Rubin CPX
AI GPU for massive-context inference with GDDR7 memory
The Rubin CPX is a graphics processing unit from NVIDIA designed for long-context AI tasks. It uses GDDR7 memory and NVFP4 precision to handle million-token workloads in coding and video generation, launching in late 2026.
- Features 128GB of GDDR7 memory for large context windows
- Delivers up to 30 petaflops of NVFP4 compute performance
- Includes integrated video encoders and decoders
- Designed for million-token coding and generative video
- Expected to be available at the end of 2026
What it is
The Rubin CPX is a specialized graphics processing unit created by NVIDIA. It targets artificial intelligence systems that need to process vast amounts of data at once, such as million-token software code or hour-long video files. Unlike standard GPUs that rely heavily on high-bandwidth memory, this chip uses cost-efficient GDDR7 memory to manage large context windows during inference. It serves as a co-processor within larger server racks, working alongside central processors and other graphics units to boost speed and efficiency for generative applications.
How it came about
NVIDIA announced the chip at its AI Infra Summit, positioning it as a new category of processor for massive-context workloads. The company identified a split in AI tasks between compute-heavy pre-fill stages and memory-limited generation stages. By creating a dedicated unit for the context phase, NVIDIA aimed to solve bottlenecks in long-format content creation. The design reflects a shift away from homogeneous clusters toward heterogeneous racks that mix different memory architectures for better performance.
Inside the chip
This processor features a monolithic die layout packed with NVFP4 computing resources. It delivers up to 30 petaflops of compute power and includes 128 gigabytes of GDDR7 memory. The chip integrates four video encoder and decoder engines to handle visual data directly. It offers three times faster attention capabilities compared to previous GB300 systems, allowing models to reason across longer sequences without slowing down. The architecture is optimized for energy efficiency while maintaining high throughput for specific inference tasks.
How it was made
The Rubin CPX uses a monolithic die design rather than a multi-chip module approach. NVIDIA states this method allows for a cost-efficient build while packing powerful computing resources into a single unit. The chip relies on GDDR7 memory technology, which provides a balance of capacity and speed suitable for holding large key-value caches. Specific manufacturing details such as the foundry or exact transistor count are not disclosed in the available information.
Where it went
The chip supports AI coding assistants that need to understand entire software projects at once. It also accelerates generative video tools, enabling creators to produce cinematic content with longer durations and higher quality. Companies like Cursor, Runway, and Magic are exploring its use for developer productivity and autonomous software engineering agents. The processor runs multimodal models from the Nemotron family, helping enterprises deploy advanced reasoning agents in data centres.
Versions and family
The Rubin CPX is part of the broader Vera Rubin platform. It appears in the Vera Rubin NVL144 CPX configuration, which combines these GPUs with Vera central processing units and standard Rubin graphics cards in a single rack. A dedicated compute tray is also offered for customers who wish to upgrade existing systems. This setup allows for heterogeneous mixing of HBM-based GPUs and GDDR7-based CPX units to optimize different stages of AI processing.
Why it matters
NVIDIA claims the platform enables companies to generate significant revenue from token processing relative to their infrastructure investment. By splitting tasks between specialized chips, the system aims to lower the cost of inference for long-context applications. This approach could influence how future AI clusters are built, moving towards mixed-memory architectures that address specific bottlenecks in model reasoning and content generation rather than relying on a single type of accelerator.
Problems and reception
Industry partners have expressed interest in the chip for accelerating their specific workflows. Software firms note it will help unlock faster code generation, while creative studios see it as a leap forward for video production. The announcement highlights a strategic move by NVIDIA to address the growing demand for processing millions of tokens simultaneously, a task that traditional hardware struggles to handle efficiently.
Questions people ask about Rubin CPX
What makes Rubin CPX different from standard GPUs?
It uses GDDR7 memory instead of HBM and is optimized for the context phase of AI inference. This allows it to handle million-token workloads more cost-effectively than traditional high-bandwidth memory systems.
What kind of tasks is Rubin CPX best suited for?
It is designed for massive-context processing, such as analyzing entire codebases for software development or generating hour-long videos. It excels at tasks requiring long attention spans in AI models.
How much memory does the Rubin CPX have?
Each Rubin CPX GPU comes with 128 gigabytes of GDDR7 memory. This capacity helps store the large key-value caches needed for long-context inference without dropping speed.
When will the Rubin CPX be available?
NVIDIA expects the Rubin CPX to be available at the end of 2026. It will be part of the Vera Rubin NVL144 CPX platform and other configurations.
Good reading and sources
There is no English Wikipedia article on Rubin CPX yet, so the facts on this page come from the sources listed above: nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu, servethehome.com/nvidia-rubin-cpx-is-an-ai-gpu-for-next-gen-nvidia-ai-. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.