Who are the AI companies, and what does each one sell?

Who are the AI companies, and what does each one sell?

The AI industry sorts into four layers, and each layer sells a different thing. Labs train models and sell access to them. Cloud providers rent the computers everything runs on. One chipmaker, Nvidia, sells the specialized chips both of those groups depend on. And a much larger crowd of application companies rents a model from a lab and builds a product around it.

The layers matter for hiring because "AI company" appears on thousands of resumes and describes work at every one of them. An engineer at a lab, an engineer at a cloud, and an engineer at a startup that rents a model all worked in AI, and they did completely different jobs. This lesson maps the layers so a company name on a resume tells you which job it was.

Which companies train large models?

OpenAI, Anthropic, Google, and Meta train large language models, along with a short list of others: Mistral in France, DeepSeek and Alibaba in China, xAI, and a handful more. Worldwide, the companies training models at the frontier number in the dozens.

The group is small because training is expensive in every direction at once. A frontier training run costs millions to hundreds of millions of dollars, takes months, and needs thousands of specialized chips running the whole time. It also needs a rare kind of engineering team: people who work on training itself, on the systems that keep a months-long run alive, and on serving the finished model to millions of requests. The labs lesson covers what training actually is; the point here is what it costs, because the cost is what draws the line between the labs and everyone else.

That line is the single most useful cut when reading a resume. Time at a lab usually means training work, serving work, or the research around them. Time at any other AI company usually means product work on top of a model someone else trained. Both are real engineering. They are different skills, practiced by different people, and the company name is often the only clue on the page.

What is a foundation model?

A foundation model is a large general-purpose model trained once, at great cost, and then used as the base for many different products. GPT, Claude, Gemini, and Llama are foundation models. The name comes from construction: the expensive part gets poured once, and everything else stands on it.

The word signals a business structure as much as a technology. Because training is so expensive and the result is so general, it makes sense for a few labs to train foundation models and for everyone else to build on them, the same way a few utilities generate power and everyone else plugs in. When a job posting says a team "builds on foundation models," it is saying the team rents its intelligence and spends its engineering on the product around it.

A foundation model is general on purpose. The same base model answers legal questions, writes code, and summarizes meetings, because it was trained on broad data rather than for one task. Companies then shape it toward their use with instructions and with the data they feed it, which is application work, done far from the lab that poured the foundation.

Why is Nvidia in every AI conversation?

Nvidia makes most of the GPUs that models are trained and run on. A GPU, short for graphics processing unit, is a chip originally designed to draw video-game graphics, which happens to require millions of small calculations at once. Training a model requires the same shape of work at enormous scale, so the gaming chip became the engine of the AI industry. That accident of fit is why a graphics company sits underneath every lab and every cloud.

The chips are scarce and expensive. A single top-end GPU costs tens of thousands of dollars, training runs use thousands of them, and for stretches of the AI boom demand has outrun supply. Access to GPUs sets the ceiling on what any lab can train and any company can serve, which is why Nvidia comes up in AI conversations that have nothing to do with its own products. When people say "compute" in an AI context, GPUs are mostly what they mean.

On a resume, GPU work is its own specialty. An engineer who writes software for GPUs directly, or who keeps large fleets of them busy and healthy, holds a scarce skill that labs and clouds compete hard for. That work sits closer to the hardware than most software jobs and is a different profile from building products on a rented model.

What do the cloud providers sell in AI?

AWS, Google Cloud, and Azure sell AI the same way they sell everything else: they rent the machines. The cloud lesson covered how companies rent computers instead of owning them, and AI made those rentals bigger. The clouds buy GPUs by the hundred thousand, install them in their data centers, and rent them to labs for training and to everyone for running models.

They also resell the models themselves. Azure carries OpenAI's models, AWS and Google Cloud carry Anthropic's, and each cloud offers a menu of models behind its own API. A company already running on AWS can add Claude to its product through the same account it already pays. This is why so much of the money in AI lands with the clouds: whether a lab wins or an application company wins, the computers were rented either way.

For hiring, cloud names on a resume need the same layer-reading as lab names. "AI infrastructure at Google Cloud" most often means the machinery around models: the systems that schedule GPUs, move data, and keep model-serving fast. That is infrastructure work, and it is deep, valuable engineering that transfers well to any company running models at scale. It is a different day-to-day from training models and different again from building products on them.

What are open weights, and what is Llama?

When a lab finishes training, the result is a set of files called the weights: billions of numbers that hold everything the model learned. Most labs keep their weights private and sell access through an API. An open-weight model is one where the lab publishes the files instead, so anyone can download the model and run it on their own hardware.

Llama is the best-known open-weight family, published by Meta. Mistral and DeepSeek publish open-weight models as well. Meta gives Llama away because its business is advertising rather than model access, and a widely used free model shapes the industry in its favor.

Teams choose open weights for control. A hospital, a bank, or a defense contractor often works under rules that say patient records or client data cannot leave the company's own network. A rented model through an API means sending data out; an open-weight model runs inside the walls, on machines the company controls. Open weights also let a team retrain a model on its own data and avoid per-request fees at high volume.

What do the rest of the AI companies do?

They build on rented models, and they are the overwhelming majority. For every company that trains a model, hundreds build products by calling a lab's API: a legal-drafting tool on Claude, a sales assistant on GPT, a support product that routes between several models. The model is an ingredient they buy, the way earlier companies bought payment processing from Stripe instead of building a bank.

This is honest and standard, and the engineering in it is real. The work is shaping the product around the model: getting the right company data in front of it, checking its output, handling its failures, and keeping response times and per-request costs inside what the business can afford. Those skills are close cousins of ordinary backend work, which is why strong AI-application engineers so often come from teams that never touched AI.

So when a resume says "AI company," place it on the map before reading further. A lab means training and serving. A cloud means the machinery underneath. Nvidia and its neighbors mean hardware. Everyone else means product work on a rented foundation model, and the next lesson shows how to tell, from the outside, which kind of AI work a company actually does.