Where Facility Meets Computer
Liquid cooling moved the facility to the chip. With little thermal inertia at the cold plate, a coolant shortfall shows up as throttled accelerators and lost throughput, not as a facility alarm. That turns coolant temperature and electrical headroom into scheduling inputs. The vendors have noticed. Vertiv, Schneider Electric and a long list of partners now build to NVIDIA reference designs, and NVIDIA's March 2026 Vera Rubin DSX design names a component whose whole purpose is to pass signals between IT and operational technology. Faraday's read: the reference design is now the unit of competition, and an operator should know exactly which layers of its control plane it owns and which it has borrowed from a supplier.
Knowledge check
Why does a liquid-cooled AI hall make thermal state a scheduling problem rather than only a facility problem?
Show the answer
At the cold plate a liquid loop has little thermal inertia, so heat from a synchronized ramp arrives almost immediately. If flow or temperature is inadequate, chips throttle and the job slows. A scheduler that knows thermal headroom can place work where it will run at full speed, which is why coolant and computation are better controlled together.