All these efficiencies just increase OpenAI's margin on inference. Of course it's not "one cluster per customer" and of course a customer can't saturate a clust

All these efficiencies just increase OpenAI's margin on inference. Of course it's not "one cluster per customer" and of course a customer can't saturate a cluster by themselves, my illustration was only to point out that the economics work.