With the emergence of the low-cost generative artificial intelligence (AI) model DeepSeek, concerns that the demand for high-performance graphics processing units (GPUs) would slow down have been dispelled as Jensen Huang, the chief executive officer (CEO) of NVIDIA, presented a roadmap for next-generation AI semiconductors that surpasses the performance of the current state-of-the-art GPU 'Blackwell'.
On the 18th (local time), Huang CEO expressed the intention to pursue extreme performance with the next-generation GPU 'Rubin Ultra', which will be launched from 2027, consisting of four dies instead of the existing two dies (Die, small square pieces), at NVIDIA's annual conference 'GTC 2025' held in San Jose, California. Since this product requires high-bandwidth memory (HBM) with higher capacity and performance, it is likely to be a boon for SK hynix, Samsung Electronics, and others.
NVIDIA unveiled the next-generation AI semiconductors, Rubin and Blackwell Ultra, which were highlighted as the key points that day. Huang CEO personally introduced the performance and features of the next-generation products. Huang CEO noted, "We are at an important turning point in the era of AI semiconductors" and presented a blueprint to enhance the performance of AI semiconductors after Blackwell.
The industry evaluates that NVIDIA's announcement this time is not significantly different from the existing methods of enhancing the hardware performance of AI semiconductors by increasing the new manufacturing process and the integration density of chips. The next-generation AI semiconductor Vera Rubin, which follows Blackwell, is a next-generation AI accelerator that combines a self-designed central processing unit (CPU) Vera and the next-generation GPU Rubin, named after the astronomer Vera Rubin who discovered dark matter. Huang CEO explained, "In terms of standard performance for AI factories (data centers), Blackwell is expected to be 68 times and Rubin 900 times better than the previous chip H100 ‘Hopper,’ while the expenses are only 13% for Blackwell compared to Hopper, and 3% for Rubin."
NVIDIA plans to start mass production of GPUs based on the Rubin architecture in the second half of next year. Rubin will feature a new CPU called 'Vera' instead of the existing CPU 'Grace'. In 2027, an upgraded version of Rubin called 'Rubin Ultra' will be released, and in 2028, a new AI chip named 'Feynman' is expected to be launched, according to Huang CEO. However, specific specifications for Feynman were not disclosed that day.
Huang CEO added, "The upgraded version of Blackwell, 'Blackwell Ultra,' is expected to be released in the second half of this year." Blackwell Ultra is a product that increases the existing 192 gigabyte (GB) 5th generation HBM, HBM3E, to 288GB, a 50% increase. 'Blackwell Ultra' will be offered in combination with NVIDIA's Arm-based CPU 'GB300' and GPU version 'B300.'
The roadmap presented by Huang CEO suggests that NVIDIA will stick to the existing method of maximizing hardware performance despite industry forecasts that the demand for high-priced GPUs will decrease following the emergence of DeepSeek. However, the industry's response has been somewhat lukewarm. Unlike last year, when Blackwell was introduced and raised market expectations, NVIDIA's stock closed down 3.43% that day, more than double the 1.7% decline in the Nasdaq index on the same day.
Recently, as the Chinese AI model DeepSeek has successfully implemented high-performance AI with fewer AI chips and lower performance AI, questions have arisen about whether NVIDIA's high-performance, high-priced semiconductors are still as necessary as before. Huang CEO remarked about DeepSeek, "DeepSeek is rather a boon for NVIDIA semiconductors," adding, "DeepSeek utilized semiconductors for inference, and to produce better results in this process, more semiconductors are necessary."
Recently, doubts have been raised about whether NVIDIA can respond to the customized semiconductors (ASICs) for AI pursued by major American tech companies such as Meta and Google with the same ultra-high-performance GPU roadmap as before. An IT industry official stated, "The big tech companies, which manage massive AI investment expenses annually, are actively considering Broadcom's ASICs as a cost-effective option," and added, "Given the unavoidable impact of U.S. semiconductor sanctions on sales in China, it is uncertain whether NVIDIA's strategy to maximize hardware performance will have the same power as in the past."