The AI industry looks crowded from outside and is surprisingly concentrated inside. A small number of companies build the models everything else runs on, and knowing who does what makes the rest of the landscape legible.
The model builders
These are the companies training frontier models. Everyone else is, to some degree, a customer.
| Company | Models | Position |
|---|---|---|
| OpenAI | GPT | First mover, largest consumer mindshare |
| Anthropic | Claude | Safety-focused, strongest writer |
| Google DeepMind | Gemini | Search and Workspace integration |
| Meta | Llama | Open weights, ecosystem play |
| xAI | Grok | X integration, looser guardrails |
| DeepSeek | DeepSeek | Open weights, very low cost |
| Mistral | Mistral | European, efficiency-focused |
| Alibaba | Qwen | Open weights, strong multilingual |
The layer underneath
Easy to miss, and arguably where the most durable value sits.
Nvidia makes the chips nearly all of this trains and runs on. Whatever happens to any individual model company, the compute is sold by a very small number of suppliers.
The cloud providers — Microsoft Azure, Amazon AWS, Google Cloud — own the data centres and hold large stakes in the model companies. Microsoft/OpenAI and Amazon/Anthropic are the most significant of these relationships.
Chip challengers — AMD, Google's own TPUs, and a set of startups — matter because Nvidia's position is the industry's largest single dependency.
The application layer
Companies building products on top of someone else's models. Vastly more numerous, individually smaller, and the layer where most people actually encounter AI.
This includes writing tools, coding assistants, support platforms, image and video products, and aggregators — All Chatbots AI sits here, connecting to the model builders' official APIs.
The dynamic worth understanding: this layer has low barriers to entry and depends on suppliers who are also potential competitors. Products survive here by being genuinely better at a specific job than a general assistant is, not by having access to a model that anyone can license.
How to read industry news
"Beats GPT on benchmarks" usually means little. Contamination, cherry-picked tasks, and margins inside the noise. See how to read model comparisons.
Valuations are not revenue. Several prominent AI companies have enormous valuations against modest income and very large compute bills.
Partnership announcements are often distribution deals, not technical breakthroughs.
Open-weight releases reshape pricing more than they reshape capability. When a good open model appears, hosted prices fall.
What would actually change the landscape
Worth watching, as opposed to the daily noise:
- Open models closing the frontier gap. They keep narrowing it. If it closes, the economics of the whole application layer change.
- Compute costs falling. More than model quality, this determines what becomes viable.
- Regulation, particularly in the EU, which sets rules others often follow.
- Chip supply diversifying away from a single dominant supplier.
- Consolidation in the application layer, which is overcrowded and underfunded relative to its size.
Common questions
What are the best AI companies? For models: OpenAI, Anthropic, Google DeepMind, Meta, xAI, DeepSeek, Mistral and Alibaba. Nvidia and the cloud providers own the infrastructure underneath.
Which company has the best AI model? No single one leads everywhere. Anthropic leads on writing, Google on live information, and the largest tiers of each on hard reasoning.
Who owns OpenAI and Anthropic? Both are independent companies with major cloud investors — Microsoft in OpenAI, Amazon and Google in Anthropic.
Is Nvidia an AI company? It makes the chips nearly all AI trains and runs on, which arguably makes it the most structurally important company in the field.
Which AI companies are open source? Meta, DeepSeek, Mistral and Alibaba release open weights. None releases full training data. See open-source AI models.
How do AI companies make money? Subscriptions, per-token API access, cloud contracts, and hardware. Several frontier labs are not yet profitable.
Is the AI industry a bubble? Valuations are high relative to current revenue and compute costs are enormous. The technology is clearly useful; whether current valuations reflect that is a different question.