Segment
AI compute: the silicon layer, and what it is waiting on
Nearly all of the money spent building AI capability ends up here, in a few classes of chip and the machines built around them. This branch breaks that layer into the parts that are actually bought separately, and shows which of them the rest of the stack is waiting on.
In one sentence
AI compute is the hardware layer that executes neural-network arithmetic: the accelerators, the memory co-packaged with them, and the server systems that turn both into something a data centre can rack.
It is tempting to treat AI compute as one purchase, because it is often quoted as one price. It is not. An accelerator is a die from one supply chain, memory stacks from a second, an interposer and packaging step from a third, and a system integration from a fourth — and at various times each of those four has been the limiting factor rather than the die itself.
Separating them matters because they behave differently. Wafer capacity can be booked years ahead and is expensive to add. Packaging capacity has been added faster but from a much smaller base. Memory is made by three companies. System integration is competitive and adds relatively little scarcity. A shortage narrative that does not say which of these it means is not saying much.
The other reason to separate them is that they are diverging. Training and serving now pull the specification in different directions, and the parts of the industry serving each are starting to look like different businesses.
How this breaks down
Split by what is bought and manufactured separately, rather than by what arrives in one box.
- AI acceleratorsThe processors that run training and inference, and why there are three kinds of them.Breaks down into: Merchant GPUs · Custom AI ASICs · Inference silicon3 pages belowChokepoint
- AI memoryBandwidth, capacity and the storage under them — the part of the system that is usually the bottleneck.Breaks down into: DRAM manufacturing · NAND and SSDs · High-bandwidth memory · Server memory · AI storage5 pages belowChokepoint
- AI server systemsThe integration step — power delivery, cooling, backplanes — that decides how many chips can share one problem.Breaks down into: Compute trays · Rack power delivery · Rack integration3 pages belowChokepoint
What this depends on
3 of these are marked as a chokepoint: a handful of qualified suppliers, a multi-year lead time, or a single geography.
- Supply chainChokepoint
Leading-edge wafers
Every part in this branch is manufactured on the most advanced nodes available.
Wafer fabrication - Supply chainChokepoint
Advanced packaging capacity
Die and memory have to be joined before there is a product. This has been the binding constraint more often than wafer supply.
Advanced packaging - Supply chainChokepoint
Data-centre power and cooling
A rack of these parts draws several times what a conventional rack draws and cannot be air-cooled at density. Compute that cannot be powered or cooled cannot be deployed.
Data centres - Supply chain
Outsourced assembly and test
Packages, memory modules and boards still have to be assembled onto substrates and electrically tested at volume. That step is bought from a small set of assembly and test houses, and a shortage there holds finished parts as surely as a wafer shortage does.
OSAT - Technology
The software stack that drives it
Compute nobody can program is inventory. Every part of this layer is chosen for the compilers, kernels and schedulers written against it as much as for its specification.
Software stack
What depends on this
Other pages in this map that name AI compute as something they cannot do without.
- Robotics and automation · Humanoids and embodied AIEmbodied learningTraining these policies is a conventional large-scale training job on the same hardware as any other.
- Artificial intelligence · Model architectureScaling and pre-trainingThe run is defined by how many accelerator-hours can be kept working in parallel.
- Robotics and automation · Control and softwareSimulation and digital twinsRunning many simulated environments in parallel is an accelerator workload, and it is what makes the method practical.
Companies across AI compute
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.
97 more companies appear at a single step each; they are named on the pages for those steps.
How these pages are written
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