AI's unit of account just moved from the chip to the rack

The unit of account for AI infrastructure is moving from the chip to the rack, and the bill now arrives in millions of dollars per box rather than thousands per GPU. Microsoft just took the first production delivery of NVIDIA's next-generation Vera Rubin AI rack system, a generation past Blackwell, and the shift is no longer a roadmap slide. Bajarin's projection frames it directly: rack-scale compute sits below 10% of the total accelerator installed base today, and he expects 35–40% within a few years. That is the move worth naming.

One rack is now a complete AI supercomputer, not a server with parts. The NVL72 box holds 72 Rubin GPUs and 36 Vera CPUs, pulls power at 800 volts DC, requires 100% liquid cooling, weighs roughly 4,000 pounds, and costs $7 million to $8 million (per NVIDIA via memeburn). Hyperscalers used to buy chips and assemble the rest. They now buy the rack and the room around it, and the room has to be built to its spec. Power, cooling, weight, and floor load become procurement variables, not afterthoughts.

Most readers will read the Microsoft news as a vendor customer win. The pattern underneath is a buying unit: capacity contracts, depreciation schedules, and cloud pricing will all start to be quoted per rack. The chip is no longer the SKU.