Nvidia Corp. announced new iterations of its artificial intelligence chips at its annual GTC conference on Tuesday, introducing the Blackwell Ultra GPU and previewing its forthcoming Rubin family of processors. The moves come as the semiconductor giant aims to cement its dominance in AI infrastructure while grappling with rising investor scrutiny and global trade headwinds.
CEO Jensen Huang unveiled the Blackwell Ultra GPU, scheduled to ship in the second half of 2024, touting it as a significant leap in performance. Huang also previewed Rubin, Nvidia's next-generation GPU family set for release in 2026, alongside Vera, Nvidia's first custom CPU design. The announcements underscore Nvidia's commitment to an annual release cadence and its focus on supplying powerful chips to hyperscalers including Microsoft, Google, Amazon, and Meta.
"This last year is where almost the entire world got involved," Huang said during the conference, referring to the exponential growth in AI computing demand. "The computational requirement, the scaling law of AI, is more resilient, and in fact, is hyper-accelerated."
The Blackwell Ultra chip, paired with Nvidia's Grace CPU as part of the GB300 superchip, is designed for what Huang called the "age of AI reasoning." The chip delivers 1.5 times the performance of its predecessor and offers cloud providers the opportunity to generate up to 50 times more revenue per server rack compared to Nvidia's earlier Hopper generation chips, according to the company.
The GB300 NVL72 rack server, equipped with 72 GB300 superchips, is capable of processing 1,000 tokens per second when running reasoning models such as China's DeepSeek R1. This is a tenfold increase over the capabilities of the Hopper chip, slashing response times from 90 seconds to about 10 seconds per query.
Nvidia's push to support reasoning models comes as competitors experiment with developing more efficient AI systems using fewer chips. DeepSeek's R1 model, launched in January, initially rattled Nvidia investors by demonstrating high performance with lower chip usage. However, Huang argued that reasoning models inherently require more computational power during deployment, playing to Nvidia's strengths. "It can reason about how to answer or how to solve a problem," Huang said.
Nvidia's GTC event, held in San Jose, California, featured more than 25,000 attendees and presentations from partners like Waymo, Microsoft, Ford, and General Motors. The event served as a showcase of Nvidia's expansive role in the AI ecosystem, from cloud data centers to automotive applications.
Looking further ahead, Nvidia introduced Rubin, its next-generation GPU architecture, and Vera, its custom CPU design based on the Olympus core. Rubin is projected to manage 50 petaflops of inference performance and support up to 288 gigabytes of high-speed memory. Nvidia will further expand the Rubin family with Rubin Next in 2027, combining four GPU dies into one package, doubling performance. A subsequent generation, named Feynman after physicist Richard Feynman, is slated for release in 2028.
Nvidia's stock has experienced volatility in recent months, falling 11% year-to-date despite a 36% increase over the past year. The drop reflects investor concerns about overinvestment by cloud providers and the threat of new U.S. tariffs. President Donald Trump has floated a potential 25% tariff on overseas semiconductors, adding further uncertainty to the chipmaker's outlook.
Nvidia's revenue soared to $39.3 billion in its most recent quarter, with $11 billion attributed to Blackwell chip sales alone. But questions remain about the sustainability of this growth amid economic and geopolitical pressures.
