At the annual GTC conference, NVIDIA (NVDA, Financial) introduced its latest chip family, Blackwell Ultra, aimed at enhancing the construction and deployment of AI models. CEO Jensen Huang announced that these chips would start shipping later this year. The company also revealed a next-generation GPU architecture named Vera Rubin, expected to launch in 2026.
Since the release of OpenAI's ChatGPT at the end of 2022, NVIDIA's business has seen significant growth, with sales increasing over sixfold. This surge is attributed to NVIDIA's dominance in the advanced AI model development market with its large-scale GPU chips. Developers and investors are closely monitoring NVIDIA's new chips to assess their performance and efficiency, which are crucial for major cloud companies like Microsoft, Google, and Amazon in building data centers centered around NVIDIA's technology.
Jensen Huang emphasized the global involvement in AI development, noting the robust and accelerating demand for AI computing. NVIDIA is now aiming to update its chip architecture annually, a shift from its previous biennial schedule. Following the Rubin architecture, the next generation will be named after physicist Richard Feynman, with a projected release in 2028.
During the GTC conference, NVIDIA showcased other innovative products, including new laptops and desktops equipped with its chips, such as the AI-focused DGX Spark and DGX Station. The company also introduced new network components and a software package named Dynamo to optimize chip performance.
The Vera Rubin chip system, comprising the Vera CPU and Rubin GPU, is a significant advancement. The Vera CPU features NVIDIA's first self-designed architecture, Olympus, offering double the performance of previous generations. Combined with the Rubin GPU, it achieves 50 petaflops of inference performance, surpassing the current Blackwell chip's 20 petaflops, and supports up to 288GB of high-speed memory.
NVIDIA plans to redefine its GPU approach with the Rubin architecture, adopting a dual-chip design that will be termed as multiple independent GPUs. A future version, Rubin Next, is expected in late 2027, integrating four chips to double performance.
The Blackwell Ultra chip, a new version of the Blackwell series, significantly enhances content generation efficiency, allowing cloud providers to offer high-end AI services for time-sensitive applications. It is said to have up to 50 times the revenue potential of the 2023 Hopper chip. Currently, the use of Blackwell chips by major cloud companies is over three times that of Hopper chips.
Despite investor concerns about the DeepSeek R1 model posing a threat, NVIDIA argues that it highlights the demand for its products. The DeepSeek model relies on a reasoning process requiring greater computational power, which the new Blackwell Ultra chip supports more efficiently.
Jensen Huang concluded by stating that the past few years have seen significant breakthroughs in AI, ushering in the "agentic AI" era, characterized by reasoning and problem-solving capabilities, which will drive future demand for computational performance.
