With exaflop-scale compute, massive memory, and faster attention mechanisms, the technology unlocks long-context reasoning and next-level productivity for developers and creators alike.
A new class of GPU has been introduced to tackle one of the toughest challenges in artificial intelligence: massive-context inference for coding and video. The Rubin CPX, unveiled at the AI Infra Summit, is built to process million-token workloads, enabling AI systems to reason across vast stretches of code or long-format video content with speed and efficiency.
NVIDIA describes Rubin CPX as the first CUDA GPU designed specifically for massive-context AI, positioning it as a complement to the Rubin GPU line. By integrating NVFP4 compute cores, a monolithic die design, and 128GB of GDDR7 memory, Rubin CPX achieves up to 30 petaflops of performance and three-times faster attention mechanisms than earlier systems. The hardware also includes integrated video encoding and decoding, critical for generative video and video search at the scale of hours of content.
The key specifications are:
- Built on the Vera Rubin NVL144 CPX platform
- Integrates Rubin CPX GPUs with Vera CPUs in a single MGX-based rack
- Delivers 8 exaflops of AI compute
- Provides 100TB of fast memory and 1.7PB/s bandwidth
- Achieves 7.5x performance boost over NVL72 systems
- Balances compute density and memory capacity for long-context AI workloads
Beyond raw performance, the system promises unprecedented monetization. According to projections, enterprises could generate up to $5 billion in token revenue for every $100 million invested, making Rubin CPX not only a technological leap but also an economic catalyst for AI infrastructure.
The GPU is already drawing attention from leading AI innovators. Cursor aims to boost developer productivity with lightning-fast code generation; Runway is preparing to expand cinematic AI workflows; and Magic sees Rubin CPX as key to powering autonomous software agents with 100-million-token context windows.
Software support will be equally critical. Rubin CPX will run the full NVIDIA AI stack, including the Dynamo platform for inference scaling and Nemotron multimodal models for enterprise-grade reasoning. Integration with CUDA-X libraries and NVIDIA AI Enterprise ensures compatibility across cloud and data center deployments.
