Nvidia’s next-generation VR200 NVL72 rack now costs hyperscalers $7.8 million, driven largely by a 435% surge in memory costs that now account for 25% of total system expenditure.
Cost Escalation
Morgan Stanley Research estimates that each VR200 NVL72 rack—built around Nvidia’s forthcoming Rubin GPUs and Vera CPUs—costs major cloud service providers approximately $7.8 million. This represents a near-doubling from the roughly $4 million price tag of the previous-generation GB300 NVL72 system. The increase stems not only from higher GPU and CPU pricing but from a dramatic rise in memory content. Nvidia is charging $55,000 per Rubin GPU and $5,000 per Vera CPU in volume, while the Oberon chassis now incorporates more sophisticated switching, networking, PCB, cooling, power, and packaging components, all inflating the bill of materials.
Memory Dominance
Memory—including DRAM, NAND, and HBM—now constitutes roughly $2 million of each VR200 NVL72 system, up 435% from the GB300 NVL72. This shift is driven by two factors. First, each rack packs 54 TB of LPDDR5X memory, a threefold increase over the 17 TB in the GB200 NVL72. At an estimated $8 per GB from Nvidia’s suppliers, that alone accounts for $408,000; if prices rise to $10 per GB, the LPDDR5X content reaches $540,000, excluding Nvidia’s markup. Second, each VR200 NVL72 now includes approximately $1 million in 3D NAND storage, compared to virtually none in prior generations. HBM4 memory onboard the Rubin GPUs adds further cost.
Supply and Pricing Pressures
Memory prices are under upward pressure from surging AI demand. SemiAnalysis estimates Nvidia paid $8 per GB for LPDDR5X in Q1 2026, but that figure may climb as demand intensifies, particularly for SOCAMM2 modules—expensive to manufacture and test—used exclusively by Vera CPUs. For context, contract pricing for DDR5 memory currently ranges from $12 to $16 per GB, with spot prices averaging $20 per GB. LPDDR5X is inherently more costly, and Nvidia’s markup further elevates the total. Memory manufacturers have resisted adding capacity, suggesting the current pricing environment may persist.
Strategic Implications
The rising memory cost share signals a structural shift in AI infrastructure economics. Hyperscalers must now budget for memory as a primary cost driver, not a peripheral component. As Nvidia pushes toward denser, higher-bandwidth memory configurations to support increasingly complex models, the total cost of ownership for AI racks will continue to climb. For enterprises and cloud providers, this underscores the importance of optimizing memory utilization and exploring alternative architectures—before memory becomes the single largest line item in their data center budgets.
