Comment by rescbr

If they need to divert inference resources to train models, this counts as a compute deficit to me.

I'm using GLM-5.2 daily for my own stuff, and during Chinese business hours, specially on their afternoon, it's a festival of rate limits.