All news

GPT-3's 300 apps turn language models into a metered utility

A row of white taps on a graphite-blue wall, one with a cobalt handle running a thin stream of water into a basin

Nine months after launch, OpenAI's API serves more than 300 apps and 4.5 billion words a day. Large language models are becoming a layer others build on.

More than 300 applications now run on GPT-3, and the model generates an average of 4.5 billion words a day. OpenAI gave those figures on 25 March, nine months after it launched the API as its first commercial product, with tens of thousands of developers now building on the platform.

GPT-3, with 175 billion parameters, is sold as a service rather than released as research, and that choice turns a language model into a metered utility. The supplier carries the training cost once and spreads it across hundreds of products; each developer converts a research budget it could never afford into a usage bill. Pricing power moves to the few companies able to train such models, at least until open alternatives close the gap. The 300 apps mark the start of a platform layer in which an application keeps only as much value as it owns beyond the model: its users, its workflow and the data it collects.

The range of uses shows how far the barrier has fallen. Products run from productivity and education tools to games, where one studio has GPT-3 write live dialogue for an interactive character. Many of the developers have no traditional machine learning background, and new features let them ground answers in their own documents and classify text without training a model. A language product no longer needs a research team to ship.

The utility also sets the rules. OpenAI reviews every application before it goes into production, so a launch now waits on another company's approval policy, and that dependency is part of the price of renting capability.

Cheap access cuts both ways. The same low barrier helps every competitor, so a thin layer over a prompt is easy to copy. Durable products gather feedback and data the model supplier never sees, keep the option to switch models and test their margins against changes in the supplier's prices.

Competition at the model layer would loosen the supplier's grip, and it has already begun. Four days before OpenAI's update, EleutherAI, an independent research collective, released open-source language models of up to 2.7 billion parameters, built to replicate the GPT-3 design at a smaller scale. How fast open models close the gap with GPT-3 decides how long prices hold, and comparable models from the largest cloud providers would turn a single supplier into a market. Usage gives the other reading: if it keeps compounding as OpenAI widens access beyond its waitlist, the platform layer is real whichever model wins.

The model layer looks set to stay concentrated and capital-hungry, while the application layer can support many winners. When Fanaura weighs a product built on a rented model, the conversation with its founders starts with what it owns beyond that model, and our thesis favours AI-first products that own the workflow.

For founders

Building an AI-first product on large language models that owns its own data loop? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.