Via dwglogo.com
Memory's share of Nvidia's next-gen AI superchip costs jumped from 53% on the prior generation, raising questions about the economics of scaling AI infrastructure.
Building the most powerful AI chips in the world turns out to be a lot like building a luxury home: the most expensive part isn’t the architecture, it’s what goes inside. Goldman Sachs estimates that memory components now account for roughly 62% of the total material costs for Nvidia’s upcoming Vera Rubin superchip, a significant jump from about 53% on the previous-generation GB300 systems.
That nine-percentage-point leap sounds modest until you look at the dollar figures. The estimated bill of materials for a full Vera Rubin NVL72 rack system clocks in at approximately $7.8 million, with memory costs alone approaching $2 million. Some analyses peg the year-on-year increase in memory spending at over 435%.
What’s driving the memory cost surge #
The Vera Rubin platform pairs two types of cutting-edge silicon. On the GPU side, Nvidia’s Rubin chips pack up to 288 GB of HBM4, the latest generation of high-bandwidth memory, capable of delivering 22 TB/s of bandwidth. The GPU itself is a 336 billion transistor design fabricated on TSMC’s 3 nm process.
On the CPU side, the custom Vera processor houses 88 Olympus cores alongside up to 1.5 TB of LPDDR5X memory, connected via memory-coherent NVLink-C2C with 1.2 TB/s of bandwidth.
What this means for the supply chain #
Goldman’s analysis highlights a dynamic that extends well beyond Nvidia’s own balance sheet. When memory represents 62% of a system’s material cost, the fortunes of memory manufacturers become directly tied to the AI infrastructure buildout. Companies like Samsung, SK Hynix, and Micron, the three dominant producers of HBM and LPDDR chips, find themselves in an unusually powerful position.
SK Hynix in particular has been the leading supplier of HBM chips to Nvidia, and the transition to HBM4 represents both an opportunity and a manufacturing challenge. Producing these chips requires advanced packaging capabilities that are in limited supply globally, which creates a natural constraint on how quickly Nvidia can scale production.
The cost pressure is already prompting Nvidia and its partners to explore potential optimizations. These could include adjustments to memory capacities on certain SKUs, more aggressive binning of memory chips, or architectural changes that improve memory utilization so that less raw capacity is needed per unit of compute.
Nvidia announced the Vera Rubin platform in 2025, with production ramping set to begin in 2026. The system is designed specifically for high-performance agentic AI workloads.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our