Nvidia Unveils Blackwell Ultra AI Chip and GB300 Superchip at GTC 2025, Targeting the ‘Age of AI Reasoning’
At its annual GTC conference in San Jose, California, Nvidia (NVDA) CEO Jensen Huang unveiled the company’s next-generation Blackwell Ultra AI chip, alongside the GB300 superchip, marking a significant leap in AI processing capabilities. The GB300 superchip integrates two Blackwell Ultra GPUs with Nvidia’s Grace CPU, designed to power advanced AI systems for hyperscalers like Amazon (AMZN), Google (GOOG, GOOGL), Microsoft (MSFT), and Meta (META), as well as global research labs.
Recent Post
Nvidia claims the Blackwell Ultra delivers 1.5 times the performance of its predecessor, Blackwell, and represents a 50x increase in data center revenue opportunities compared to the Hopper chip. This boost is attributed to its enhanced AI capabilities, positioning it as a cornerstone for the “age of AI reasoning”—a phase where AI systems mimic human thought processes to draw conclusions. This concept gained traction with the debut of DeepSeek’s R1 AI model, followed by reasoning models like OpenAI’s o1 and Google’s Gemini 2.0 Flash Thinking.
DeepSeek initially made waves by developing its AI models at a fraction of the cost of Silicon Valley giants, using less advanced chips. However, Nvidia has countered this narrative, emphasizing that reasoning models benefit significantly from powerful GPUs, enabling faster and more accurate responses to user queries.
The Blackwell Ultra will be integrated into Nvidia’s NVL72 rack server, which combines 72 GB300 superchips for improved efficiency and serviceability. According to Nvidia, the GB300 NVL72 can process 1,000 tokens per second using DeepSeek’s R1 AI model, a tenfold increase over the Hopper chip’s 100 tokens per second. This translates to answering user queries in approximately 10 seconds, compared to Hopper’s 1.5 minutes, showcasing a substantial performance upgrade.
Additionally, Nvidia plans to incorporate the GB300 into its DGX SuperPod, an AI supercomputer that clusters multiple NVL72 servers into a single powerhouse. Each SuperPod will feature 288 Grace CPUs, 576 Blackwell Ultra GPUs, and an impressive 300TB of memory, further solidifying Nvidia’s position as a leader in AI infrastructure.
The announcement underscores Nvidia’s commitment to advancing AI reasoning capabilities, catering to the growing demand for faster, more efficient AI systems across industries.
