Anyone running >1000 servers with >80% utilization should consider going partially colocated.
Comma.ai and Tesla come to mind who run their own training clusters.
If you are serious about compute and cost of compute and storage is ~100s of engineers, then moving off cloud is a wise strategy.
Cloud providers have huge margins. One can run a very profitable business if they get a pie of that margin.
Anyone running >1000 servers with >80% utilization should consider going partially colocated.
Comma.ai and Tesla come to mind who run their own training clusters.
If you are serious about compute and cost of compute and storage is ~100s of engineers, then moving off cloud is a wise strategy.
Cloud providers have huge margins. One can run a very profitable business if they get a pie of that margin.