**NVIDIA’s AI Chip Roadmap Announced: Ultra, Rubin, and Feynman Set for Release from Second Half of This Year Until 2028**
Jensen Huang: “100 Times More Computing Power Needed with Emergence of Inference Models”
Expanding Dominance to Robots and Personal Supercomputers; Photon Communication Networking Chip Also Unveiled
GM Stock Drops; Jensen Huang’s Influence as the ‘Midas Touch’ of the Stock Market Weakens
(San Francisco = Yonhap News) Correspondent Kim Tae-jong = NVIDIA, a leading player in AI chips, announced its roadmap for future AI chips on the 18th (local time).
Jensen Huang, CEO, revealed the AI chip roadmap until 2028 at the annual developer conference GTC 2025 held at SAP Center in San Jose, California.
NVIDIA, which has been releasing the Blackwell since last year, announced in June last year that it would launch a new AI chip called ‘Rubin’ next year and detailed it further today.
CEO Huang stated that from the second half of this year until 2027, upgraded versions of Blackwell, Rubin, and an upgraded version of Rubin would be introduced, with a new AI chip being launched in 2028.
He noted that “in terms of AI factory (data center) performance, Blackwell is 68 times, and Rubin is 900 times more powerful than the previous chip, Hopper,” adding that “cost for the same function is only 13% with Blackwell and 3% with Rubin compared to Hopper.”
He also dismissed concerns about production setbacks due to design flaws, asserting that the production of the advanced AI chip ‘Blackwell,’ released late last year, is “fully operational.”
He explained that the four major cloud companies—Amazon, Microsoft, Google, and Oracle—purchased 1.3 million units of the previous generation Hopper chip last year and are projected to purchase 3.6 million units of Blackwell this year, highlighting its popularity.
CEO Huang stated that an upgraded version of Blackwell, ‘Blackwell Ultra,’ is set to launch in the second half of this year. Blackwell Ultra has increased the 5th generation HBM, HBM3E, from 192GB to 288GB, a 50% increase.
‘Blackwell Ultra’ comes with NVIDIA’s Arm-based CPU ‘GB300’ and a GPU version known as ‘B300.’
In the second half of next year, a new architecture AI chip called ‘Rubin’ will be introduced. Rubin will incorporate a new CPU called ‘Vera,’ replacing the previous CPU, Grace, found in existing chips.
In 2027, Rubin’s upgraded version, Rubin Ultra, will be launched, with a new AI chip named ‘Feynman’ expected to debut in 2028. Rubin is 3.3 times more efficient than the current Blackwell, and Rubin Ultra offers a significant 14 times improvement. However, specific specifications for Feynman were not disclosed.
CEO Huang announced plans to expand AI’s dominance beyond data centers to robots and personal supercomputers.
He unveiled ‘Isaac Groot Blueprint,’ a VR generation software for developing humanoid robots.
In collaboration with Walt Disney and Google DeepMind, he introduced ‘Newton,’ a software ‘physics engine’ as part of a robot development project. Disney’s robot, ‘Blue,’ also appeared on stage.
CEO Huang stated, “Physical AI (used in robots, etc.) will open up a new trillion-dollar market.”
NVIDIA is working with General Motors (GM) to use AI in next-generation cars, factories, and robots. It also collaborates with Hyundai Motor Group in simulating Boston Dynamics’ Atlas robot used in production lines utilizing its robot development platform.
He showcased the personal supercomputer ‘Spark’ created using its own CPU Grace and Blackwell chip, adding that PC and server manufacturers like Dell Technologies and HP plan to release personal supercomputers using it.
Furthermore, in collaboration with TSMC, the world’s first silicon photonics networking chip enabling computer-to-computer communication via photons is to be released within this year. Silicon photonics is a technology that facilitates faster transmission and lower power consumption by enabling photon-based computer communication instead of electronic-based.
Despite the shock from the introduction of Chinese AI startup DeepSec’s R1 model, CEO Huang predicted a need for even greater computing power.
He explained that “inference model introduction requires 100 times more computing power than before,” and “AI inference models and AI agents will significantly increase demand for NVIDIA chips.”
He emphasized the need for “scaling up existing servers before horizontal expansion through additional servers is maximized.”
NVIDIA’s stock fell 3.4% on the New York Stock Exchange compared to the previous day. GM’s stock, which announced cooperation with NVIDIA, also fell 0.7%.
Bloomberg News commented, “Jensen Huang’s ‘Midas touch’ influence in the stock market is waning,” stating that while mentioning names of customers and partner companies used to cause a surge in stock prices, his keynote speech did not elicit a notable reaction this time.