Nvidia announces Blackwell Ultra and Vera Rubin AI chips

nvidia At the annual GTC conference on Tuesday, the new chips were announced for building and deploying artificial intelligence models.

CEO Jensen Huang revealed the family of chips, Blackwell Ultra, later this year, and Bera Rubin, a GPU scheduled to ship in 2026 (the company’s next-generation graphics) (GPU).

Nvidia sales have risen more than six times since the business changed with Openai’s ChatGPT release in late 2022. This is because “Big GPUs” have most of the market for developing Advanced AI, a process known as training.

Software developers and investors are looking closely at the company’s new chips to see if it offers enough additional performance and efficiency to convince the company’s biggest customers. Microsoft, Google and Amazon – To continue to spend billions of dollars building data centers based on nvidia chips.

“This last year, it’s where almost the world was involved. The computational requirements, which are AI scaling methods, are more resilient and in fact hyper-accelerated,” says Huang.

Tuesday’s announcement also tests Nvidia’s new annual release cadence. The company strives to announce new chip families every year. Before the AI ​​boom, Nvidia released a new chip architecture every other year.

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

The event, Nvidia’s second in-person meeting since the pandemic, will feature 25,000 participants and hundreds of companies discuss how to use company hardware for AI. This includes Waymo, Microsoft, and Fordespecially. General Motors We have also announced that Nvidia services will be used for next-generation vehicles.

The Chip Architecture, named after Rubin, is named after physicist Richard Feynman, and said Tuesday it continues the tradition of naming the Chip Family after scientists. According to the slides displayed by Huang, Nvidia’s Feynman chip is expected to be available in 2028.

Nvidia will showcase other products and services at the event.

For example, Nvidia has announced a new laptop and desktop using the chip, including two AI-centric PCs called DGX Spark, which can run large AI models such as Llama and Deepseek. The company also announced updates to networking parts to tie hundreds or thousands of GPUs together, so they have announced a software package called Dynamo that will work as one and help users make the most of their chips.

Nvidia plans to launch a delivery system with its next-generation GPU family in the second half of 2026.

This system has two main components: a CPU called Vera and a new GPU design called Rubin. It is named after astronomer Vera Rubin.

Vera is Nvidia’s first custom CPU design, which the company says, and is based on the core design it calls Olympus.

Previously, when a CPU was needed, Nvidia used a ready-made design from arm. Companies that have developed custom arm core designs such as Qualcomm and Apple say they can become more tailored and unleash better performance.

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

When combined with Vera, Rubin manages 50 PetaFlops while making inferences. That’s more than twice the company’s current Blackwell chip 20 petaflops. Rubin can also support 288GB of high speed memory, one of the core specifications that AI developers view.

Nvidia is also making changes to what is called a GPU. Rubin is actually two GPUs, Nvidia said.

The Blackwell GPUs currently on the market are actually two independent chips that are assembled together and act as one chip.

Starting with Rubin, Nvidia will say that when you combine two or more dies to create a single chip, they will call them individual GPUs. In late 2027, Nvidia plans to release a “Rubin Next” chip that combines four dies to create a single chip, double Rubin’s speed, calling it four GPUs.

Nvidia said it would enter a rack called the Vera Rubin NVL144. An earlier version of Nvidia’s rack was called NVL72.

Nvidia has also announced that it will call the new version of the Blackwell Family of Chips Blackwell Ultra.

That chip will be able to generate more tokens per second. This means that the chip can generate more content in the same time as its predecessor, the company said in its briefing.

Nvidia means that cloud providers can use Blackwell Ultra to provide premium AI services to time-sensitive applications, meaning they can earn 50 times the revenue from new chips, 50 times the Hopper Generation shipped in 2023.

The Blackwell Ultra features two paired versions of an Nvidia ARM CPU called the GB300 and a GPU-only version called the B300. There is also a version with eight GPUs on a single server blade and a rack version with 72 blackwell tips.

The top four cloud companies roll out three times the number of blackwell chips, like hopper chips, Nvidia said.

The Chinese Deepseek R1 model may have scared Nvidia investors when it was released in January, but Nvidia has adopted software. Chipmakers use models to benchmark several new products.

Many AI observers said Deepseek’s model, which reportedly has fewer chips than models made in the US, reportedly threatened Nvidia’s business.

But Huang said earlier this year that Deepseek is actually a good sign for Nvidia. This is because DeepSeek uses a process called “Reasoning.”

According to Nvidia, the new Blackwell Ultra chip is suitable for inference models.

Because we developed chips to make inferences more efficient, Nvidia’s chips can handle them if new inference models require more computing power when deployed.

“In the last two or three years, there have been major breakthroughs and fundamental advances in artificial intelligence. We call it Agent AI,” Huang said. “It can infer how to answer or how to solve the problem.”

Watch: Nvidia launches GTC meeting: Committee discusses how to trade it