- SemiAnalysis reports Nvidia’s next-generation Kyber NVL144 AI server rack has been pushed back more than twelve months to 2028 due to circuit-board manufacturing problems.
- The PCB midplane — the central board connecting all components — has proven extremely difficult to produce without defects.
- Asian suppliers fell sharply: Ibiden dropped as much as 10 percent, Kingboard Laminates 18 percent, Elite Material 10 percent, and Samsung Electro-Mechanics 11 percent.
- Nvidia has also scrapped the NVL72x2 rack design and the four-die version of Rubin Ultra, and a key interconnect won’t arrive until the Feynman generation.
What Happened
Analyst firm SemiAnalysis reported on X that Nvidia’s next-generation AI server rack, Kyber NVL144, has been delayed by more than a year to 2028 because of manufacturing problems, as covered by The Decoder on July 6, 2026. The issue is the PCB midplane — the central circuit board that connects all the rack’s components — which has proven extremely difficult to produce without defects. Nvidia CEO Jensen Huang had shown off Kyber NVL144 just three months earlier at the company’s GTC conference.
Why It Matters
The report hit an already jittery investor base: after years of AI-fueled gains, even modest setbacks now trigger sharp sell-offs. Japanese PCB maker Ibiden, which counts Nvidia as its largest customer, dropped as much as 10 percent; Kingboard Laminates fell 18 percent in Hong Kong, Elite Material lost 10 percent in Taiwan, and Samsung Electro-Mechanics slid 11 percent in South Korea. Those falls follow enormous run-ups — Samsung Electro-Mechanics had gained more than 600 percent this year, Kingboard more than 470 percent.
Technical Details
SemiAnalysis lists further setbacks beyond the headline delay. The NVL72x2, an alternative design that would have placed two Oberon racks back to back, has been scrapped entirely after cloud providers and large data-center operators pushed back on the unusual form factor and operational overhead. The four-compute-die version of the upcoming Rubin Ultra chip has been canceled, leaving only the two-die variant with roughly half the real-world compute. And CPO-NVSwitch, the interconnect that links many chips into one large system, won’t arrive until the generation after next, called Feynman — meaning Nvidia currently lacks a proven way to scale Rubin Ultra into very large systems. Nvidia plans to compensate by selling more Oberon-Rubin racks in the existing form factor.
Who’s Affected
The direct impact falls on Nvidia’s Asian supply chain — the PCB and component makers whose stocks sold off — and on the cloud providers planning data-center buildouts around Kyber’s timeline. The delay also hands competitors an opening: AMD’s MI500X and Google’s TPUv8i Broadfly get room to move while Nvidia’s most ambitious system slips. Shawn Oh of NH Investment & Securities said the growing uncertainty around Nvidia’s expansion plans gives alternative AI platforms more space to compete.
What’s Next
Gary Tan of Allspring Global Investments told Bloomberg the delay doesn’t necessarily mean overall AI spending will shrink — it shows Nvidia’s most ambitious system is taking longer than expected, with the current stock weakness mostly profit-taking. The dates to watch are 2028 for Kyber NVL144 and the Feynman generation for CPO-NVSwitch; in the interim, Nvidia’s answer is volume on the existing Oberon-Rubin form factor. Nvidia has not publicly confirmed the SemiAnalysis timeline.
