I wouldn’t call it an air gap, but when you use a model on Bedrock or Vertex the inference runs on Amazon or Google’s servers respectively. The model provider gives them the weights, and is not otherwise in the loop for serving individual requests.