Nvidia announces Blackwell Ultra and Vera Rubin AI chips

Nvidia CEO Jensen Huang arrives to attend the opening ceremony of Siliconware Precision Industries Co. (SPIL)’s Tan Ke Plant site in Taichung, Taiwan Jan. 16, 2025.

Ann Wang | Reuters

Nvidia announced new chips for building and deploying artificial intelligence models at its annual GTC conference on Tuesday.

CEO Jensen Huang revealed Blackwell Ultra, a family of chips shipping in the second half of this year, as well as Vera Rubin, the company’s next-generation graphics processing unit, or GPU, that is expected to ship in 2026.

Nvidia’s sales are up more than sixfold since its business was transformed by the release of OpenAI’s ChatGPT in late 2022. That’s because its “big GPUs” have most of the market for developing advanced AI, a process called training.

Software developers and investors are closely watching the company’s new chips to see if they offer enough additional performance and efficiency to convince the company’s biggest end customers — cloud companies including Microsoft, Google and Amazon — to continue spending billions of dollars to build data centers based around Nvidia chips.

“This last year is where almost the entire world got involved. The computational requirement, the scaling law of AI, is more resilient, and in fact, is hyper-accelerated,” Huang said.

Tuesday’s announcements are also a test of Nvidia’s new annual release cadence. The company is striving to announce new chip families on an every-year basis. Before the AI boom, Nvidia released new chip architectures every other year.

The GTC conference in San Jose, California, is also a show of strength for Nvidia.

The event, Nvidia’s second in-person conference since the pandemic, is expected to have 25,000 attendees and hundreds of companies discussing the ways they use the company’s hardware for AI. That includes Waymo, Microsoft and Ford, among others.

Nvidia will also showcase its other products and services at the event.

For example, Nvidia announced new laptops and desktops using its chips, including two AI-focused PCs that will be able to run large AI models such as Llama or DeepSeek. The company also announced updates to its networking parts for tying hundreds or thousands of GPUs together so they work as one, as well as a software package called Dynamo that helps users get the most out of their chips.

Vera Rubin

Jensen Huang, co-founder and chief executive officer of Nvidia Corp., speaks during the Nvidia GPU Technology Conference (GTC) in San Jose, California, US, on Tuesday, March 18, 2025.

David Paul Morris | Bloomberg | Getty Images

Nvidia said on Tuesday that it expects to start shipping systems on its next-generation GPU family in the second half of 2026.

The system has two main components: a CPU, called Vera, and a new GPU design, called Rubin. It’s named after astronomer Vera Rubin.

Vera is Nvidia’s first custom CPU design, the company said, and it’s based on a core design they’ve named Olympus.

Previously when it needed CPUs, Nvidia used an off-the-shelf design from Arm. Companies that have developed custom Arm core designs, such as Qualcomm and Apple, say that they can be more tailored and unlock better performance.

The custom Vera design will be twice as fast as the CPU used in last year’s Grace Blackwell chips, the company said.

When paired with Vera, Rubin can manage 50 petaflops while doing inference, more than double the 20 petaflops for the company’s current Blackwell chips. Rubin can also support as much as 288 gigabytes of fast memory, which is one of the core specs that AI developers watch.

Nvidia is also making a change to what it calls a GPU. Rubin is actually two GPUs, Nvidia said.

The Blackwell GPU, which is currently on the market, is actually two separate chips that were assembled together and made to work as one chip.

Starting with Rubin, Nvidia will say that when it combines two or more dies to make a single chip, it will refer to them as separate GPUs. In the second half of 2027, Nvidia plans to release a “Rubin Next” chip that combines four dies to make a single chip, doubling the speed of Rubin, and it will refer to that as four GPUs.

Nvidia said that will come in a rack called Vera Rubin NVL144. Previous versions of Nvidia’s rack were called NVL72.

Blackwell Ultra

Nvidia also announced new versions of its Blackwell family of chips that it calls Blackwell Ultra.

That chip will be able to produce more tokens per second, which means that the chip can generate more content in the same amount of time as its predecessor, Nvidia said in a briefing.

Nvidia says that means that cloud providers can use Blackwell Ultra to offer a premium AI service for time-sensitive applications, allowing them to make as much as 50 times the revenue from the new chips as the Hopper generation, which shipped in 2023.

Blackwell Ultra will come in a version with two paired to an Nvidia Arm CPU, called GB300, and a version with just the GPU, called B300. It will also come in versions with eight GPUs in a single server blade and a rack version with 72 Blackwell chips.

The top four cloud companies have deployed three times the number of Blackwell chips as Hopper chips, Nvidia said.

DeepSeek

Jensen Huang, co-founder and chief executive officer of Nvidia Corp., speaks during the Nvidia GPU Technology Conference (GTC) in San Jose, California, US, on Tuesday, March 18, 2025.

David Paul Morris | Bloomberg | Getty Images

China’s DeepSeek R1 model may have scared Nvidia investors when it was released in January, but Nvidia has embraced the software. The chipmaker will use the model to benchmark several of its new products.

Many AI observers said that DeepSeek’s model, which reportedly required fewer chips than models made in the U.S., threatened Nvidia’s business.

But Huang said earlier this year that DeepSeek was actually a good sign for Nvidia. That’s because DeepSeek uses a process called “reasoning,” which requires more computing power to provide users better answers.

The new Blackwell Ultra and Vera Rubin chips are better for reasoning models, Nvidia said.

It’s developed its chips to more efficiently do inference, so when new reasoning models require more computing power at the time of deployment, Nvidia’s chips will be able to handle it.

“In the last 2-3 years, a major breakthrough happened, a fundamental advance in artificial intelligence happened, we call it agentic AI,” Huang said. “It can reason about how to answer or how to solve a problem.”