NVIDIA VRNVL72 Reportedly "Downgraded": SOCAMM Memory Capacity May Be Halved

According to foreign media reports, the latest analysis from SemiAnalysis indicates that NVIDIA is adjusting the memory configuration of its next-generation Vera Rubin NVL72 rack system, potentially halving the Vera CPU's originally planned 192GB SOCAMM solution to 96GB.

Capacity Halved: Per-Rack SOCAMM DRAM Drops from 55TB to 28TB

The original Vera Rubin NVL72 included 36 Vera CPUs, each equipped with eight 192GB SOCAMM LPDDR5X modules, bringing the total SOCAMM capacity per rack to approximately 55TB. However, according to the latest SemiAnalysis report, the specifications may be revised to 36 Vera CPUs each with eight 96GB SOCAMM modules, reducing per-rack SOCAMM capacity to about 28TB — a reduction of nearly half.

Cost Reduced: Rack Cost Falls from $7.6 Million to $6.8 Million

This memory configuration adjustment brings significant cost reductions. The report shows that the estimated rack cost lowers from $7.6 million to $6.8 million, saving $800,000 per rack with a 10% cost reduction.  Total cost of ownership reduces from $4.16/hour/GPU to $3.90/hr/GPU. For cloud providers and AI enterprises planning large-scale deployments, this capital expenditure saving is meaningful. Lower customer deployment barriers are expected to drive overall system shipments and accelerate the adoption of Vera Rubin.

It is worth noting that if the "SOCAMM capacity halving" becomes reality, the training efficiency of some ultra-large models with over a trillion parameters or MoE architectures may be somewhat affected. 
It remains to be tested by the market whether customers are willing to trade performance for cost advantages.

Following this news, U.S. memory-related stocks fell overnight, while NVIDIA closed higher. At market close, Micron dropped over 7%, Sandisk and WD fell more than 3%, Seagate declined over 1%, while NVIDIA rose more than 1%.