[News] NVIDIA Reportedly Halves Vera Rubin SOCAMM Capacity as Memory Costs Near 29% of System BOM

Rising memory costs are prompting AI hardware vendors to rethink their product strategies. According to Wccftech, citing analysis from GF Securities, NVIDIA is reportedly reducing the memory capacity of its Vera Rubin NVL72 rack-scale AI system to cope with elevated memory prices and ongoing supply shortages. Without these adjustments, memory could account for around 29% of the Vera Rubin VR200 system’s estimated US$2.1 million bill of materials (BOM), well above NVIDIA’s preferred level of 20%.

The cost burden is significant. Wccftech, citing an earlier Bernstein report, notes that a single Vera Rubin NVL72 rack could cost as much as US$9.1 million, with memory accounting for a significant share of the cost. Bernstein also forecasts HBM4 pricing to rise to US$53 per GB by 2027.

To lower costs and ease supply constraints, NVIDIA may adopt 96GB SOCAMM modules for each Vera CPU in its Vera Rubin NVL72 racks, down from the previously planned 192GB, the report says. As a result, total Vera CPU memory would decline from about 55TB to 28TB, while GPU HBM4 capacity would remain unchanged at 20.7TB per rack.

The revised memory configuration could deliver significant savings. As noted by Wccftech, GF Securities originally estimated LPDDR5X costs for the VR200 at around US$1.2 million based on the original memory configuration. Under the revised configuration, the cost is estimated to fall to around US$586,000 and could decline further to as low as US$293,000 if memory capacity is reduced to one-quarter.

According to a press release released by TrendForce in early June, based on preliminary allocation plans from Samsung, SK hynix, and Micron, NVIDIA is expected to receive enough LPDRAM to meet only about 60% of its estimated demand, with limited room for further increases. As a result, the company may reduce the SOCAMM memory configuration of its Vera Rubin Superchip platform to increase Vera CPU production and mitigate the risk of prolonged supply shortages.

Strengthening Long-Term Memory Partnerships

As NVIDIA navigates these cost and supply challenges, it is also strengthening its long-term partnerships with South Korea’s memory industry. According to Tom’s Hardware, NVIDIA and SK Group signed letters of intent formalizing a strategic partnership valued at more than US$500 billion. The collaboration spans multiple areas, including a long-term memory supply agreement with SK hynix that was announced in June.

Read more

• NVIDIA Cuts Vera CPU Memory Configuration, Highlighting Persistent LPDRAM Supply Constraints and Rising Long-Term Demand, Says TrendForce
• [News] AMD’s Helios AI Rack Challenges NVIDIA with HBM4 Memory Edge; Microsoft Joins as Latest Customer

(Photo credit: NVIDIA)