According to the latest report from SemiAnalysis, a semiconductor industry research organization, NVIDIA's rack-level architecture product Kyber NVL144 for the next generation Rubin Ultra chip will be postponed from its original launch of 2027 to 2028 due to technical difficulties encountered in the manufacturing process of key circuit boards. This delay adds another variable to the already tight supply chain and product rhythm, and also intensifies external concerns about the contradiction between Nvidia's high-intensity annual update rhythm and manufacturing capabilities.
Kyber is a cabinet-level system for data centers. The design goal is to integrate 144 of NVIDIA's most powerful AI chips in a cabinet, allowing them to work together as a "supercomputer" to provide a computing power basis for training and inference of the most advanced artificial intelligence models. Unlike traditional horizontally placed GPUs, Kyber installs the GPU vertically in the computing tray to increase system density and reduce latency. It was originally planned to be officially launched in 2027 with Nvidia's next-generation rack-level system Vera Rubin Ultra.
According to SemiAnalysis, the core reason for this delay is a dedicated multi-layer printed circuit board (PCB midplane) at the heart of the system. This midplane is responsible for connecting various electronic modules within the system. There are currently significant challenges in manufacturability. "The Kyber NVL144 rack architecture has been pushed back to 2028 due to ongoing manufacturability challenges with the PCB midplane," the report states.
In terms of larger-scale configurations, the NVL576 system, which uses optical interconnects to connect eight cabinets together, is also considered likely to be delayed or will only be available in minimal shipments for some time to come. Nvidia has not yet responded to the above reports.
This is not the first time Nvidia’s product line has been exposed to schedule pressure. SemiAnalysis pointed out that Kyber's delay combined with the tight situation of many previous product lines highlights that Nvidia is facing the upper limit test of manufacturing capabilities and process complexity while maintaining the "annual update" rhythm. While encountering obstacles on the manufacturing side, a set of "alternative plans" previously proposed by the company also failed - the plan was to physically "bundle" two current-generation cabinet systems to achieve a computing power scale close to Kyber's.
However, this “splicing solution” has not been recognized by cloud service providers and hyperscale data center customers. SemiAnalysis quoted that the plan was eventually canceled because its design form was considered "weird" by customers, and the operating costs and operation and maintenance burden were too high, triggering strong opposition from cloud service providers (CSPs) and hyperscale customers. Accordingly, the research institution believes that NVIDIA currently "does not have a proven solution to expand the scale-up world size of Rubin Ultra", which may open a rare technology gap for competitors in the high-end market.
In the field of high-performance AI computing, Nvidia’s main competitors include self-developed chip manufacturers such as AMD and Google. SemiAnalysis pointed out that the high-end products of these opponents and the self-developed chips used internally by Google have won some business in leading artificial intelligence laboratories. Kyber's delay may provide them with further opportunities to enter the high-end market.
Despite this, NVIDIA’s current-generation Rubin systems are still on track for mass production and deployment. According to reports, this generation of systems has entered full production and is expected to begin shipping to eight cloud partners including Amazon Cloud Service (AWS), Microsoft Azure and Google Cloud this fall. SemiAnalysis also expects revenue from Nvidia's data center computing business to be about 20% higher than Wall Street consensus estimates in the second half of fiscal 2027.