3x cheaper per FLOP
The racks got cheaper to run.
The racks got cheaper to run.
So they run harder.
This is Jevons paradox for GPUs: cut the cost per FLOP and demand does not shrink to match, it expands past it. Every rack that got cheaper just got rented out twice as long.
NODE_ID
0xJEV_RACK04_STILL_SPOOLING
Compute consumed, cluster-wide
UTILIZATION 340%
utilization only climbs from here