NVIDIA's upcoming Rubin Ultra series racks are poised to break records in server costs, fueled predominantly by the expense of cutting-edge memory technology. Estimates show that a single Rubin Ultra rack could cost as much as $21 million, largely due to the inclusion of high-bandwidth memory (HBM4e), which alone is projected to account for more than $1.5 million per unit.
The Rubin Ultra rack, internally codenamed "Kyber," features immense hardware densities, including GPUs containing 576 GB of HBM4e memory each. Combined across the full rack, this sums to over 82 TB of HBM4e memory, pushing the price of memory components far beyond previous server generations. This marks a steep rise not just in total cost but also in the price per gigabyte, reflecting broader trends of rising semiconductor and memory prices in the industry.
By comparison, the Rubin standard "Oberon" rack, equipped with 72 GPUs featuring 288 GB of HBM4 memory per GPU, holds roughly 20.7 TB of memory. This HBM4 component alone accounts for approximately $382,000, with prices per gigabyte climbing to $18.40—nearly double the cost observed in earlier Blackwell B200 models which employed HBM3e memory. The total estimated cost for a Rubin "Oberon" rack stands at around $6 million, doubling the price of the Blackwell rack.
Memory is not the only factor in these price increases. The Rubin servers incorporate significant CPU memory as well, with each Vera CPU socket outfitted with 1.5 TB of LPDDR5X memory. Additionally, infrastructure elements including networking, cooling systems, power supply, and high-speed interconnects contribute substantially to the total bill of materials, reflecting the complexity and scale of these AI-focused data center platforms.
This surge in costs underscores the growing investments required in advanced memory subsystems to meet the demands of large-scale AI workloads, particularly those driven by the rapidly evolving GPU architectures NVIDIA is launching. The Rubin series represents a step change in performance but also in the expense associated with cutting-edge server deployments.