Why does a private endpoint take so long to create?

Creating an endpoint involves multiple stages:

• Planning and provisioning the server;
• Downloading vLLM and launching the container;
• Downloading model weights [1];
• Loading weights into GPU memory;
• Computing CUDA graphs;
• Starting inference and tooling engines;
• Verifying system routes /health and /v1/models;
• Configuring the load balancer.

[1] Weight downloads depend on the load of foreign services and network connectivity, and currently may take longer than usual.

That's why endpoint creation takes more time than provisioning a regular server without preinstalled software.