Compal put a number on the part of the Vera Rubin story that cannot be rendered in a keynote animation: about 24 kilowatts of system power in a 2U HGX Rubin NVL8 server, with direct liquid cooling attached to the sentence. [1]
That makes this a follow-up to the paper's June 1 account of Nvidia turning token throughput into a business metric. Yesterday's question was whether the AI-factory pitch had enough customer, power, and throughput detail to test it. Today's answer is partial but useful: the constraint moves from the chip to the rack.
Compal says its Rubin NVL8 system supports 400 PFLOPS of NVFP4 compute, up to 2.3TB of GPU memory, 176TB/s of memory bandwidth, and direct liquid cooling for high-density deployment. [1] Nvidia's own technical blog frames Rubin as six chips in one AI supercomputer, pairing GPUs, CPUs, networking, security, and rack-scale design around lower cost per token and higher inference throughput. [2]
The divergence is simple. The market story treats Vera Rubin as a GPU cycle. The operating story treats it as plumbing, power delivery, thermal design, and uptime. If a deployment needs liquid loops and tens of kilowatts per 2U system, the bottleneck is not only whether Nvidia can ship silicon. It is whether data-center shells, power rooms, cooling plants, and facilities teams can absorb the density.
That does not make the supplier page a customer receipt. Compal is describing a system it is showing and building around Nvidia's platform; it is not proving cloud utilization, margins, or adoption. But it is a better receipt than the usual AI-factory adjective because it names a physical burden. A rack-scale platform that promises more throughput per token still arrives as equipment with heat to remove and power to feed.
The practical question for readers is not whether Vera Rubin is impressive. It is whether AI capacity claims now require the same skepticism usually reserved for factories: what voltage, what cooling, what floor plan, what service schedule, what failure mode. Nvidia's platform language says supercomputer. Compal's spec sheet says building operations.
That is where AI governance quietly becomes infrastructure governance. The system can be an agentic factory only after someone builds the room around it. [2]
-- DARA OSEI, London