[Exclusive] NVIDIA's "Rubin CPX" Launch in Doubt… No Memory or Substrate Orders

The launch of NVIDIA's inference GPU "Rubin CPX" has become uncertain. Although the company had initially announced plans to release it in the second half of this year, it has been confirmed that there are no moves to order or develop the associated memory and substrates. This has led to speculation that NVIDIA may have withdrawn or substantially revised its original plans.

According to multiple industry sources on the 27th, NVIDIA is currently not placing orders for memory or substrates related to Rubin CPX. In particular, no activity has been confirmed around GDDR7 graphics DRAM, the memory it was reported to use. Nor have there been any related development requests.

A memory industry source said, "NVIDIA stated that it would use GDDR7 in Rubin CPX, but at the moment there are no discussions about it whatsoever," adding, "There was also a possibility that HBM adoption was being considered, but with no progress made, the actual development direction is unclear."

A substrate industry source said, "With no movement around Rubin CPX, the industry effectively regards the project as cancelled," adding, "We had expected GDDR7's range of applications to expand, but it's disappointing that the market never opened up."

NVIDIA announced at the AI Infra Summit last September that it would release Rubin CPX in the second half of this year. It specifically stated that Rubin CPX would carry a total of 128GB of GDDR7 memory — a configuration of eight 16GB GDDR7 chips. Unlike high-bandwidth memory (HBM), which is mounted inside the GPU package, it was planned to use an onboard approach, with the memory placed around the GPU substrate.

Because inference AI servers do not require the extreme memory bandwidth that training GPUs do, GDDR memory was adopted. HBM offers fast data processing but carries heavy cost and power burdens, and is also more difficult to package. GDDR, by comparison, has a relatively lower cost burden and is easier to scale in supply.

The memory industry had expected that Rubin CPX would trigger a full-scale expansion in GDDR7 supply. Currently, GDDR7 is used mainly in a handful of high-performance graphics cards such as NVIDIA's GeForce RTX 5090 and 5080, leaving its applications limited. The substrate industry likewise anticipated benefiting from expanded supply of memory substrates for GDDR7.

However, NVIDIA removed Rubin CPX from its roadmap at GTC 2026, held this past March — a change in direction roughly six months after it was unveiled with a second-half launch target. Some foreign media outlets asked about the reasons behind this, but NVIDIA offered no answer.

Instead, the company put front and center the LPU (Language Processing Unit) from Groq, with which it signed a licensing agreement late last year. The "Groq 3 LPX" abruptly became the core inference product of the Vera Rubin platform. Last year, NVIDIA struck a deal worth $20 billion with Groq and absorbed its core inference technology and engineering talent. The industry views this as NVIDIA having effectively acquired Groq.

Rubin CPX was a product NVIDIA had been preparing in response to the expanding inference market. Whereas NVIDIA's existing GPUs focused on large-scale data training, Rubin CPX was specialized for inference — the execution stage of AI services. It was a strategy aimed at a market shift in which inference computing is taking up a growing share amid the spread of AI agents.

As NVIDIA strengthens its Groq-centered inference strategy, observers suggest that the very direction of Rubin CPX may have changed. Some also raise the possibility that NVIDIA could redesign Rubin CPX in a new form down the line.

$NVDA

likes: 230 | retweets: 27 | replies: 9 | views: 55832