loading
loading
The extra delay when a server or function spins up from idle to handle the first request.
Serverless platforms and on-demand GPU endpoints shut down idle instances to save money, so the next request pays a startup tax: loading the runtime, and for models, pulling tens of gigabytes of weights into VRAM. That can be milliseconds for an edge function or many seconds for a large model. You hide it with warm pools, provisioned concurrency, or a keep-alive ping. It's the classic gotcha behind 'why was the first request after lunch so slow?'
Plainly
Think of Cold Start as roads and power for the app city. The extra delay when a server or function spins up from idle to handle the first request.
In practice
Use it when local behavior needs to become a reachable, reliable deployed service. In practice, define the owner, input, output, and failure mode before you rely on it.