OpenAI and Anthropic are the only pure play AI companies, and each runs four businesses at once: lab, product, services and scaling.
To understand where AI is moving, you have to understand how the pure play labs are built, and how Anthropic and OpenAI differ.
Of all the hyperscalers, only OpenAI and Anthropic are pure play AI companies. Amazon, Microsoft, Google, Meta and even Oracle are adding AI to businesses that already exist and already earn revenue. AI is only about 15% of big tech cloud revenue. At AWS, it's $25 billion of AI inside $169 billion of revenue. At OpenAI and Anthropic, it's 100%.
Inside each lab, four businesses run at once. The lab does the research and development behind the next model. The product delivers the models already released, through the API and subscriptions. Scaling gathers the infrastructure to serve inference in the future. Services, the newest of the four, gets businesses using AI, and using it better across more of what they do.
The lab
The lab does research and training to build the next model. On its own it generally earns nothing, but it drives everything the labs sell. Models keep getting bigger and cheaper to operate, while building them keeps getting more expensive.
Product
Product is how you deal with the labs: the ChatGPT app, the Claude app on your phone, the APIs businesses call, and the subscriptions a small fraction of people pay for to get the more advanced features. OpenAI's enterprise revenue passed its consumer revenue in August. Anthropic is now running at $80 to $100 billion a year.
Services
Services is new for both labs. They bring in consultants and advisors to work inside companies and put AI into their workflows, from products and services to the back office. Anywhere a business needs help getting AI up and running, the labs will step in, for a price.
Where it's heading: competition with their own ecosystem. The advisors and consultants who helped the labs get into big businesses will now compete with them directly. That may open more room for non-frontier models in corporate America.
Scaling
Scaling gets most of the attention. It's the data center build out, and where the labs get their compute. It happens three ways. They build sites themselves or with a partner. They lease capacity from providers like AWS, Google or CoreWeave. Or they sign purchase commitments for future compute. Those deals can keep the numbers off the balance sheet, but it's all money they will have to pay.
Pricing is mostly tied to compute deals still 12 to 18 months out. Recent changes in the rules on data center building are likely to move that timeline.
Four businesses
Lab, product, services and scaling. These are the four core functions at the heart of the pure play AI companies. They're where these companies spend their time, and what will drive the AI future for all of us.