Nvidia's next-generation AI server rack has reportedly been pushed back by about a year, according to a report from the research firm SemiAnalysis.
The delayed product is Kyber, a rack architecture designed to house a large number of Nvidia's Rubin Ultra chips. According to SemiAnalysis, its release has slipped to 2028. Data Center Dynamics, summarizing the report, said the delay follows manufacturing concerns.
Racks like Kyber matter because modern AI systems aren't built around single chips anymore. They are assembled as tightly packed racks — in this case referred to as Kyber NVL144 — that link many processors together to train and run the largest AI models. A delay to the rack, not just a chip, can ripple through the plans of the cloud providers and AI companies that buy this hardware in bulk.
The news moved Nvidia's stock. As Yahoo Finance reported, NVDA shares swung in premarket trading on the report. Notably, the reaction was not uniformly negative across the sector: Finimize reported that tech futures rose even as Nvidia's AI rack faced the delay.
It's worth stressing what is and isn't confirmed here. The reported 2028 timeline and the manufacturing concerns behind it come from SemiAnalysis and secondary coverage of that report, not from an official Nvidia announcement in these sources. Details such as the exact original launch date were not specified in the items available.
Why it matters: Nvidia's roadmap sets the pace for the entire AI infrastructure buildout, so even a rumored one-year slip in its flagship rack is enough to move markets and reshape how much compute the industry can expect — and when.