Nvidia kicked off GTC 2025 on Tuesday with a series of announcements. The main highlights include a new physics engine called Newton, a fresh AI foundation model for humanoid robots named Groot N1, and two “personal AI supercomputers” under the DGX lineup.
Jensen Huang, Nvidia’s founder and CEO, led the keynote in San Jose, California, spotlighting collaborations with Disney Research and Google DeepMind that are intended to reshape how robots move and learn.
Disney, whose Star Wars-inspired BDX droids have generated curiosity for years, will be among the first to use Newton. During Tuesday’s keynote, one of those droids waddled onstage alongside Huang.
According to Disney, these robots may soon appear in theme parks worldwide. Nvidia notes that Newton is highly customizable and will integrate with Google DeepMind’s existing robotics tools. An early open-source version of Newton will be released later in 2025.
Nvidia framed Newton as an engine that developers can use to simulate how robots interact with items like food, sand, cloth, and other deformable objects. It was also stressed that Newton incorporates features from DeepMind’s MuJoCo, a physics engine that supports multi-joint robot movements.
By combining these components, Nvidia expects to give robotics developers a more efficient way to program tasks that might be difficult to replicate in real-world test environments. The company also highlighted how this partnership with Google DeepMind reinforces a larger focus on AI-driven robotics.
In addition to Newton, Nvidia introduced Groot N1. Billed as an AI foundation model for humanoid robots, Groot N1 is designed to support robots in “perceiving and reasoning about their environments.” The company says this foundation model should help robots move seamlessly through spaces that might involve unpredictable shifts in terrain, lighting, or human activity. Although additional details on Groot N1’s technical specifications remain sparse, Nvidia suggests that it will play a role in how robots across multiple industries operate in the near future.
Huang reveals DGX personal AI supercomputers
At the same keynote, Huang discussed new AI chips and personal supercomputers, underscoring how much the company has grown since OpenAI’s ChatGPT launched in late 2022. Over that span, Nvidia has seen a sharp rise in the demand for its GPUs, with major cloud providers including Microsoft, Google, and Amazon being among its largest customers.
One of the most talked-about products revealed at GTC 2025 is the DGX Spark, part of Nvidia’s new “personal AI supercomputers” family. Powered by the Grace Blackwell chip platform, DGX Spark offers up to 1,000 trillion operations per second of AI computing. It is equipped with the GB10 Grace Blackwell Superchip, making it a robust machine for AI tasks executed at the edge.
The second device, DGX Station, features a GB300 Grace Blackwell Ultra Desktop Superchip and 784GB of memory. While DGX Spark is already available, DGX Station is slated for release later this year. Companies such as Asus, Boxx, Dell, HP, and Lenovo will serve as manufacturing partners.
Huang characterized these products as groundbreaking, stating, “This is the computer of the age of AI. This is what computers should look like, and this is what computers will run in the future.” Analysts attending the conference noted that these so-called personal supercomputers could help both large enterprises and smaller outfits move from AI experimentation to production more quickly.
Nvidia announces Blackwell Ultra and Vera Rubin
In parallel, Nvidia announced two new chip families aimed at building and deploying advanced AI models: Blackwell Ultra, scheduled for the latter half of this year, and Vera Rubin, expected to ship in 2026. The announcements are significant for cloud companies that rely on Nvidia hardware to run massive AI training workloads. Blackwell Ultra is meant to generate more tokens per second, allowing for faster content production during AI inference or other language-based processes.
With this capability, Nvidia suggests that cloud providers can introduce higher-tier AI services using the new chips, potentially bringing up to 50 times more revenue compared to what was possible with the company’s Hopper generation chips introduced in 2023.
The Blackwell Ultra lineup will include versions that pair with an Arm CPU, labeled GB300, and versions that focus solely on the GPU (B300). Additional configurations will have eight GPUs in one server blade or 72 GPUs in a single rack. Nvidia says that the top four cloud providers have already deployed three times the number of Blackwell chips compared to the previous generation.
Named after astronomer Vera Rubin, this system is expected to start shipping in the second half of 2026. It centers on two main parts: a CPU, called Vera, and a GPU design, called Rubin. Vera is Nvidia’s first custom CPU design, built on a new core architecture labeled Olympus. The company claims this custom CPU will be twice as fast as the Grace Blackwell CPU released in earlier iterations.
Rubin, the GPU component, will be capable of performing 50 petaflops of inference, surpassing the 20 petaflops of Nvidia’s current Blackwell chips. In addition, Rubin can support up to 288GB of fast memory, which is an important metric for AI developers dealing with large-scale models. Another shift in naming convention emerges here: while Blackwell GPUs feature two separate dies assembled as a single chip, Rubin will formally acknowledge each die as an individual GPU.
The company says it plans to release a “Rubin Next” version in the second half of 2027 that will include four dies in one chip, thereby offering double the speed of the initial Rubin version. That product will be housed in a rack setup called Vera Rubin NVL144, building on the NVL72 standard of previous racks.
Adding to the roadmap, Huang presented a slide that revealed the name of the chip family after Rubin: Feynman, named after the physicist Richard Feynman. Nvidia expects to release Feynman chips in 2028, rounding out a multi-year plan that includes incremental and major architectural changes in AI hardware design.
Cryptopolitan Academy: Want to grow your money in 2025? Learn how to do it with DeFi in our upcoming webclass. Save Your Spot
