NVIDIA has unveiled its latest breakthrough in AI hardware, the Rubin CPX GPU.
Built to handle highly demanding tasks like generating videos and assisting in software development, this new chip shows how NVIDIA is making strides to stay at the forefront of AI.
Announced yesterday, the Rubin CPX GPU is designed for huge workloads, making it one of the most powerful AI processors NVIDIA has created to date.
The chip is built on NVIDIA’s new Rubin architecture, tailored for tasks that require ultra-long context windows. It can handle up to one million tokens, which is roughly equivalent to an hour of video.
This makes it especially valuable for advanced generative tasks that need to maintain coherence across large datasets, from long-form video generation to complex coding sessions.
Key Features
- 30 PFLOPs of NVFP4 compute power for each GPU
- 128 GB of GDDR7 memory designed for efficient and cost-effective scaling
- Built-in video encoding and decoding support
- Optimized for step-by-step reasoning, with Rubin CPX managing context and standard Rubin GPUs focusing on generation and decoding
Vera Rubin NVL144 CPX
The Rubin CPX GPU isn’t built to work on its own. It’s part of NVIDIA’s Vera Rubin NVL144 CPX platform, a rack-level system that brings together:
- 144 Rubin CPX GPUs
- 144 Rubin GPUs
- 36 Vera CPUs
According to NVIDIA, this setup delivers 7.5X the performance of previous-generation Blackwell systems, setting a new benchmark for AI infrastructure.
NVIDIA believes the new chip will also be a smart investment. The company estimates that a $100 million setup could return as much as $5 billion in revenue from AI tasks like training and running large models.
The chips and systems are expected to hit the market by late 2026, once fabrication at TSMC is complete. NVIDIA will offer them both as standalone GPUs and integrated in full rack-scale deployments.
