NVIDIA rolls out new AI superchips Blackwell Ultra and Vera Rubin

US-based chipmaker NVIDIAhas introduced its latest artificial intelligence (AI) chips, Blackwell Ultra and Vera Rubin, designed to build and run advanced AI models. 

CEO Jensen Huang announced the new lineup of chips during the company’s annual GTC conference in San Jose, California. 

Blackwell Ultra, which will be shipped later this year, offers improved AI model performance and efficiency.

Looking ahead, Nvidia also revealed a next-generation family of graphics chips called Vera Rubin, scheduled for release in 2026.  

Named after an astronomer who discovered evidence of dark matter, Vera Rubin will feature Nvidia’s first custom-built CPU 'Vera', alongside a new GPU design called 'Rubin'.

Pairing with Vera, Rubin GPU will deliver 50 petaflops for AI inference—more than double the 20 petaflops of Blackwell, and support up to 288 GB of high-speed memory. 

Huang observed that the computing space is at a $1-trillion inflection point as AI demand surges, fuelled by the demand for reasoning and agentic AI. He added that the increase in AI workloads is prompting significant investments in data centres around the world.

The Silicon Valley-based firm is aiming to roll out new GPUs, CPUs, and computing advancements on an annual basis to fuel its AI infrastructure, including the upcoming NVIDIA Vera Rubin architecture. 

Looking ahead, Huang revealed that, after Rubin chips, Nvidia would introduce Feynman chips in 2028. 

The dominance of GPUs in AI has been recently called into question following the launch of low-cost models by Chinese startup DeepSeek, which triggered a market shakeup that erased $592.7 billion from Nvidia’s valuation.

Huang noted that with the firm’s new open source software, Nvidia Dynamo, Blackwell chips can run DeepSeek R1 30 times faster.

“Almost the entire world got it wrong. The computation requirement, the scaling law of AI, is more resilient and in fact hyper-accelerated,” said Huang.