OpenAI's API turns its largest model into rented infrastructure

OpenAI chose to sell access to its GPT-3 models through an API instead of releasing them. Builders gain capability without training costs, and a dependency.
OpenAI has decided to rent out its largest language models rather than release them. Its first commercial product, launched on 11 June, is an API that gives developers access to new models from the GPT-3 family through a general-purpose text in, text out interface. Given a prompt and a few examples, the model continues the pattern across many English-language tasks.
That choice creates a new layer in the market. A few model providers hold the weights and the compute, and a much larger set of products gets built on top through an API. The model itself becomes infrastructure, and application value depends on what a company adds that the model cannot: proprietary data, workflow and distribution.
Scale explains the decision. Two weeks before the launch, a research paper from OpenAI described GPT-3 at 175 billion parameters, ten times more than any earlier language model of its kind, and models of that size cost more to train and serve than almost any application company can afford. OpenAI gives three reasons for renting rather than releasing: the revenue pays for its research, the models are too large and expensive for most companies to run, and an API lets it cut off harmful uses. Keeping the model in-house and selling access is the move Amazon made in March by selling its store checkout models to other retailers.
Builders gain speed and take on a landlord. A team that wants language capabilities no longer needs to train a large model and can prototype in days on rented access. The provider, in turn, sets who gets access, the rules of use and, soon, the price, and it can end access for uses it judges harmful. Founders on the API should own what the model does not provide, such as data from their users, evaluation of output quality in their domain and the workflow around the text the model writes. That ownership is the first thing Fanaura checks in founders who build on another company's model.
Price is the missing number. OpenAI has not published one yet, and it will set the margins of everything built on the API, which decides whether products on top can earn software margins at all. Access is a private beta with a waitlist, early users include Quizlet and Reddit, and how fast that waitlist turns into paying applications in production, and in which tasks, will show where rented language actually earns. Dependence on a single supplier is the other risk, and it eases only if other large labs open their own models through similar interfaces.
The API marks the point where the most capable language models became something companies rent rather than build. Lasting value will gather in applications that turn rented capability into owned data and customers, the layer the Fanaura mandate targets. Renting the model is cheap, and owning the customer is the business.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building a product on rented language models that owns its data and its users? 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.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.