At the 2025 AI Infrastructure Summit, NVIDIA took the wraps off the Rubin CPX GPU, a product specifically crafted for AI inference tasks. Built upon the Rubin architecture, this GPU boasts a monolithic design and is equipped with a whopping 128GB of GDDR7 memory. In terms of performance, it can deliver up to 30 petaflops of compute power when operating in NVFP4 precision. This makes it highly suitable for long - context inference tasks that involve processing millions of tokens.

The Rubin CPX has a great deal of flexibility in system integration. It can be paired with the Vera CPU and Rubin GPU to create the Vera Rubin NVL144 CPX platform. This platform is a powerhouse, offering an impressive 8 exaflops of AI compute power. Additionally, it comes with 100TB of high - speed memory and provides a memory bandwidth of 1.7PB/s per rack. When compared to the GB300 NVL72 system, its performance is 7.5 times greater.

Although the Rubin CPX GPU is primarily aimed at enterprise - grade AI workloads, there's an interesting aspect. The fact that it retains raster units indicates that it has the potential to act as the base for future RTX 6090 graphics cards. Market analysts expect it to be released to the market by the end of 2026.