Model as a Service

An AI program that writes texts or generates images needs very expensive specialized computers just to run at all. Hardly any company buys these computers itself. Instead, a provider operates the finished program in its own data centers. Customers send their request there over the internet and get the answer back. This exact business model is called Model as a Service, abbreviated MaaS. So you rent the capability of software, not the software itself.

Why hardly anyone buys their own AI computers anymore

The computers on which large AI systems run often cost several tens of thousands of euros apiece. For a larger system you don’t need just one of them, but hundreds. On top of that come electricity, cooling, and the specialists who keep the whole thing running. A mid-sized company could hardly manage that. Through a service, it instead pays only for the requests it actually makes.

This lowers the entry barrier enormously. A student-run company can today use the same AI as a major corporation, because both call the same service. The only difference is the volume of requests. That’s why, in recent years, a great many small companies have emerged whose product is essentially a good idea plus a rented AI service.

For providers, this is a very attractive business. They train a model once and then sell access to it millions of times over. That’s exactly why companies like OpenAI, Anthropic, or Google appear so often in business news. Their revenue depends on how many requests run through their data centers every day.

The path of a request to the data center

The connection runs through what’s called an interface, technically an API. This is a fixed address on the internet to which a program can send a message in a defined format. The provider receives the message, has its model compute a result, and sends the result back. For the customer, this usually takes less than a second. What happens in the meantime inside the data center remains completely hidden.

Payment is usually based on amount of text. The billing unit is called a token, a word fragment about four characters long. Both the question and the answer are counted. One provider, for example, charges a few euros per million tokens. Anyone running an app with many users sees these costs clearly on the monthly bill.

MaaS should not be confused with open-source models. With open source, you’re allowed to download the model’s files and run them on your own computers. With Model as a Service, you never get to see the model itself, only its answers. Both exist side by side, and some providers offer the same model in both variants.

MaaS in apps, invoices, and headlines

Very many programs that advertise themselves with AI are not themselves AI at all. A translation button in an online shop, a summary in an email program, a chat window in customer service: behind it is often just a call to a rented model. The visible part is the interface; the actual work happens elsewhere.

In business news, you mostly encounter the term in connection with dependency. If a provider raises its prices, shuts down a model, or has an outage, this immediately affects thousands of companies at once. Data protection also plays a role, because requests leave the company’s own premises. European companies therefore often make sure that the data centers are located within the EU.

A typical misconception is the assumption that a rented model learns along from your own data. By default, this is usually ruled out for business customers. The model stays the same for all customers and does not change due to individual requests. Adjustments for a specific company go through separate procedures that must be explicitly commissioned and paid for.