Nvidia reduces specifications on Rubin Ultra GPU due to high HBM memory prices, indicating even flagship products constrained by memory cost and availability
Over the past weekend, an institutional report from investment research firm SemiAnalysis sparked widespread discussion across the AI industry. Following the firm's late-June disclosure that Nvidia's original 4-die Rubin Ultra design would be scaled down, SemiAnalysis has now revealed that Nvidia provided key customers with a Rubin Ultra preview featuring further specification reductions from earlier expectations.
Rubin Ultra was positioned as Nvidia's flagship processor, announced at GTC 2026 to handle extreme-scale AI model training and inference by integrating multiple dies and high-bandwidth memory. According to SemiAnalysis's latest disclosure, Nvidia has adjusted its design philosophy for the product.
Over the past two years, high-bandwidth memory has become the most critical component in AI infrastructure expansion. With explosive demand for Nvidia's H100, H200, and Blackwell accelerators, HBM evolved from a niche product into the most constrained link in the entire supply chain. SK Hynix, Samsung, and Micron have all expanded HBM investments while continuously breaking revenue records, yet supply-demand imbalances continue driving prices higher.
For AI chip manufacturers, HBM's importance is paramount—while GPUs handle computation, HBM provides the high-speed data throughput that determines the efficiency of AI model training and inference. However, HBM has become so expensive that it now affects the overall economics of AI systems. HBM3 pricing illustrates this surge: a single module cost $180–220 in Q2 2025, rose to $600–700 by Q1 2026 (contract price), and reached $700–850 in Q2 2026 (spot price).
According to SemiAnalysis, as HBM prices climbed, the bill-of-materials cost for a single Rubin Ultra rack rose from approximately $6.6 million to $8 million. After the design adjustments, Nvidia can bring this back down to roughly $6.4 million.
The specifications of the redesigned Rubin Ultra tell the story: peak compute remains 35 PFLOPs, but memory is cut to 192GB—lower than the original Rubin's 288GB—with bandwidth improving by only 1 TB/s, delivering limited performance gains. More significantly, HBM's cost share of the total drops from nearly 40% to 28%, while interconnect cost share rises from 4% to 12%.
The strategic shift is toward system-level interconnect capabilities. The NVL576 architecture supported by Rubin Ultra can connect up to 576 GPUs via NVLink into a unified compute domain—eight times the scale of the standard 72-card configuration—designed to compensate for reduced individual chip specifications through larger-scale system expansion.
The market reacted immediately. Korean memory stocks collapsed at open this morning: as of 11:45 Beijing time, SK Hynix and Samsung—the two HBM leaders—both fell approximately 8%, while the KOSPI index dropped about 5%.
The concern is clear: if Nvidia's Rubin Ultra adjustments are confirmed (Nvidia has not yet publicly acknowledged the information), they could signal that Nvidia, as the most pivotal buyer in AI infrastructure, is reducing HBM demand. Over the past two years, HBM vendors have benefited from AI expansion as chipmakers continuously increased HBM configurations, driving explosive revenue growth and strengthening vendors' pricing power throughout the supply chain.
Nvidia's choice may signal that AI chip manufacturers are considering reducing individual chips' reliance on high-capacity HBM through optimized hardware design. If this approach proves viable, the room for continued price increases by HBM vendors could become significantly constrained.
AI infrastructure construction is inevitably moving past the era of indiscriminate spec expansion and unfettered price increases. Even Nvidia—commanding the computing power landscape—has begun counting every penny.