Searching for "AI companies" returns ranked lists, and ranked lists of a fast-moving private market are stale the week they publish. More importantly they flatten a real distinction: the businesses on those lists are not in the same industry.
A company designing accelerators and a company selling a scheduling tool with a model call in it share a label and nothing else — not capital requirements, not margins, not defensibility, not the questions you should ask before trusting them.
This page is the map rather than the list. For the list, with published scores, see the company directory.
The six layers
Read this as a stack. Money and physical constraint concentrate at the bottom; count of companies concentrates at the top.
| Layer | What it actually sells | Capital needed | What makes it defensible |
|---|---|---|---|
| 1. Silicon and compute | Accelerators, interconnect, fabrication capacity | Extreme | Fabrication access, software ecosystem lock-in, multi-year design cycles |
| 2. Cloud and infrastructure | Training and serving capacity, and the power and cooling behind it | Extreme | Physical footprint, energy contracts, existing enterprise relationships |
| 3. Foundation models | General-purpose models, sold by API or licence | Very high | Talent, compute access, distribution — capability leads erode quickly |
| 4. Data and labelling | Training corpora, annotation, human preference data, evaluation | Moderate, labour-heavy | Rights to data, workforce quality, domain expertise |
| 5. Tooling and orchestration | Retrieval, agent frameworks, evaluation, observability, guardrails | Low | Developer adoption — thin, and routinely absorbed by layers 2 and 3 |
| 6. Applications | Software solving a specific problem, with models inside | Low | Domain knowledge, workflow integration, customer relationships, regulatory position |
Two observations that the ranked lists cannot express.
Defensibility does not track excitement. Layers 1 and 2 are the hardest to enter and the hardest to displace, because they are constrained by physics, capital and time. Layer 5 is the easiest to enter and the easiest to lose, because what is cheap to build is cheap to rebuild — and because the layers below it keep absorbing that functionality into their platforms.
Most "AI companies" are layer 6. That is not a criticism. Applying a general capability to a specific problem, with the domain knowledge to know what "correct" means, is where most of the realised value sits. But it means the label describes what a company uses rather than what it is, and a business whose AI is one supplier's API among several is a software company with a dependency.
The layer decides which ethics questions apply
This is the part that matters for anyone assessing a company rather than investing in one, and it is why a single questionnaire sent to every vendor produces so little.
| Layer | The questions that actually bite |
|---|---|
| Silicon and compute | Supply chain and materials sourcing; who is sold to and under what controls; energy and water at the fabrication end |
| Cloud and infrastructure | Energy source and grid impact; water consumption for cooling; siting decisions and the communities affected; what customer data is retained |
| Foundation models | Training data provenance and consent; what evaluations were run and what they missed; release and access policy; whether capability claims are independently checkable |
| Data and labelling | Working conditions and pay for annotators; psychological safety on harmful content; whether data subjects consented; whether rights were actually held |
| Tooling | What the tool logs and where it sends it; whether "guardrails" claims are measured or asserted; failure behaviour |
| Applications | Deployment context and who is affected; meaningful human oversight; how a person wrongly treated finds out and gets redress; whether the domain is regulated |
Ask an application vendor about training data provenance and you will usually get an honest shrug — they did not train the model. Ask a foundation model developer about redress for an individual and you are asking the wrong party, because they never saw the individual. Both answers sound evasive and neither is. The question was misaddressed.
The practical version of this is in the vendor risk assessment, and the layer above your vendor usually matters too: if your supplier is layer 6 reselling a layer 3 model, the model's limitations are yours whether or not your contract mentions them.
The layer nobody writes about
Layer 4 — data and labelling — attracts the least coverage and carries some of the heaviest questions.
Models are shaped by data that was collected, cleaned, annotated and ranked, and much of that is done by people. Where those people are, what they are paid, what they are shown, and what support exists when the content is distressing are questions with clear answers that are rarely published. So is whether the rights to a corpus were held rather than assumed.
The reason this layer stays invisible is structural: it sits behind a company you have heard of, and the disclosure obligations follow the brand rather than the supply chain. Any serious assessment of a foundation model developer is partly an assessment of suppliers it may not name.
How to judge a company at any layer
Independent of layer, four questions separate a substantiated claim from a marketed one.
- What does it control, and what does it rent? A company whose core capability is another company's API has that company's economics and constraints, one step removed.
- What is checkable? Prefer claims someone outside the company could verify — published evaluations, third-party audits, documented policy — over adjectives. "Responsible" is not a claim.
- What happens when it is wrong? Every one of these systems is sometimes wrong. A company that can describe its failure modes and what follows from them has thought about it; a company that cannot has not.
- Who is accountable, by name? A published owner for AI decisions is a stronger signal than any policy document — see who owns AI risk.
What we publish, and what we do not
We score companies rather than count them, and it is worth being precise about the difference.
Our directory is curated, not a census. It does not attempt to list every company in the industry, and it should not be read as a market map. What it offers instead is scoring against a published methodology — three pillars, weighted Ethics 30%, Innovation 35% and Profitability 35% — with the weights, the factors, the data sources and the limitations all stated so that you can disagree with them specifically rather than in general. The rankings apply that method across the entries.
You will notice this page names no companies. That is deliberate, for two reasons worth stating rather than hiding. A named list inside an article is wrong within months and has no mechanism for correction, whereas the directory is maintained and dated. And any factual claim about a named private company is a claim we would have to stand behind — so we make those where they can be scored, sourced and revised, not in prose.
The bottom line
"AI company" is a label, not a category. Before comparing two of them, establish which layer each occupies — the answer changes what good performance would even look like, and it changes which questions produce a real answer.
If you are buying, start at vendor risk. If you are assessing an industry, start with the stack above. If you want scored entries rather than a landscape, that is the directory.