I'm going to assume they are saving on inference by rolling it out faster.
Why do you think rolling out faster -> more inference savings?
alt Hacker News
Why do you think rolling out faster -> more inference savings?