Nvidia's power pitch: 40,000 Rubin GPUs in 100MW, or 2.5kW a GPU at the meter

Divide the two headline numbers and you get the figure an operator will actually plan against: 2.5 kW of site power for every Rubin GPU installed ^[11], or 2.5 MW per SuperPOD ^[12]. That is measured at the meter, so it has to cover cooling, networking, conversion losses and everything else in the shell, not just the accelerator. Run the same division on the performance claim and the site delivers roughly 20 TFLOPS per watt for NVFP4 inference and 14 for training ^[10]. Those are the numbers to hold Nvidia to, and Tom's Hardware has already said the FLOPS half is estimated rather than measured, with real workloads landing significantly lower ^[1].

The measured part of the pitch is smaller and more useful. In Nvidia's GB300 example, provisioning statically against a 135 kW peak fills a 540 kW budget with four racks and leaves as much as 170 kW unspent ^[4], which is 31 percent of the allocation sitting behind a breaker doing nothing ^[13]. Reclaim it and a fifth rack fits inside the same budget ^[5]. That is 25 percent more installed compute per contracted megawatt ^[14], at an average allocation of 108 kW per rack rather than 135 ^[15].

Where that 25 percent comes from decides whether it travels. The control loop harvests the gap between what a group of racks could draw and what it does draw, watching power at chip, rack and group-of-racks level and moving slack to whichever system needs it ^[6]. Nvidia's own description of the problem is a load differential between racks, one busy and one not ^[7]. That differential is a property of the workload mix, not of the silicon. A hall running one synchronous training job across every rack has less to redistribute than a mixed inference fleet with idle capacity, and the rack-level profiles Nvidia ships for inference, training and memory-bound work ^[8] are an admission that the answer changes by job type.

Then there is the reserve. The five-rack configuration works partly because much less power is held back for peaks ^[5]. Static provisioning was wasteful, and it also failed safe: that stranded 170 kW was, incidentally, insurance. Dynamic allocation converts the insurance into inventory and moves the guard band into software. That is a defensible trade for an operator who can measure it. Note also what DSX stands for, which is Land, Power, Shell ^[2], and what that implies about where Nvidia now expects to be consulted.