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

[Anthony]

7.8 million per rack times the number of racks in a gigawatt data centre is a number I need a spreadsheet to even approach. The capital density of modern AI infrastructure is genuinely hard to comprehend

[DotEXE]

Memory going from 10-15 percent to 25 percent of rack cost is the data point that explains everything about SK Hynix's stock price this year. They own the memory that makes these systems work

[Tel86]

At 55,000 dollars per Rubin GPU in volume purchases I would love to know what the retail equivalent is. Nvidia's margins on this product line must be extraordinary given that TSMC advanced node wafers cost roughly 25 to 30 thousand each

[ReacherBadger]

The 2 million dollars in memory per rack explains why Micron and SK Hynix are both in IPO mode right now. Their customers are spending amounts of money on their products that simply could not have been projected two years ago

[Ronan_34]

LPDDR5X in server racks is interesting architecturally. That is mobile-derived memory technology being deployed in hyperscale data centres because the bandwidth density is higher than traditional server DRAM. The lines between segments are dissolving

[Leopard85]

The price per rack doubling from GB300 to VR200 while capabilities presumably improve significantly means the total spend at hyperscalers has to either scale with it or something gets rationed. Something always gets rationed eventually

[Amy96]

I want to understand the power consumption per rack on Vera Rubin. If the price doubled the power requirements probably went up significantly too. Data centre power infrastructure is already the binding constraint for most new builds