On September 9th, NVIDIA dropped a "bombshell" at the AI Infra Summit, unveiling its new GPU—the Rubin CPX. Designed specifically for long-context inference and video generation applications, it stands as an "efficiency multiplier" in the current AI inference space, poised to excel in scenarios demanding ultra-long context windows, such as programming and video generation.
NVIDIA founder and CEO Jensen Huang couldn’t hide his excitement, praising: “Just as RTX revolutionized graphics and physical AI, Rubin CPX—the first CUDA GPU custom-built for massive-context AI—can handle inference for millions of knowledge tokens simultaneously, opening a new chapter.”
Rubin itself is NVIDIA’s next-gen flagship computing chip, set to launch next year, and industry attention is high. The Rubin CPX based on it is expected to ship officially only by the end of 2026. NVIDIA’s next-gen flagship AI server, the NVIDIA Vera Rubin NVL144 CPX, takes performance further by integrating 36 Vera CPUs, 144 Rubin GPUs, and 144 Rubin CPX GPUs—truly a performance beast.
Looking at Rubin CPX’s specs: it packs 128GB GDDR7 memory and delivers 30 PFLOPS of AI computing power at NVFP4 precision. This robust performance is “tailor-made” for long-context processing (over 1 million tokens) and video generation tasks.
As for the Vera Rubin NVL144 CPX platform: within a single rack, it combines 144 Rubin CPX GPUs, 144 Rubin GPUs, and 36 Vera CPUs, offering up to 8 EFLOPS of AI performance (NVFP4 precision), 100TB of fast memory, and memory bandwidth soaring to 1.7 PB/s.
Performance comparisons highlight its edge: the platform’s AI performance is over 2x that of NVIDIA’s Vera Rubin NVL144 platform, and a massive 7.5x higher than the Blackwell Ultra-based GB300 NVL72 system—with attention mechanism processing speed 3x faster.
ICgoodFind：the launch of Rubin CPX is expected to reshape the long-context AI inference market landscape and inject strong momentum into the development of related applications—its future performance warrants continued attention.
