Industry
Artificial intelligence: the stack, layer by layer
Almost every claim made about AI is a claim about one layer of a stack that has six of them. This map separates those layers, explains each in plain language, and states what each one depends on — including the dependencies that sit outside the industry entirely, in semiconductor packaging and in grid power.
Artificial intelligence, as an industry, is a stack of six layers that are bought, built and constrained separately. At the bottom is silicon: accelerators and the memory beside them. Above that, the networks that join thousands of chips into one machine. Above that, the models themselves — the architecture, the training recipe, the post-training that makes them usable. Then the data those models are made of, the serving layer that answers requests, and the software that makes any of it run at all.
Separating the layers is not a presentational choice. It is what makes the industry legible. A shortage in packaging, a licensing agreement over training data, a rewritten kernel and a substation that cannot be energised are all constraints on AI, but they are constraints on different layers with different time constants — and a claim that does not say which layer it is about cannot be evaluated.
Two dependencies run outside this branch entirely and are worth naming at the top. Every accelerator in the compute layer depends on advanced packaging capacity in the semiconductor industry, which has been the binding constraint more often than chip supply itself. And every deployment depends on power and cooling that a data centre can actually deliver, which is now a multi-year procurement problem rather than a line item.
Each page below explains its subject on its own terms, then states what it depends on and what depends on it. Follow the dependency links and you can walk from a serving optimisation down to the lithography step that made the chip possible, or from a training run out to the transformer on the substation that powers it.
How this breaks down
Split by layer of the stack — the axis along which the dependencies actually run.
- Artificial intelligencethis page
- AI computeAccelerators, the memory beside them and the systems they ship in — the layer where the capital goes.Breaks down into: AI accelerators · AI memory · AI server systems14 pages belowChokepoint
- AI networkingThe fabrics that let accelerators exchange gradients and weights fast enough to stay useful.Breaks down into: Switch silicon · SerDes and retimers · Cables and connectors · Scale-up fabrics · Cluster fabrics · Optical interconnect6 pages belowChokepoint
- Model architectureThe architecture, the scaling recipe and the post-training that turn compute into a usable model.Breaks down into: Model evaluation · Transformers · Mixture of experts · Scaling and pre-training · Post-training5 pages belowChokepoint
- Data layerCorpora, pipelines, human data operations and the retrieval infrastructure around them.Breaks down into: Data processing frameworks · Training corpora · Data pipelines · Human data operations · Embeddings and vector search5 pages belowChokepoint
- Inference and servingThe runtime layer where models meet users, and where the recurring cost lives.Breaks down into: Serving engines · Context and caching · Quantisation · Retrieval-augmented generation4 pages below
- Software stackProgramming models, training frameworks and cluster schedulers — the layer that decides how much hardware you actually get.Breaks down into: Kernel libraries · Container platforms · Programming models · Training frameworks · Cluster orchestration5 pages below
What this depends on
4 of these are marked as a chokepoint: a handful of qualified suppliers, a multi-year lead time, or a single geography.
- Supply chainChokepoint
Advanced semiconductor packaging
Accelerators and their memory must be joined on an interposer. This step, not wafer supply, has repeatedly set the ceiling on how many parts exist.
Advanced packaging - Supply chainChokepoint
Leading-edge foundry capacity
Every competitive accelerator is built on the most advanced process node available, from a very small number of fabs.
Foundry and capacity - Supply chainChokepoint
Data-centre power and cooling
AI racks draw several times what conventional racks draw. Securing power, and removing the heat, is now the pacing item for deployment.
Data centres - Supply chainChokepoint
High-bandwidth memory
Made by three companies, consuming disproportionate wafer area, and fused into the accelerator package at assembly.
High-bandwidth memory - Standard
Export controls on advanced accelerators
Which parts may be sold into which country is set by governments, not by the vendors. A rule change re-routes demand overnight and strands designs specified against the previous one.
Geography and controls - Resource
Capital at industrial scale
A frontier training run and the building it happens in are financed years before either produces revenue. Without capital willing to sit through that gap the compute is not built, whatever the chips cost.
Companies across Artificial intelligence
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.
Custom AI ASICs · Inference silicon · Scale-up fabrics · Cluster fabrics · Model evaluation · Transformers · Mixture of experts · Scaling and pre-training · Post-training · Data processing frameworks · Data pipelines · Embeddings and vector search · Serving engines · Context and caching · Quantisation · Retrieval-augmented generation · Kernel libraries · Container platforms · Programming models · Training frameworks · Cluster orchestration
Merchant GPUs · Inference silicon · Compute trays · Rack integration · Switch silicon · Scale-up fabrics · Cluster fabrics · Optical interconnect · Transformers · Mixture of experts · Scaling and pre-training · Data processing frameworks · Data pipelines · Serving engines · Context and caching · Quantisation · Kernel libraries · Container platforms · Programming models · Training frameworks · Cluster orchestration
Custom AI ASICs · Model evaluation · Transformers · Scaling and pre-training · Post-training · Data processing frameworks · Data pipelines · Embeddings and vector search · Serving engines · Quantisation · Retrieval-augmented generation · Container platforms · Training frameworks · Cluster orchestration
- OpenAIPrivate10 steps
Custom AI ASICs · Model evaluation · Transformers · Mixture of experts · Scaling and pre-training · Post-training · Embeddings and vector search · Context and caching · Kernel libraries · Programming models
- Huawei TechnologiesPrivate8 steps
Merchant GPUs · Inference silicon · Switch silicon · Scale-up fabrics · Cluster fabrics · Kernel libraries · Programming models · Training frameworks
- Hugging FacePrivate8 steps
Model evaluation · Post-training · Data processing frameworks · Training corpora · Data pipelines · Serving engines · Quantisation · Training frameworks
193 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