Nvidia Vera Rubin VR200 Rack Costs $7.8 Million Per Unit as AI Infrastructure Spending Hits Absurd Scale

[RusticDaemon]

Morgan Stanley Research published analysis this week showing that Nvidia's next-generation Vera Rubin-based VR200 NVL72 rack will cost hyperscale cloud providers approximately 7.8 million dollars per unit, up from roughly 4 million dollars for the prior GB300 generation. Memory now accounts for around 25 percent of total system cost, or about 2 million dollars per rack, driven by a threefold increase in LPDDR5X content and around 1 million dollars in 3D NAND storage. Each individual Rubin GPU is priced at approximately 55,000 dollars for volume hyperscaler purchases.

These numbers provide useful context for understanding why the Micron earnings and SK Hynix listing are happening now. When a single GPU rack costs 7.8 million dollars and hyperscalers like Microsoft and Google are spending 175 to 190 billion dollars on AI infrastructure this year, the demand for memory components at the top of the performance curve is essentially unlimited. The 25 percent memory cost share is also a meaningful shift. Previous generations had memory at 10 to 15 percent of rack cost. The Vera Rubin architecture is fundamentally more memory-intensive to deliver the performance improvements it promises.

For anyone trying to understand why the memory chip market has been so extreme, this is the structural explanation. AI training and inference at frontier scale requires bandwidth and capacity that pushed the state of the art for HBM in every successive generation. The three Nvidia GPU generations from Ampere to Hopper to Blackwell to now Vera Rubin have each approximately doubled memory bandwidth requirements, which flows directly into SK Hynix, Micron and Samsung's order books.

Nvidiawww.tomshardware.com