Model as a Service (MaaS)

An AI model is a computer program that has learned from a very large number of examples to carry out certain tasks. It can, for example, write texts, generate images, or convert speech into text. Building such models costs many millions of euros and requires special computers. With Model as a Service, a provider takes on this work and makes the finished model available over the internet. Other companies then send it their requests and receive the answers back. Payment is based on usage, similar to electricity from the grid.

Why hardly anyone builds their own model anymore

A few years ago, a company that wanted to use AI had to hire its own experts and buy its own computers. Today, access to someone else’s model is often enough. This lowers the barrier to entry enormously. A small start-up with five people can build an application that runs on the same technology as that of a large corporation.

For providers, this is a very attractive business. They build the model once and then sell access to it millions of times over. This is exactly what underlies the high valuations of companies like OpenAI, Anthropic, or Mistral. That’s why terms like recurring revenue—income that flows in again every month—show up in financial news.

There is, however, a downside. Anyone who bases their product entirely on someone else’s model becomes dependent on that provider. If the provider raises prices or discontinues a model, the customer has to react. Experts call this vendor lock-in. That’s why some companies deliberately keep two providers running in parallel.

The journey of a request from customer to data center

Access runs through an interface, known in technical jargon as an API. This is a fixed address on the internet to which a program sends a request. The request contains the text or image in question. In the provider’s data center, the model processes it and sends the result back. For the customer, this usually takes only a fraction of a second.

For language models, billing is normally based on tokens. A token is a piece of text, such as a short word or a syllable. A thousand tokens correspond to roughly 750 words of English. Prices vary by model, ranging from a few cents to several euros per million tokens. Small, fast models are significantly cheaper than large ones.

Many providers also allow customization using your own data. You supply examples from your own operations, and the model is fine-tuned accordingly. It’s important to distinguish this from open-source models: these you download and run yourself. That involves no usage fee, but it does require your own computers and your own expertise.

MaaS in apps, cloud catalogs, and quarterly earnings

Most of the AI features people use every day run on this model. The translator in a travel app, the text suggestions in a writing program, the chatbot in an online shop’s customer service: behind these there is often a rented model from an external provider. The app itself contains no AI—it merely poses the questions.

The major cloud providers sell MaaS as well. Microsoft, Amazon, and Google list dozens of models in their catalogs, some their own, some from others. A developer selects a model there just like a product in an online shop. This revenue shows up in the corporations' quarterly reports and moves their share prices.

A common misconception is that MaaS automatically means the provider keeps training on the data it receives. For business customers, this is usually excluded by contract. Nevertheless, for European companies it’s crucial which country the servers are located in, since data protection law requires this. That’s why providers now explicitly advertise data centers within the EU.