Nvidia’s Vera Rubin rack bill shows memory prices going bananas as AI hardware turns into a very expensive arms race.
Nvidia’s bill of materials for its coming Vera Rubin NVL72 rack shows a huge surge in memory prices. Memory now makes up 26 per cent of the total system cost, which is quite the slap for anyone pretending AI infrastructure is getting cheaper.
Rising demand and tight supply have pushed memory costs sharply higher for Vera Rubin compared with Grace Blackwell racks. The memory share rises to 26 per cent on Vera Rubin, compared with nine per cent on Grace Blackwell.
Vera Rubin is in production and is confirmed for first shipments in the third quarter of 2026, with volume ramping in the fourth quarter. While Nvidia cooks up its grandest AI platform so far, the price tag is getting suitably ridiculous.
Morgan Stanley Research has shared an estimated bill of materials for Nvidia’s Vera Rubin VR200 NVL72 rack. The rack will feature 72 Vera Rubin GPUs, with each tray housing four Rubin GPUs and two Vera CPUs. Nvidia’s NVL72 rack is called Oberon and uses 72 GPUs. A single Vera Rubin tray holds four Rubin GPUs and two Vera CPUs.
Two GPUs and one CPU sit on a motherboard, which Nvidia calls a Superchip. There are 36 Superchips in the NVL72 rack, giving it 72 GPUs and 36 CPUs. Each Rubin GPU carries 288GB of HBM4 memory, while each Vera CPU comes with 1.5TB of LPDDR5X memory.
That means an NVL72 rack has 20.7TB of HBM4 and 54TB of LPDDR5X. Plenty more goes into the rack, including networking, cooling, power, interconnects and the usual costly plumbing needed to keep AI kit alive.
Morgan Stanley’s estimates suggest the Rubin GPU is expected to be the highest single-cost component, at nearly $4 million per rack.
That is a 57 per cent jump over the Blackwell NVL72 B300 rack, where the GPU cost was about $2.5 million. It works out at about $55,000 per GPU. The second-highest cost is memory, which should surprise nobody watching LPDDR and HBM shortages turn into an industry-wide mugging.
With Rubin, memory costs rise 435 per cent, from $373,939 in Grace Blackwell to more than $2 million on Vera Rubin. That figure combines the costs of LPDDR5X and HBM4 memory.
The Vera CPUs account for $180,000 of the total, putting each chip at roughly $5,000. Together, GPUs, CPUs, and memory push the rack cost into nosebleed territory before the rest of the system is factored in.
The total bill of materials is estimated at $7.8 million, driven by pricier GPUs and the daft memory bill. The remaining costs include NVLink switches, networking chips, cooling, power supplies, PCBs, ABF substrates, MLCCs and other bits.
PCBs see the second biggest jump, rising 233 per cent from $35,100 in Blackwell to $116,730 on Rubin. Nvidia’s Vera Rubin NVL72 rack marks a beefy new chapter in AI infrastructure, but it comes with a grand price.
Memory now accounts for more than a quarter of the whole system cost, showing how HBM4 and LPDDR5X supply pressure is warping the AI buildout. As production ramps for shipments in the third quarter of 2026, Vera Rubin is set to deliver silly levels of performance at an equally crazy price.
