The map, who does what

Five layers

Think of it as a stack. Each layer sells to the one above it, and you are at the top.

1. Chips. The physical processors that do the maths. Training a big model takes tens of thousands of them running for months; answering your question takes a slice of one for a second. Nvidia makes most of them. Google, Amazon, AMD and a few others make the rest.

2. Clouds. Warehouses full of those chips, rented by the hour. Amazon Web Services, Microsoft Azure and Google Cloud are the big three. A model company does not buy a building; it rents part of one of these.

3. Model makers. The labs that train the models: OpenAI, Anthropic, Google DeepMind, xAI, Meta, Mistral, DeepSeek, Alibaba. This is the layer with the famous names. They spend the money on layers one and two and sell what comes out.

4. Apps and tools. Where a model becomes something you can use. Some are made by the model makers themselves (ChatGPT, Claude, Gemini, Grok). Some are made by others on top of a rented model: Cursor for writing code, Perplexity for search, Notion's AI features, the assistant in your bank's app. Most products you meet in a normal day sit here.

5. You. Your instructions, your connectors, your account. On this site, the agent you build is layer five: a set of standing orders that runs inside a layer four app, on a layer three model, in a layer two building, on layer one chips.

Why one company shows up on several layers

The confusing part is that the biggest companies are on more than one rung, sometimes all of them.

Google makes chips (TPUs), runs a cloud, trains Gemini, and ships the Gemini app. Amazon makes chips (Trainium), runs the biggest cloud, does not train a headline model of its own, and instead has put more than thirty billion dollars into Anthropic. Microsoft runs a cloud, owns about a quarter of OpenAI, and ships OpenAI's models inside Office and Windows as Copilot. Meta makes a model (Llama), gives it away, and uses it inside WhatsApp and Instagram.

So when you read "Amazon invests in Anthropic" or "Apple picks Gemini for Siri", the question to ask is: which layer is buying from which?

Who pays whom

Follow the money and the stack explains itself:

• You (or your employer) pay a subscription or a per-use fee to an app.
• The app pays the model maker per token, or is the model maker.
• The model maker pays the cloud for compute, by far its largest cost.
• The cloud pays the chip maker.

Every dollar flows downhill toward Nvidia, which is why Nvidia became one of the most valuable companies in the world without most people ever using a Nvidia product directly.

Where the competition actually is

Each layer has its own fight:

• Chips: Nvidia has roughly four fifths of the market. The threat is not a rival chip company so much as the clouds building their own chips to stop paying Nvidia.
• Clouds: the big three plus specialist clouds that only do AI. The model makers play them against each other; Anthropic, for example, runs on Amazon, Google and Nvidia hardware at once.
• Models: five or six labs at the frontier, a new generation from each every year, and open-weight models from China and Europe that keep prices falling for everyone.
• Apps: thousands of companies, most of them renting the same handful of models. The ones that win are the ones that fit into your day, not the ones with the best benchmark.

What this means for you

Two things.

First, the model is not the product. Two apps can run the same model and feel completely different. Pick apps by what they connect to and how they fit your life.

Second, do not marry a layer. The instructions you build here are plain text. They run in Claude today and can run in ChatGPT or Gemini tomorrow, because the site writes for all of them. Anything you build that only works in one company's app is a bet on that company. Keep your work portable and the stack can shuffle all it likes.

What to remember

• Five layers: chips, clouds, model makers, apps, you.
• Big companies sit on several layers, which is why the deals sound tangled.
• Money flows down toward the chips. Attention flows up toward the apps.
• The model is not the product. Keep what you build portable.