Serverless

The name is misleading: servers still exist. What is meant is that nobody has to take care of them anymore. Anyone running an app or a website normally needs a computer that runs constantly and is reachable at all times. This machine has to be set up, supplied with security updates, and kept under observation. With serverless, a large provider such as Amazon, Microsoft, or Google takes over this completely. You just upload a piece of program code and specify when it should be executed.

Paying by the millisecond instead of a monthly rent

The most important difference lies in the costs. A rented server costs money whether it is working or not. At three in the morning, when nobody is visiting the site, the bill keeps running just the same. With serverless, only the actual computing time is billed, often in steps of milliseconds. Without requests, there are practically no costs.

The second advantage is automatic adaptation to load. If a video suddenly goes viral and, instead of a hundred people, a million call up the page, the provider simply starts more copies of the code in the background. A classic server would have collapsed under this load. For small teams this is enormously valuable, because they don’t need their own specialists for server operations.

There is a downside, though. Anyone who builds everything with one provider can later find it hard to switch, because every provider has its own rules and tools. Experts call this lock-in. In addition, serverless often becomes more expensive than a fixed rented server under very large, steady, continuous load.

Functions, triggers, and the cold-start problem

The code is broken down into small, self-contained units called functions. A function handles exactly one task: resizing an uploaded photo, writing an order to the database, sending an email. Each function waits for a trigger, that is, a specific event. This can be a click on a website, a new file in storage, or a fixed point in time.

If nothing happens, the function exists only as a stored file. When a request comes in, the provider starts a small, isolated environment within fractions of a second, executes the code, and then tears it down again. This booting up from nothing is called a cold start and costs noticeable time, sometimes half a second or more. For frequently used functions, the environment stays warm for a while, so the delay is avoided.

This leads to an important property: serverless functions must not remember anything. Anything that needs to be stored has to go into an external database or storage service. Experts call this stateless. Runtime is also limited, often to a few minutes. For that reason, the technology is unsuitable for long computing tasks.

From the phone app to the AI query

Serverless underlies many services without being visible. When a streaming platform automatically generates thumbnails after an upload, such a function is often running in the background. Contact forms, chat notifications, and billing in online shops are frequently built this way too. The user notices nothing, just seeing a website that responds.

In business news, the term mainly appears in connection with cloud providers, i.e. companies that rent out computing power over the internet. Amazon Web Services kicked off the development in 2014 with the Lambda service, and Microsoft and Google followed suit. Serverless is considered one of the business areas with which these corporations tie customers to themselves in the long term.

Increasingly, the idea is also being applied to AI. Instead of permanently renting an expensive graphics card, you call an AI model via an interface and pay per request. Providers advertise this with terms like serverless inference. The principle remains the same: someone else’s hardware, your own task, billing by usage.