Comment by danpalmer

I imagine there wouldn’t bd much of a cost to the provider on the API call there so much longer times may be possible. It’s not like this would hold up the LLM in any way, execution would get suspended while the call is made and the TPU/GPU will serve another request.