Nvidia has delayed Kyber, the next-generation "rack-scale" architecture meant to power its most advanced AI systems, according to research firm SemiAnalysis. The system, formally called Kyber NVL144, has been pushed back by more than 12 months, moving its launch to 2028.
Kyber was designed to house Nvidia's Rubin Ultra chips, which are due in 2027, per reporting from CNBC's Anniek Bao citing SemiAnalysis. SemiAnalysis attributes the slip to problems manufacturing printed circuit boards (PCBs) — the layered boards that connect the components inside a server. As part of the same shift, SemiAnalysis reports that Nvidia has also canceled a related design known as NVL72x2.
The news rippled quickly through the supply chain. Bloomberg reports that the delay report sent shares of Asian PCB makers sliding, a sign of how closely those suppliers' fortunes are tied to Nvidia's roadmap.
Nvidia has built its dominance partly on an aggressive, predictable release cadence for AI hardware, and its roadmap is watched closely by data-center operators and investors alike. Outlets including Techzine and GuruFocus framed the delay as a setback to that ambitious schedule.
A "rack-scale" architecture bundles many chips into a single tightly connected unit, effectively turning a whole server rack into one giant accelerator. Delays to that kind of flagship product can affect when cloud providers and AI companies get access to Nvidia's fastest hardware.
Why it matters: Nvidia's timelines set the pace for the entire AI computing industry, so a year-plus slip in a marquee product — and the immediate hit to its suppliers' stocks — signals that even the sector's leader can be tripped up by manufacturing bottlenecks.