Yep.

Also, large companies are still experimenting on how to integrate LLMs into their workflows. Due to the fast cadence of releases, people forget that LLMs became robust (regardless of the capability level/parameter count/data size) enough to use semi-reliably in company-specific ways only 1 year ago. And bigger the org, the slower the process. I don't expect it to settle and get productive used across the majority of very large companies for another year atleast.

High prices are mostly a result of DC capacity. As more and more DCs get built out, prices will drop. At a unit level the inference business is extremely sound regardless.