Price Pressure Unbearable! NVIDIA Slashes CPU Memory in Vera Rubin NVL72 by Half

According to the latest foreign media reports, LPDDR5X supply remains tight and prices continue to surge—soaring even beyond the reach of long-contracted players like NVIDIA.

China's 'Kuaikeji' reports that NVIDIA's next-generation Vera Rubin NVL72 rack-scale AI system will drastically reduce CPU-side system memory: SOCAMM modules will be halved from 192GB to 96GB, cutting the total Vera CPU memory from 54–55TB down to 28TB. All CPU racks will be uniformly equipped with 96GB SOCAMM. Meanwhile, GPU-side HBM4 memory per cabinet remains unchanged at 20.7TB.

Bernstein estimates that due to rising prices of both LPDDR5X and HBM4, the per-cabinet manufacturing cost has increased from Morgan Stanley’s earlier projection of $7.8 million to $9.1 million. HBM4 prices could reach $53 per GB by 2027. Without specification changes, memory costs for Vera Rubin 200 would account for approximately 29% of the total bill of materials (BOM), around $2.1 million—far exceeding the comfortable 20% threshold.

Halving the capacity would reduce LPDDR5X costs from $1.2 million to $586,000; further reducing to a quarter would bring it down to $293,000, significantly easing BOM pressure per cabinet.

NVIDIA unveiled the Vera Rubin NVL72 in January this year. CoreWeave testing showed a throughput of 800,000 tokens per second per megawatt when processing MoE workloads—10 times higher than the Blackwell cabinet’s 80,000 tokens per second.

The report also notes that the flagship computing system is reducing only CPU memory while maintaining GPU memory, effectively signaling that memory price inflation has now penetrated even the specifications of the most powerful AI cabinets.