NVIDIA Rubin Ultra Allegedly Downgraded: HBM Shortage Stalls AI Computing Power, Flagship Chips Cannot Remain Low

The AI computing arms race is colliding with the toughest piece of the industry's supply chain. According to the latest report from the US tech media The Information, Nvidia is evaluating a rather aggressive adjustment plan—to launch a version of the next-generation flagship AI GPU Rubin Ultra with a lower HBM (High Bandwidth Memory) configuration than originally planned, to alleviate production pressure from the shortage of high-end HBM. This means that even Nvidia, which holds absolute dominance in the GPU market, has to compromise between product specifications and supply capabilities. News of this quickly led to a vote by the market: on Thursday local time, Nvidia's stock price fell by 0.1%, while HBM giant SK Hynix stock price plummeted by 4.97%. ## I. Flagship Chips Forced to "Shrink": Multiple Downgraded Versions Under Testing Sources indicate that over the past few weeks, Nvidia has tested at least three different versions of the Rubin Ultra GPU, some of which utilize lower memory capacity than initially planned. According to the latest industry survey by TrendForce (CB Insights), Nvidia had originally planned to use HBM4e 12Hi memory across the entire Rubin Ultra lineup, but is now evaluating lower-spec options such as HBM4e 8Hi and HBM4 12Hi, HBM4 8Hi, and other low-spec configurations. The cost of downsizing is directly reflected in bandwidth. One of the core goals for the Rubin Ultra is to achieve a breakthrough in interface bandwidth—if the top-spec configuration is maintained, its I/O speed can jump from 8-11.7Gbps in the previous generation to 14-16Gbps. If forced to use the lower-spec HBM4 configuration, the speed increase might only be maintained around 11-12Gbps. For large-scale AI training that seeks ultimate computing power, this is undoubtedly a significant concession. In fact, this is not the first time Nvidia has "reduced" in terms of storage. Previously, due to the forecast that LPDDR5X memory shortages will continue until 2027, Nvidia decided to halve the capacity of the SOCAMM in the Vera Rubin super chip module. Memory shortages are reshaping Nvidia's product roadmap across consumer to data center levels. ## II. Why HBM Has Become the "Bottleneck" HBM has become one of the most scarce core components in the current AI industry chain. The rapidly growing demand for large model training is driving continuous increases in GPU shipments, while the HBM required by GPUs is supplied by a small number of manufacturers such as Samsung Electronics, SK Hynix, and Micron. Due to the complexity of HBM manufacturing processes, high yield requirements, and long expansion cycles, the rate of supply growth has consistently failed to keep up with the explosive growth in AI demand. The TrendForce report points out that although HBM shipment capacity is expected to grow by 50% to 60% year-over-year in 2027, this growth still cannot fill the explosive demand gap in the AI computing market. More critically, the most advanced HBM4e 12Hi solution still has uncertainties regarding validation progress and mass production yield. With multiple factors compounding, Nvidia is forced to prepare alternative solutions to counter supply risks. Industry insiders generally anticipate that HBM will remain in a supply-demand imbalance in the coming years, which is one of the major reasons for the persistently high prices of AI servers. Nvidia's decision to consider lowering memory configurations this time clearly demonstrates that even with the strongest bargaining power in the supply chain, product planning is still constrained by upstream storage supply—this is a reality of the industry that must be recognized when understanding the current AI market. ## III. Cascading Effects of Downsizing: Cost Pressure Spreads Downstream For AI training, HBM not only determines the data throughput capacity of GPUs but also directly impacts the scale of large models that a single card can accommodate. If Rubin Ultra ultimately adopts lower memory configurations, customers may need to deploy more GPUs to complete computational tasks that could have been finished with fewer chips. Although Nvidia can still partially compensate for performance losses by increasing GPU computing power and optimizing interconnect bandwidth, system deployment costs and cluster complexity will still increase. For large cloud vendors like Microsoft, Meta, Amazon, and Google, who are continuously increasing AI capital expenditures, this means that the cost of future data center construction could rise further. In fact, the "grabbing" of AI infrastructure is accelerating: media reports indicate that AI star company Anthropic just signed a $10 billion, six-year compute procurement agreement with infrastructure provider Volta Infra Holdings. The frenzy on the demand side and the tightness on the supply side are creating a self-reinforcing price spiral. The impact of HBM price increases also spills over beyond AI servers. As key components like GPUs and HBM remain scarce, hardware costs across the entire technology industry are rising. Some hardware manufacturers, including Apple, are absorbing supply chain cost pressures by increasing the price of end products—the battle for high-end storage is evolving into a cost redistribution affecting the entire industry. ## IV. The "Three Kingdoms Kill" in the Storage Market Changes: Landscape is Being Reshaped In this storage cycle driven by AI demand, the seating of suppliers is quietly changing. The Counterpoint Research report on Global Storage Tracking shows that Samsung Electronics captured 39% of the global DRAM market in the second quarter, regaining its high position since 2024; while SK Hynix, which led last year, saw its market share drop from 39% to 26%; and Micron Technology's share rose to 25%, quietly falling behind SK Hynix. Analysis indicates that while traditional DRAM prices rose significantly quarter-over-quarter this year, HBM prices declined year-over-year—due to price reductions in existing HBM3E products and delays in HBM4 production, the average price of HBM dropped noticeably. Since SK Hynix has the highest proportion of HBM in its shipments and revenue, and signed long-term agreements based on fixed prices earlier, it suffered the greatest impact from the drop in HBM prices. However, as traditional DRAM prices continue to rise, the currently "relatively low" HBM prices are expected to see a significant increase in 2027—which is precisely the underlying logic for Nvidia to adjust product configurations before price hikes and shortages. It is noteworthy that Counterpoint believes that SMIC has the potential to surpass SK Hynix and take the position of the world's second-largest DRAM manufacturer. Meanwhile, China's DRAM leader ChangXin Technology has become the fastest-growing DRAM supplier with a year-over-year growth of 716%, shifting the market focus from "whether ChangXin can enter the top three" to "when ChangXin will enter the top three." The pattern of the "three giants vying for dominance" in the memory industry may change significantly within the next year. ## V. Implications for US Stock Investors From an investment perspective, this news conveys at least three signals: First, computing power demand remains strong. Nvidia is willing to maintain the Rubin Ultra on schedule, which precisely demonstrates the robust downstream demand; the prosperity of AI computing power investment has not cooled down but is instead challenging the supply chain to its limits. Second, pricing power in the memory segment is evident. Against the backdrop of insufficient HBM supply, memory manufacturers with leading process technology and capacity advantages are gaining unprecedented pricing power, and the memory sector is expected to become one of the segments with the highest "certainty premium" in the AI industry chain. Third, the center of computing power costs is shifting upwards. For cloud vendors and AI application companies, rising hardware costs will test their ability to realize profits; conversely, companies with capabilities in computing power leasing and energy supply may gain excess returns. In the short term, the Philadelphia Semiconductor Index still exhibits significant volatility after the sharp pullback in July. The rhythm of HBM supply, Nvidia product shipment guidance, and cloud vendor capital expenditure remain the three main lines determining the sector's direction. For investors focusing on US stock AI computing power opportunities, rather than chasing daily fluctuations, it is better to closely monitor the gap between HBM capacity release and GPU shipments—this is precisely the core variable for the next round of computing power market trends.