A PCB manufacturing defect has pushed Nvidia's next-generation AI server rack back more than twelve months, rattling Asian suppliers and opening questions about Rubin Ultra's scalability.
Nvidia’s Kyber NVL144 — its next-generation AI server rack — has been pushed back more than twelve months to 2028 due to circuit-board manufacturing problems, according to a report from analyst firm SemiAnalysis posted on X [1].
The delay matters for builders and buyers of large-scale AI infrastructure because Kyber NVL144 is Nvidia’s most ambitious rack system to date, and the setback arrives alongside a string of related cancellations that collectively reduce Nvidia’s near-term compute roadmap [1].
What’s breaking and what’s been cut
The root cause, SemiAnalysis said, is the PCB (printed circuit board) midplane — a central board that connects all individual components inside the rack — which has proven extremely difficult to manufacture without defects [1]. Nvidia chief executive Jensen Huang had demonstrated Kyber NVL144 just three months earlier at the company’s GTC conference [1].
The problems extend beyond the delay itself. A planned alternative design called NVL72x2, which would have placed two Oberon racks back to back, has been scrapped entirely after cloud providers and large data center operators objected to its unusual form factor and high operational overhead [1].
The more powerful four-die version of the upcoming Rubin Ultra chip has also been canceled; only the smaller two-die variant remains, delivering roughly half the real-world compute performance of the original design [1]. A key interconnect technology called CPO-NVSwitch (co-packaged optics NVSwitch), which links many chips into a single large system, won’t arrive until the generation after next — a platform called Feynman — leaving Nvidia without a proven path to scale Rubin Ultra to very large systems in the near term [1].
To compensate, Nvidia plans to sell more Oberon-Rubin racks in the existing form factor [1].
Supplier stocks take a sharp hit
The SemiAnalysis report landed hard on Asian suppliers already sitting on enormous year-to-date gains. Japanese PCB maker Ibiden, which counts Nvidia as its largest customer, dropped as much as ten percent [1]. Kingboard Laminates fell 18 percent in Hong Kong, Elite Material lost ten percent in Taiwan, and Samsung Electro-Mechanics slid eleven percent in South Korea [1]. The declines follow steep run-ups: Samsung Electro-Mechanics had gained more than 600 percent this year, and Kingboard Laminates more than 470 percent [1].
Competitive implications
Gary Tan of Allspring Global Investments told Bloomberg that a Kyber delay does not necessarily mean overall AI spending will shrink, but does show that Nvidia’s most ambitious system is taking longer than expected [1]. Shawn Oh of NH Investment & Securities pointed to growing uncertainty around Nvidia’s expansion plans as a factor that gives alternative AI platforms more room to compete [1].
SemiAnalysis noted that the CPO-NVSwitch gap could give rivals such as AMD’s MI500X or Google’s TPUv8i Broadfly an opening to move in on large-system deployments [1].
Sources
This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.
