NVIDIA Rubin CPX's Second Half Launch in Doubt, Supply Chain Says No Memory and Substrate Orders Yet

According to The Elec, the prospects for the release of NVIDIA's Rubin CPX GPU for inference have become uncertain. Several industry chain sources have indicated that NVIDIA has not placed any memory and substrate orders related to the Rubin CPX, nor has it made any corresponding development requests.

NVIDIA had announced at the AI Infra Summit in September last year that the Rubin CPX would be released in the second half of 2026, with plans to equip it with 128GB of GDDR7 memory (comprising 8 units of 16GB GDDR7 chips), and to place the memory around the GPU substrate in a board-mounted fashion.

The Elec cited a memory industry source stating that while NVIDIA had clearly stated that the Rubin CPX would use GDDR7, the industry chain has not seen actual progress.

“NVIDIA had said that the Rubin CPX would be equipped with GDDR7, but there isn’t even a discussion about it now. There was previously a possibility of evaluating the use of HBM, but there has been no progress so far, and the actual development direction is not clear.”

Similar feedback was given from the substrate industry. An individual from the substrate industry stated that with no movement on the Rubin CPX, the industry has essentially considered the project as cancelled.

“Due to the lack of movement regarding the Rubin CPX, the industry actually views the project as having been cancelled. We had anticipated that the application range of GDDR7 would expand, but the market ultimately did not open up, which is regrettable.”

Alternative Shift to Groq LPU

The Rubin CPX was initially positioned for AI inference rather than large-scale training scenarios. As inference servers have lower requirements for extreme memory bandwidth, NVIDIA opted for the more cost-effective and less supply-chain-intensive GDDR7 instead of HBM, which is significantly more expensive, power-consuming, and difficult to package. The memory and substrate industries had hoped that the Rubin CPX would promote GDDR7 into mass production, but currently, GDDR7 is still mainly limited to a few consumer-grade high-end graphics cards such as the GeForce RTX 5090 and 5080, with far less application than expected.

The Elec notes that NVIDIA has formally removed the Rubin CPX from its product roadmap at GTC 2026, just about six months since the initial announcement. In response to inquiries from foreign media, NVIDIA did not respond. Instead, it has turned to Groq’s LPU solution—after signing a licensing agreement with Groq late last year, NVIDIA has positioned the Groq 3 LPX as the core inference product of the Vera Rubin platform. It is reported that the two parties reached a deal valued at $20 billion last year, and NVIDIA has also absorbed Groq's core inference technology and engineering talent, which is widely seen in the industry as a de facto acquisition.