The Stack of Controls
The control plane of an AI factory is not a product. It is a stack assembled by two industries that rarely spoke: facilities engineers with building management and power monitoring systems, and computer scientists with schedulers such as Slurm and Kubernetes. Accelerator telemetry, typified by NVIDIA's Data Center GPU Manager, now sits between them. Each layer works. The trouble is the border. A scheduler that cannot see the cooling plant will place work where it cannot run at full speed, and a facility that cannot see the job queue will react instead of prepare. Faraday's read: the next gains come from the seams, not the layers, and whoever defines the interface between them will shape the industry.
Knowledge check
Why did the traditional arrangement between facility teams and IT teams, a fixed power and cooling promise per rack, begin to fail with AI workloads?
Show the answer
AI loads are dense, synchronized and expensive. Density turns small cooling errors into large thermal excursions, synchronization makes thousands of chips change power draw together, and cost makes every idle minute a measurable loss. Each property requires information to cross the old border between facility and IT systems.