Nvidia announces "Rubin Ultra" and "Feynman" AI chips for 2027 and 2028

At the GTC 2025 conference, Nvidia CEO Jensen Huang announced the upcoming release of two new AI chips: Rubin Ultra and Feynman. The Rubin Ultra, set for the second half of 2027, will feature GPUs capable of delivering 100 petaflops of FP4 precision per chip, with a total rack performance of 15 exaflops for inference and 5 exaflops for training. This chip will utilize a new NVL576 rack configuration and include 1TB of HBM4e memory per GPU. The Rubin chip, named after astronomer Vera Rubin, is expected to launch in late 2026 and will significantly outperform its predecessor, Grace Blackwell, with 50 petaflops of FP4 inference performance per chip. Additionally, Huang introduced the Feynman architecture, expected in 2028, which will also utilize a Vera CPU. He emphasized the future of AI, envisioning data centers as "AI factories" and predicting that Nvidia chips will support billions of digital agents. Huang also mentioned the imminent launch of the Blackwell Ultra B300 chip in late 2025, which will enhance performance over the current Blackwell configuration.

- Nvidia announced the Rubin Ultra and Feynman AI chips for 2027 and 2028, respectively.

- Rubin Ultra will deliver 100 petaflops of FP4 precision per chip and 15 exaflops at the rack level.

- The Rubin chip is set to launch in late 2026, offering significant performance improvements over its predecessor.

- The Feynman architecture will utilize a Vera CPU and is expected to be released in 2028.

- Nvidia envisions a future where AI powers billions of digital agents and humanoid robots.