3 Comments
User's avatar
Paresh Dave's avatar

"the GPU power bill will run anyway so any revenue generated in this timezone is better than zero." how so? there's no power draw if there's no inferencing happen, correct?

Sriram Krishnan's avatar

the capacity bill runs anyway, GPu clusters are leased on fixed contracts so eg, you can't dial down a H100 pod at 2am. So any APAC inference revenue is contribution margin against sunk fixed costs, same logic as a half-empty plane still burning near-full fuel

Paresh Dave's avatar

Gotcha that makes sense