Nvidia despecs Rubin Ultra memory as HBM hits 40% of rack cost
- Source
- x.com
- Date
Nvidia has despecced Rubin Ultra's HBM: from HBM4E 12-Hi (384GB) down to HBM4 8-Hi (192GB). Why strip memory out of your flagship rack-scale system? Because after the latest HBM and DRAM price hikes, memory had quietly become ~40% of total capital cost of ownership. We ran the TCO math in a recent conference keynote (1/3)🧵
The despec cuts HBM cost by >50% — even with the 2026 HBM price hike already baked in — and takes memory from ~40% to ~28% of total capital cost. But that spend doesn't vanish. It's redirected into scale-up networking: on the NVL576 NPO SKU, scale-up triples from 4% to 12% of rack spend as optics take over rack-to-rack interconnect. (2/3)
The takeaway: despec is a symptom of the supply crunch, not weakening demand. When the industry's largest, best-positioned buyer is stripping memory content just to manage cost and availability, pricing power sits firmly with suppliers — and the shortage is broadening, not easing. (3/3)
Context
SemiAnalysis says NVIDIA has cut the memory planned for its Rubin Ultra system from 384GB to 192GB of high-bandwidth memory (HBM), the fast memory stacked next to the processor, and estimates that this changes where the cost of a rack goes. In its table, memory (HBM, DRAM and NAND combined) is 41% of capital cost of ownership in the 384GB design and 28% in the 192GB design, while scale-up networking, the links that tie accelerators together into one larger system, rises from 4% to 12%. Analysis and charts.
The two columns are different rack designs, not only a memory change: the table compares an NVL72 rack with HBM4E 12-Hi stacks against an NVL572 rack with HBM4 8-Hi stacks, which the thread elsewhere labels NVL576. SemiAnalysis argues that the saved memory spend is redirected into networking. These are the analyst's modeled costs, so they illustrate a spending tradeoff rather than establish a vendor's quoted system price.
Checking sign-in…
Loading comments…