"The roadmap is intact!" NVIDIA quickly refutes rumors of delays in the next-generation AI rack, while Wall Street criticizes it as a "publicity stunt."

Zhitong
2026.07.07 02:51

NVIDIA refuted rumors that the next-generation AI architecture "Kyber" has been postponed to 2028, with a spokesperson stating that its roadmap remains intact. Previously, a SemiAnalysis report indicated that the Kyber architecture and NVL72x2 design were canceled or delayed due to customer objections and CPO challenges

According to the Zhitong Finance APP, NVIDIA (NVDA.US) has refuted a report claiming that its next-generation "Kyber" artificial intelligence (AI) rack will be delayed until 2028. A company spokesperson stated, "Our roadmap is intact."

On Sunday, SemiAnalysis released a report stating that the Kyber rack-level architecture, designed to accommodate its 2027 Rubin Ultra chip, will be delayed by more than 12 months, pushing it to 2028.

SemiAnalysis wrote in a post on X, "Just three months after Jensen Huang showcased the Kyber NVL144 at GTC, it is facing significant setbacks and has been delayed by more than 12 months to 2028." In the following, we will explain why Kyber is facing a substantial delay and why NVIDIA's NVL72x2 back-to-back rack architecture has also been canceled, thereby limiting the expansion domain of Rubin Ultra.

SemiAnalysis also stated that NVIDIA canceled the originally planned NVL72x2 back-to-back rack design as an alternative plan due to strong objections raised by hyperscale cloud service provider customers during actual operations. Additionally, given the current challenges faced by CPO (Co-Packaged Optics), the larger-scale system NVL576, which connects eight Oberon racks through CPO, "may also be delayed or limited to small batch shipments."

Kyber is a server cabinet that packages 144 of NVIDIA's most powerful chips in a single unit, allowing them to work together like a giant computer, providing the power needed for AI companies to train and run their cutting-edge models.

The design installs GPUs in vertically rather than horizontally placed computing trays to increase density and reduce latency, and is scheduled to debut in 2027 alongside NVIDIA's next-generation rack-level system, Vera Rubin Ultra.

Earlier, a supply chain integrator for NVIDIA servers in Taiwan revealed that even if the Kyber rack's mass production is delayed, it does not affect NVIDIA's chip upgrades, nor does it impact NVIDIA's standing in the industry, and it poses no shock to the existing supply chain.

Mizuho Securities trading department analyst Jordan Klein also stated on Monday that investors should have seen this kind of scenario many times before, with news of NVIDIA's new products being delayed due to manufacturing issues constantly emerging, but this is all just a publicity stunt, pure noise.

Paul Triolo, a partner at consulting firm DGA-Albright Stonebridge Group, believes that the rumors of delays "should not be over-interpreted as affecting NVIDIA's long-term core importance in AI data center infrastructure." NVIDIA has faced and resolved such challenges before. However, since power supply remains one of the main constraints on U.S. AI data center spending, delays in the delivery of advanced systems could give the U.S. more time to overcome certain core power bottlenecks

Source: Zhitong The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.