A serverless program (often called a function) is invoked by a server on-demand instead of running all the time, and it doesn't listen for anything on any port. The client request is passed to it by the server (e.g. on standard input) and it's supposed to pass the response back to the server (e.g. through standard output) - and the server sends it to the client. Then the program halts.
Compare that to Django, Node.js or ASP.NET: A classic backend app exposes a HTTP server on a port and handles client connections itself - and thus it has to run all the time, it literally is a server (in the software sense).
If you know PHP, that's the original serverless. As opposed to the Python/ASP.NET/Node.js backends, your page.php is invoked by the Apache daemon only when someone opens that page, and there's no "server.listen(3000)" in page.php.
Serverless is cool because it allows the cloud provider to fully utilize a machine while the developer pays only for their portion of actual usage. You don't need to reserve a specific amount of compute resources and worry about up/down-scaling or about paying for unused hardware.
Serverless GPUs are about bringing that concept to the GPU as a service space - ideally you'd have a function that uses the GPU. That function could be invoked by sending a request to the platform's server, at which point the server would load and execute it, pass the client request to it and pass the response back to the client once the function is done.