Nvidia outlined on Tuesday at AI infrastructure Summit a new revenue-generating GPU called Rubin CPX, describing it as a GPU for “massive context” in AI acceleration that will be available at the end of 2026.
A slide shown to reporters described Vera Rubin NVL 144 CPX (comprised of Rubin and Vera and other components in a rack server that is also available in late 2026) as having the ability to raise $5 billion in revenue for every $100 million invested.
Such a massive system is designed for high value use cases with 1 million or more token length. It would be useful for advanced coding of 100,000 lines of code and video processing and generation of one hour or more of HD video.
Shar Narasimhan, director or product for data centers, said inference work in AI is best accomplished when disaggregating the serving of large context jobs with 1 million or more tokens from small and medium context.
Rubin CPX will unlock AI approaches from Cursor, Fireworks AI, Magic, Runway and together.ai, all which have endorsed Ruboin CPX. he said.
In addition, Nvidia announced its AI Factory concept at the giga-scale that will include location for power generation an storage. Nvidia is working with a range of partners on the concept include Siemens and Siemens Energy, Vertiv, Cadence and Schneider.
Also, Nvidia touted its performance with Nvidia Blackwell Ultra, which set records for a new reasoning inference benchmark on MLPerf Inference v.5.1.