Right, I still find it faster to manually provision a specific instance type, install PyTorch on it, and deploy a little flask app for an inference server.