Who is who in AI, and what each one is actually for.
Every map of this industry is either a vendor's marketing, an investor's portfolio with the labels filed off, or four hundred logos in a grid that tells you nothing about what any of them do. This one answers one question per entry: what is it, and what would you actually use it for.
No valuations, no funding rounds, no market share, no benchmark league table. Those are the parts that are wrong within a month, and a reference that goes quietly stale is worse than no reference at all. Where something is genuinely contested, it says so on the card.
The labs building frontier models
A small number of organisations training the largest general-purpose models. This is where most of the money and nearly all of the noise is.
OpenAI
Independent, with a deep commercial and compute partnership with Microsoft.The lab behind the GPT model family and ChatGPT.
ForThe default general assistant for most people, and the API most products were built against first.
Not settledIts corporate structure has been restructured more than once and is genuinely complicated. Read a primary source rather than a summary, including this one.
Anthropic
Independent, with major investment from Amazon and Google.The lab behind the Claude model family.
ForLong-document work, coding, and agent-style tool use. Published the Model Context Protocol, now the common standard for connecting tools to models.
Google DeepMind
Alphabet.Google's combined AI research organisation, behind the Gemini models.
ForModels wired directly into products a billion people already use: search, Android, Workspace. Distribution is the differentiator here, not the model.
Meta AI
Meta.Meta's research group, behind the Llama models.
ForOpen-weight models you can download and run yourself. The strategy is commoditising the layer rather than selling it.
Not settledLlama is usually called open source. The licence carries conditions, so open weights is the more accurate term.
Mistral AI
Independent.A French lab shipping both open-weight and commercial models.
ForSmaller efficient models, and the main European option where data residency matters.
xAI
Independent, closely tied to X.The lab behind the Grok models.
ForA general assistant, tightly integrated with X.
The compute layer
Models run on physical hardware in buildings that need enormous amounts of electricity. This layer is the actual constraint on the industry.
Designs the GPUs that most AI training and inference runs on.
ForThe picks and shovels. Its CUDA software stack is as much of a moat as the silicon, because years of tooling assume it.
Not settledHow durable that lead is, given serious effort from AMD and from the clouds' own chips, is one of the most consequential open questions in tech.
The hyperscalers
AWS, Microsoft Azure and Google Cloud.
ForWhere models are actually served from. All three also design their own AI chips, so they are simultaneously Nvidia's biggest customers and its competitors.
The Taiwanese foundry that manufactures the leading-edge chips.
ForNearly everything advanced is physically made here. The single most concentrated dependency in the entire stack, and the one least discussed in AI coverage.
Electricity generation and grid capacity.
ForIncreasingly the binding constraint on new data centres. When you read about AI companies signing deals with power producers, this is why.
Not settledProjections of AI electricity demand vary enormously between credible sources. Treat any single confident figure with suspicion. The IEA publishes the most careful public work on it.
The plumbing
The unglamorous layer that connects models to real data and real tools. Boring, and where most of the actual engineering happens.
An open standard for connecting models to tools and data sources.
ForWriting one integration that any supporting model can use, instead of bespoke wiring per model. Published by Anthropic in November 2024 and adopted well beyond it.
Vector databases
Databases that search by meaning rather than by keyword.
ForRetrieval: finding the passages worth showing the model. Pinecone, Weaviate, Qdrant, Chroma, and increasingly just Postgres with the pgvector extension.
Not settledWhether this stays a separate product category or becomes a feature of ordinary databases is still an open bet. The trend favours the latter.
Orchestration frameworks
Libraries for chaining model calls, tools and retrieval together.
ForGetting to a prototype fast. LangChain and LlamaIndex are the best known.
Not settledPlenty of production teams drop these and call the APIs directly once the shape of the problem is clear. That is a real pattern, not a criticism.
Evals
Test suites for model behaviour.
ForKnowing whether a change made things better or worse. The single most underrated thing on this page: without evals you are guessing, and with them you are engineering.
Open weights
Models you can download and run on your own hardware. A different bargain: you take on the operations, you keep the control.
The main public repository for models and datasets.
ForFinding, downloading and comparing open models. Effectively the package registry for this layer.
Llama, Mistral, Qwen, Gemma, DeepSeek
The most widely used open-weight model families.
ForRunning on your own infrastructure when data cannot leave, when you need a fixed unit cost, or when you want to fine-tune.
Not settledLicences differ meaningfully between these and several are not open source by the strict definition. Read the licence before shipping.
Ollama, llama.cpp, vLLM
Runtimes for serving models yourself.
ForOllama and llama.cpp for a laptop; vLLM when you are serving many users and care about throughput.
It is not complete and it is not neutral. It leaves out most of the application layer on purpose, because that list changes weekly and nobody needs another one. It is a judgement about what matters structurally, and judgements can be wrong.
Something missing, or something here that has changed? Send it over. Entries carry their own review date so you can see what is fresh, and corrections land on the entry rather than in a new post. Start here if you want the explainers first.