Nvidia (NVDA) CEO Jensen Huang announced the company’s next-generation Blackwell Ultra AI chip during its annual GTC event in San Jose, Calif., on Tuesday.
In addition to the Blackwell Ultra chip, Nvidia also announced its GB300 superchip, which combines two Blackwell Ultras with the company’s Grace central processing unit (CPU). The chips are designed to power AI systems for customers ranging from hyperscalers like Amazon (AMZN), Google (GOOG, GOOGL), Microsoft (MSFT), and Meta (META) to research labs around the world.
At close: March 18 at 4:00:00 PM EDT
According to Nvidia, the Blackwell Ultra offers 1.5 times the performance of Blackwell and represents a 50x increase in data center revenue opportunity versus its Hopper chip, thanks to its improved AI capabilities.
Nvidia says the Blackwell Ultra is designed for the “age of AI reasoning,” a type of AI processing that mimics how humans think and reach conclusions. It broke into the mainstream when DeepSeek debuted its R1 AI model. OpenAI’s o1 and Google’s Gemini 2.0 Flash Thinking are also reasoning models.
DeepSeek initially sent a shock through Wall Street when it said that it developed its AI models at a fraction of the cost that Silicon Valley heavyweights spend while using below-top-of-the-line chips. But Nvidia has fought back against that assertion, saying that reasoning models benefit from using powerful GPUs, which allow them to provide better responses to user queries faster.
Read more: How does Nvidia make money?
Like Blackwell, the Blackwell Ultra will slot into Nvidia’s massive NVL72 rack server that combines 72 GB300 superchips, which the company says provides improved efficiency and serviceability.
According to the company, the GB300 NVL72 can handle 1,000 tokens per second when using DeepSeek’s R1 AI model. That’s up from 100 tokens per second when using Nvidia’s Hopper chip. That means the GB300 NVL72 can answer users’ questions in about 10 seconds, versus the 1.5 minutes it took Hopper. In other words, Blackwell Ultra is a major step up from older Hopper systems.
On top of that, Nvidia says it will also offer the GB300 in its DGX SuperPod, the company’s AI supercomputer that combines a series of NLV72 servers into a single AI powerhouse. The SuperPods will include a staggering 288 Grace CPUs with 576 Blackwell Ultra GPUs and an incredible 300TB of memory.
Nvidia’s Blackwell chip is now in full production and, according to the company, has been its fastest ramp-up in history. In its most recent quarter, Nvidia said Blackwell contributed $11 billion to its $39.3 billion in total revenue.
