Rubin Ultra Cuts Specs, Can Nvidia Handle Rising Memory Prices?

Original | Odaily Planet Daily (++@OdailyChina++)

Author | Azuma (++@azuma_eth++)

This past weekend, a report from the well-known research firm SemiAnalysis sparked significant discussion across the AI industry.

The core content of the report states, after SemiAnalysis revealed at the end of June that Nvidia's (NVIDIA) original 4-die Rubin Ultra design would be halved, the research firm has now disclosed that Nvidia has provided a preview of Rubin Ultra to major clients, but its specifications have further declined from previous expectations.

Screenshots of the Rubin Ultra-related information leaked in the SemiAnalysis report are as follows:

• Rubin Ultra will maintain the same theoretical peak performance as Rubin, both at 35 PFLOPs;
• The memory capacity has been severely reduced, with specifications dropping to 192GB with 8-layer stacking (8-Hi), even lower than Rubin's 288GB with 12-layer stacking (12-Hi);
• The change in memory bandwidth is negligible, only increasing by 1 TB/s, which can essentially be ignored in actual high-throughput calculations;
• Chip-level power consumption has even slightly increased, with the minimum power consumption the same as Rubin (1800 W), but the maximum power consumption is higher (2600 W);
• The only significant improvement is the "scale-up world size," which has increased from 72 GPUs to 576 GPUs, meaning Nvidia has shifted the real selling point to cluster networking, allowing up to 576 Rubin Ultra GPUs to be connected into a massive "super logical GPU" (Rubin only supports interconnection of up to 72 cards).

Originally positioned as Nvidia's top flagship version announced at GTC 2026, Rubin Ultra was designed to integrate more chip dies and high-bandwidth memory to meet the extreme scale of AI model training and inference needs. However, based on the latest specification details disclosed by SemiAnalysis, it is clear that Nvidia has adjusted the design concept of Rubin Ultra.

Rapid HBM Price Increases Force Nvidia to Reassess

Over the past two years, one of the most critical components in the expansion of the AI industry has undoubtedly been HBM.

With the explosive demand for AI accelerators like Nvidia H100, H200, and Blackwell, high-bandwidth memory has transitioned from a relatively niche high-end storage product to the most scarce component in the entire AI infrastructure. SK Hynix, Samsung, and Micron have been continuously expanding their HBM investments while also refreshing their performance, but the imbalance between supply and demand continues to drive HBM prices upward.

For AI chip manufacturers, the importance of HBM is self-evident—GPUs handle computation, while HBM provides high-speed data throughput, and together they determine the efficiency of AI model training and inference. However, the problem is that HBM has become so expensive that it is starting to affect the overall economics of AI systems. For example, the price of HBM3 was only $180-220 at its low point in Q2 2025, but it has risen to $600-700 in Q1 of this year (contract price), and even higher to $700-850 in Q2 (spot price).

According to estimates from SemiAnalysis, as HBM prices rise, the pure material (BOM) cost of a single Rubin Ultra rack has increased from about $6.6 million to $8 million, but after adjusting the design specifications, the cost can drop to about $6.4 million.

For Nvidia, such a significant cost difference raises a very realistic question—if they continue to increase HBM capacity according to the original design, will it still bring performance improvements that match the costs? Are there other better options?

SemiAnalysis provided an answer in the report, stating that the adjustments to Rubin Ultra essentially represent Nvidia optimizing the cost structure of the AI system under limited resources, reducing the relatively expensive HBM configuration (cost share dropping from nearly 40% to 28%) and reallocating resources to higher-value scaling interconnect capabilities (cost share increasing from 4% to 12%).

As mentioned earlier, the core upgrade direction of Rubin Ultra has now shifted to "system-level scaling interconnect capabilities." The NVL576 architecture supported by Rubin Ultra can connect up to 576 GPUs into a unified computing domain via NVLink, aiming to compensate for adjustments in single-chip specifications with larger-scale system expansions.

Storage Stocks Plummet, Market Worries Demand Peaks

In response to this news, storage-related stocks collectively declined after the opening of the Korean stock market this morning. As of 11:45 AM Beijing time, SK Hynix and Samsung, the two leading HBM companies, both fell by about 8%, and the KOSPI index also dropped by about 5%.

The market has begun to worry that if the specification adjustments for Rubin Ultra disclosed by SemiAnalysis are true (Nvidia has not publicly confirmed this information), does it mean that Nvidia, as the core buyer in AI infrastructure, is reducing its demand for HBM?

Over the past two years, with the rapid growth of AI expansion demands, HBM has become the most scarce part of AI infrastructure. Chip manufacturers like Nvidia and AMD have continuously increased HBM configurations in AI accelerators, driving the performance of storage companies like SK Hynix, Samsung, and Micron to soar, while also strengthening their pricing power along the entire supply chain.

Nvidia's choice may indicate that AI chip manufacturers are considering optimizing hardware designs to reduce reliance on high-capacity HBM for individual chips. If this path proves feasible, the room for HBM manufacturers to continue raising prices is likely to be limited.

The construction of AI infrastructure will eventually move away from the era of "crazy stacking and mindless price increases," and now, even Nvidia, sitting atop the throne of computing power, has begun to tighten its budget.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.