There are servers — you just don't rent them
The name annoys engineers for good reason: the servers exist, operated by the provider at enormous scale. What disappears is your relationship to them. Instead of keeping an instance running to wait for requests, you register functions against events — an HTTP call, a file upload, a queue message — and the platform summons an execution environment when the event fires. No traffic, no cost, no capacity planning; a traffic spike is the platform's scheduling problem.
The constraints are the flip side of the billing model. An environment created on demand starts cold — noticeable latency on the first request after idle. Execution time and memory are capped. State must live elsewhere (a database, object storage), because the environment vanishes after the call. Serverless therefore shines for event-shaped, bursty, stateless work — APIs, image processing, scheduled jobs, glue between services — and fights you on long-lived, stateful or latency-critical paths.
Serverless beyond functions
The billing idea spread: 'serverless' databases and container platforms now scale to zero and charge by usage too. The common denominator is the economic model — pay for work done, not capacity reserved — rather than any particular technology.