
Foundry
A foundry is a service or platform through which companies can rent, customize, and integrate ready-made AI models into their own products without developing the models themselves. The term originates from the semiconductor industry, where foundries manufacture chips for other companies — the principle is the same: producing someone else's designs rather than pursuing your own.
A foundry is a platform that provides ready-made AI models as a service. Companies can rent these models, adapt them for their own purposes, and integrate them directly into their software. To do so, they need neither understand the underlying technology nor train their own models from scratch. The term comes from the semiconductor industry: there, a foundry is a factory that produces computer chips based on other companies' designs. In the AI world, the idea is the same — someone else built the foundation, and you use it.
Why foundries are changing the AI market
Developing a powerful AI model from scratch today costs hundreds of millions of euros and several years of work. Only a handful of corporations worldwide can manage that. Foundries drastically lower this barrier. A mid-sized company can thus launch a functioning AI-powered service within weeks.
This changes who can actually deploy AI at all. In the past, access to powerful AI was effectively limited to tech corporations. With foundries, hospitals, law firms, or craft businesses can also use AI tools tailored to their specific needs. The actual computational work remains with the foundry provider.
At the same time, this concentrates power among a handful of providers. Whoever controls the foundry determines which models are available, at what prices, and under what conditions. Critics therefore speak of a new kind of infrastructure dependency — similar to how companies today depend on cloud services.
What a foundry technically provides
At its core, a foundry provides three things: the model itself, the computing power needed to run it, and tools for customization. Customization usually happens through what is called fine-tuning — an existing model is further trained with additional, domain-specific data until it masters the desired task better. A model that writes general texts can thus become a specialist in legal phrasing.
In addition, foundries often offer interfaces through which developers can call the model directly from their own software. These interfaces are called APIs — put simply, these are standardized connection points, like power outlets for software. The developer writes their code, sends a request to the foundry, and receives the result back without needing to know how the model works internally.
Some foundries go even further and offer complete development environments: tools for testing, measuring model quality, and safeguarding against undesired outputs. The goal is to map the entire path from idea to finished product on a single platform.
Foundries in products and headlines
Microsoft has explicitly marketed its AI service on Azure as Azure AI Foundry. There, companies can test and deploy models from OpenAI, Meta, and other providers side by side. Google and Amazon also operate comparable platforms under other names. The term foundry has thus established itself as a marketing term for precisely this concept.
The word appears in financial reporting when technology corporations report their revenues from selling AI access. When a company announces that its AI revenue has doubled, the foundry model is usually behind it: many customers pay monthly for access to models they will never own themselves.
For consumers, the term is mostly invisible. Anyone using a chatbot in an app or reading automatically generated product descriptions is often dealing with a model that was sourced via a foundry. The company behind the app did not build the model — it rented it.