Goldman Sachs Reports Memory Costs Drive Up Nvidia's Vera Rubin Superchip Expenses

Goldman Sachs has estimated that memory components now represent approximately 62% of the total material costs for Nvidia’s forthcoming Vera Rubin superchip, a notable increase from 53% in the previous-generation GB300 systems. This shift underscores the rising importance of memory in the overall cost structure of advanced AI chips.

The estimated bill of materials for a complete Vera Rubin NVL72 rack system is around $7.8 million, with memory costs alone nearing $2 million. This represents a staggering year-on-year increase in memory spending of over 435%.

Drivers of Memory Cost Surge

The Vera Rubin platform utilizes two types of cutting-edge silicon. On the GPU side, Nvidia’s Rubin chips include up to 288 GB of HBM4, the latest generation of high-bandwidth memory, which can deliver 22 TB/s of bandwidth. The GPU is built on a 336 billion transistor design using TSMC’s 3 nm process. The CPU features a custom Vera processor with 88 Olympus cores and up to 1.5 TB of LPDDR5X memory, connected via memory-coherent NVLink-C2C with 1.2 TB/s of bandwidth.

Impact on Supply Chain

This analysis by Goldman Sachs highlights a broader trend impacting the supply chain. As memory accounts for such a significant portion of a system’s material costs, the fortunes of memory manufacturers are increasingly tied to the growth of AI infrastructure. Major producers like Samsung, SK Hynix, and Micron, which supply HBM and LPDDR chips, are positioned advantageously but face challenges in scaling production due to the advanced packaging requirements for HBM4.

Nvidia is exploring various optimizations to mitigate cost pressures, including adjustments to memory capacities and architectural changes to enhance memory utilization. The Vera Rubin platform is set to be announced in 2025, with production ramping expected to begin in 2026, specifically targeting high-performance AI workloads.