Nvidia’s 2025 GTC: new AI chips, PCs, robots and partnerships

At its annual GTC conference, Nvidia unveiled new processors for training and running AI models — Blackwell Ultra and Vera Rubin. It also introduced an engine for simulating robot motion and a range of other products.

The Blackwell Ultra family will ship in the second half of 2025. Vera Rubin is due next year.

The next-generation architecture (after Rubin) will be named after the physicist Richard Feynman. It is slated to go on sale in 2028.

Nvidia also introduced several other solutions, including:

• new desktop computers based on its chips;
• the Newton engine for simulating robot motion and the Groot N1 model;
• an update to networking components to combine hundreds or thousands of graphics processing units (GPUs) into a single system;
• the Dynamo software suite to help users get the most from its chips.

Vera Rubin

The new system, named after astronomer Vera Rubin, consists of two components: a central processing unit (CPU) called Vera and a GPU called Rubin.

Nvidia stressed that it designed the CPU itself for the first time, based on its Olympus core. Previously the company used Arm’s off‑the‑shelf designs. Thanks to the new approach, Vera is twice as fast as last year’s Grace Blackwell chips.

Paired with Vera, the Rubin graphics processor can perform 50 petaflops (50 quadrillion operations per second). The current Blackwell delivers 20 petaflops. Rubin can also support up to 288 gigabytes of high‑speed memory — a key spec for AI developers.

Technically, Rubin is two GPUs in one package. It will be followed in 2027 by Rubin Ultra, which offers a set of four GPUs delivering up to 100 petaflops.

Blackwell Ultra

The new versions of the flagship chips in the Blackwell family can produce more tokens per second, enabling heavy content generation. This should let providers offer premium AI services for time‑sensitive applications.

Personal PCs

Nvidia announced a new line of “personal AI supercomputers” based on the Grace Blackwell platform — DGX Spark (formerly Project Digits) and DGX Station. They can run large language models such as Llama or DeepSeek.

“This is the computer of the AI era. This is what PCs should look like. […] Now we have a whole lineup for enterprises, from small and tiny to workstations,” said Nvidia’s CEO Jensen Huang.

DGX Spark delivers up to one quadrillion computing operations per second thanks to a 10‑gigabyte Grace Blackwell Superchip. DGX Station is equipped with a GB300 Grace Blackwell paired with 784 gigabytes of memory.

DGX Spark is available to order; DGX Station will ship later in 2025 through partners such as Asus, Boxx, Dell, HP and Lenovo.

“AI agents will be everywhere. How they work, how enterprises and we work — everything will change radically. So we need a new line of computers. And this is it,” Huang stressed.

Newton robot engine

Nvidia is working with Disney Research and Google DeepMind on Newton, a physics engine for simulating robot movements in real‑world conditions.

Disney will be the first to apply the tool for its next‑generation entertainment bots, such as the Star Wars‑inspired BDX. One of them walked on stage with Huang during the presentation.

Nvidia plans to release an open‑source early version of Newton later in 2025.

The physics engine will help robots become more “expressive” and “learn to handle complex tasks with greater precision,” Nvidia said. Newton is designed to help developers model how bots interact with the natural world.

Model for humanoids Groot N1

Nvidia separately announced the Groot N1 model, which it calls “the foundation AI for humanoid robotics.” It implements a “two‑system architecture” for “fast and slow thinking,” inspired by human cognitive processes.

The slow‑thinking system allows a robot to perceive and make sense of its environment and instructions, then plan the right actions. Fast thinking converts that plan into actions.

“The era of general-purpose robotics has arrived,” said Huang.

Groot N1 builds on last year’s Project Groot.

Partnership with General Motors

General Motors (GM) struck an expanded partnership with Nvidia to bring artificial intelligence into the physical world. The collaboration will span all aspects of the carmaker’s business, including factories, robots and self‑driving cars.

The time for autonomous machines has come, Huang stressed.

“We look forward to building AI together with GM in all three domains. AI for manufacturing, so they can revolutionize the way they manufacture; AI for enterprise, so they can revolutionize the way they work on car design and car simulation, and AI for machines,” he said from the stage.

Nvidia will provide AI infrastructure and help GM build its own artificial intelligence. The automaker will also use Nvidia Drive AGX in its onboard hardware for future driver‑assistance systems and improved driving safety.

DeepSeek’s impact

On January 20, Chinese AI startup DeepSeek released DeepSeek‑R1, an open‑source reasoning model. The firm offered performance comparable to top models at far lower cost, which raised doubts about the need for multibillion‑dollar investments in AI and led to a plunge in Nvidia’s shares.

Huang said DeepSeek was in fact a positive signal for his company. The Chinese firm’s model uses a reasoning process that demands more compute to deliver better answers.

According to Nvidia’s CEO, Blackwell Ultra is well suited to “thinking” models. The company designed the new chips to run inference more efficiently.

Nvidia’s fourth‑quarter financial report beat expectations, and its guidance for the first quarter of 2025 was upbeat.