NVIDIA’s next-generation AI cabinet system postponed to 2028
Industry body SemiAnalysis said that core circuit board manufacturing is too difficult and the release of the Kyber NVL144 cabinet system has been delayed to 2028. This product delay has exacerbated market concerns: Nvidia's rapid iteration of new products every year has hit the upper limit of hardware manufacturing capacity and technology. Waiting for competitors to create a rare technology window in the high-end computing power market. On March 16, 2026, at the keynote speech at the NVIDIA GTC Conference in San Jose, California, NVIDIA CEO Jensen Huang stood next to the Vera Rubin Ultra Kyber computing tray and the Kyber NVLink orthogonal mid-plane and spoke. As the core manufacturer of the explosive expansion of AI computing power, NVIDIA expects that the Blackwell and Rubin series of chips will generate at least US$1 trillion in revenue for the company by the end of 2027. According to SemiAnalysis, a semiconductor industry research organization, Nvidia’s next-generation flagship computing product, the Kyber cabinet-level architecture specially built for the 2027 Rubin Ultra chip, will be delayed by more than 12 months, and its launch time will be postponed to 2028. A series of product setbacks has caused the market to cast a lot of doubts on the AI giant's product roadmap. Kyber is a high-density AI server cabinet. A single cabinet can integrate 144 top-level NVIDIA GPUs. Multiple chips work together to form a giant supercomputer, providing sufficient computing power support for AI companies to train and run ultra-large-scale models. This architecture installs the GPU vertically on the computing tray (instead of the traditional horizontal placement) to increase computing power density and reduce communication latency; it was originally planned to be released simultaneously with the new generation rack system Vera Rubin Ultra in 2027. SemiAnalysis said in an industry analysis article published on Monday that the root cause of this delay is that the mass production process of the key circuit boards at the core of the system is difficult to overcome. The original text of the agency wrote: "Due to the manufacturing process difficulties of the orthogonal midplane (PCB Midplane), the Kyber NVL144 cabinet architecture has been postponed to 2028." The orthogonal midplane is a special multi-layer printed circuit board. As a high-speed interconnection hub for the electronic modules inside the machine, the Kyber architecture adopts a super-high-density design of up to 78 layers, paired with high-end M9 The processing yield, impedance consistency, and heat dissipation control difficulty of copper-clad laminates and quartz electronic cloth are far higher than those of conventional PCB products. The agency also pointed out that the larger-scale system NVL576, which interconnects 8 cabinets through optical fiber, is likely to be delayed simultaneously, or will only be trial-produced in small batches. Nvidia did not respond to a request for comment. This product delay further exposed the production capacity pressure faced by NVIDIA's entire product line, and the market became increasingly worried: NVIDIA's aggressive annual new product iteration rhythm has exceeded the upper limit of existing hardware manufacturing processes. NVIDIA originally prepared an alternative plan: splicing two sets of existing first-generation cabinets back to back to achieve similar computing power scale. However, cloud service providers generally reported that the plan had a strange structure and high operation and maintenance costs, and this backup plan has now been completely cancelled. SemiAnalysis said: "All major hyperscale cloud vendors are strongly opposed to this design. The operation and maintenance burden is too heavy, and the plan has been scrapped." SemiAnalysis review, which means NVIDIA There is currently no mature and feasible solution , used to expand the ultra-large-scale computing cluster of Rubin Ultra chips. Institutions predict that this will bring rare high-end market technology opportunities to AMD and Google - the two companies' self-developed chips have already won orders from many leading AI laboratories. NVIDIA's current generation Rubin system is in full production and will be available this fall Delivered in batches by eight major cloud partners including Azure and Google Cloud. SemiAnalysis also predicts that Nvidia’s data center computing business revenue in the second half of fiscal year 2027 will be 20% higher than Wall Street’s consensus forecast.
Nvidia's stock price fluctuated during the pre-market trading session, with the latest price quoted at $194.79, a decrease of less than 0.1%.
Nvidia's stock price fluctuated during the pre-market trading session, with the latest price quoted at $194.79, a decrease of less than 0.1%.