts a rack, air stops being practical: the airflow required is loud, inefficient and eventually impossible. Cold plates on the hot components, fed by a coolant loop and a distribution unit, are what allow the density these systems assume — which makes liquid cooling a prerequisite rather than an upgrade.
If a model's fast-memory footprint exceeds one chassis, the accelerators must be joined into a single memory domain over a high-bandwidth link. The reach of that link is short, so the domain's size is bounded by physical layout. Designing the rack and the interconnect together is what makes the domain as large as it can be.
3 of these are marked as a chokepoint: a handful of qualified suppliers, a multi-year lead time, or a single geography.
Cold plates, distribution units, quick-disconnects and manifolds. A shortage in any of them stops a deployment as effectively as a chip shortage.
Direct-to-chip liquidRack-scale direct-current distribution is specialised hardware with a narrow supplier base relative to the volumes now being deployed.
Busway and rack distributionThe backplane that joins accelerators into one memory domain is what makes a rack more than a stack of servers.
Scale-up fabricsShared mechanical, power and management specifications are what let several integrators build compatible racks rather than each operator owning a bespoke design.
The system is built around baseboards of bought accelerators. Allocation of those modules, not chassis manufacturing, is what decides how many systems can be shipped in a quarter.
Merchant GPUsThe other half of rack-scale demand is systems built around an operator's own silicon. The mechanical, power and cooling work is the same; only the compute module changes.
Custom AI ASICsEach node carries conventional DRAM behind the accelerators for staging data and holding the operating system. It is a large line on the bill of materials and moves with the commodity memory cycle.
Server memoryA rack-scale system is thousands of terminated conductors. Harness lead times have appeared on the critical path for deliveries more than once.
Cables and connectorsSomebody has to assemble, wire, fill, leak-test and burn in each rack before it ships, then move a two-tonne object into a building. That floor space and labour is finite and sits with a small number of contract manufacturers.
Other pages in this map that name AI server systems as something they cannot do without.
Every company named on a step below this page, ordered by how many of those steps it appears at. Compiled from the pages themselves rather than written separately, so the two cannot disagree. Not a ranking and not a recommendation.
17 more companies appear at a single step each; they are named on the pages for those steps.
Whether rack-scale designs stay proprietary to a small number of integrators or standardise enough for broad competition.
Whether liquid cooling components, rather than chips, become the visible lead-time constraint on deployments.
How far rack power draw rises before facility design rather than server design becomes the limit.
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