Cold start
The delay when a service must start a new machine and load a model before it can answer the first request.
A cold start happens when no running copy of a model is ready. The platform must find a GPU, start a container and load gigabytes of weights from storage into GPU memory, which can take from a few seconds to several minutes for large models. Serverless platforms fight this with fast storage, snapshots of loaded memory and keeping a few copies warm, which costs money. Users with steady traffic often prefer reserved endpoints to avoid cold starts entirely.