Skip to main content
TechnologyMay 22, 2026· 2 min read

NVIDIA Vera Rubin: impressive numbers not only for performance, but also for prices

NVIDIA has provided an important update on the Vera Rubin NVL72, a rack system designed for next-generation data centers. The first shipments are expected in the third quarter of the year, while large-scale distribution will begin in the fourth quarter. However, the numbers emerging from the bill of materials for the rack show a significant increase in costs, particularly in terms of memory.

The NVL72 configuration, internally known as Oberon, integrates 72 Rubin GPUs and 36 Vera CPUs. Each tray hosts four Rubin GPUs and two Vera CPUs, while each motherboard includes two GPUs and one CPU in what NVIDIA defines as Superchip configuration. Therefore, a complete rack accommodates 36 Superchips.

From a technical standpoint, each Rubin GPU integrates 288 GB of HBM4 memory, while each Vera CPU has 1.5 TB of LPDDR5X memory. The final result reaches extremely high numbers: 20.7 TB of HBM4 and 54 TB of LPDDR5X within a single NVL72 rack.

According to estimates reported by Morgan Stanley Research, the total cost of the platform reaches approximately $7.8 million. The most expensive component remains the GPU, with a value close to $4 million, approximately $55,000 per individual Rubin GPU. The comparison with the Blackwell NVL72 B300 generation highlights a cost increase of 57% compared to the previous platform's estimated price of around $2.5 million.

However, the most pronounced percentage increase is recorded in memory. The technologies HBM4 and LPDDR5X increasingly affect the final price due to limited availability and rising demand in the AI sector. The cost of memory exceeds $2 million, with a growth of 435% compared to the Grace Blackwell platform, where the value was around $374,000.

Memory now represents 26% of the total system cost, indicating how much new AI architectures depend on high-bandwidth and high-capacity solutions. Other components also show significant increases. For example, PCBs rise from $35,100 in the Blackwell platform to $116,730 in the Rubin configuration, marking a 233% increase.

The Vera CPUs contribute about $180,000 to the overall rack cost, equivalent to about $5,000 per chip. The remaining part of the bill of materials includes NVLink switches, networking, power supply, cooling systems, interconnections, ABF substrates, and other essential components to support the consumption and computational density of the platform.

In conclusion, the generational leap between Vera Rubin and Grace Blackwell is nothing short of obvious, but at the same time, the cost has significantly skyrocketed. Certainly, proportionally, Vera Rubin appears more cost-effective since NVIDIA suggests performance per watt up to 10 times higher, but the price for computing power—also in terms of consumption, almost doubled with Vera Rubin—will have a considerable impact on the overall calculation.