Cloud Run has sub 30 second cold start times for some ollama models on NVIDIA L4s? It seems like not a bad idea to run your own managed ollama service.

When I joined Google, people used to spend time evaluating ideas and prototyping not for the sake of not working but not to jump on a low impact idea and burn millions of dollars. When I was leaving, company's attention was split in a million paper cuts, and new hires couldn't navigate it all and prefer to rest and vest.