Comment by zozbot234

The cloud providers are probably better at procuring hardware for inference, but on prem users are better at repurposing hardware that they'd need anyway for their existing uses. In a world where AI compute is likely inherently scarce, it makes sense to rely on both.