Nvidia Kyber NVL144 pushed to 2028, suppliers tumble

Nvidia pushed the Kyber NVL144 AI server rack back more than twelve months to 2028 after repeated failures making the system's PCB midplane, the central circuit board that links components, according to a SemiAnalysis report published alongside industry market moves on July 6, 2026. The disclosure coincided with sharp, double-digit declines in several Asian suppliers.

What was delayed and why?

SemiAnalysis says the Kyber NVL144 was delayed to 2028 because the PCB midplane proved extremely difficult to produce without defects; that manufacturing problem is the direct cause of the more-than-twelve-month slip. Jensen Huang had shown Kyber NVL144 three months earlier at Nvidia's GTC conference, but the midplane issue prevented the system from shipping on the previously expected timeline.

The supplier fallout was immediate. Japanese PCB maker Ibiden dropped as much as ten percent. 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. Those declines followed massive prior run-ups: Samsung Electro-Mechanics had gained more than 600 percent this year, and Kingboard Laminates had gained more than 470 percent.

What else was canceled and who gains?

SemiAnalysis reports Nvidia abandoned a planned NVL72x2 rack design and canceled the four-die Rubin Ultra chip, leaving only a two-die Rubin Ultra that delivers roughly half the real-world compute performance of the canceled variant. A planned interconnect called CPO-NVSwitch, intended to link many chips into a single large system, will not arrive until the generation after next, called Feynman.

Those choices remove Nvidia scaling options in the near term. SemiAnalysis says cloud providers and large data-center operators pushed back on the NVL72x2's unusual form factor and high operational overhead, which contributed to scrapping that design. With CPO-NVSwitch delayed until Feynman and the Rubin Ultra four-die canceled, competitors gain scheduling room: the report names AMD's MI500X and Google's TPUv8i Broadfly as alternative platforms that could move into that space.

Nvidia plans to offset the gaps by selling more Oberon-Rubin racks in the existing form factor, according to the report.

Why it matters

The Kyber delay shifts the timing of high-end system deliveries and briefly alters the competitive balance among AI infrastructure providers. Analysts quoted in coverage interpret the market reaction as driven largely by profit-taking rather than a drop in overall AI spending: Gary Tan of Allspring Global Investments said a Kyber delay does not necessarily mean AI spending will shrink, it shows the most ambitious system is taking longer than expected. Shawn Oh of NH Investment & Securities pointed to growing uncertainty around Nvidia's expansion plans, which creates short-term room for alternative AI platforms.

Operationally, losing the four-die Rubin Ultra and deferring CPO-NVSwitch mean Nvidia currently lacks a proven path to scale Rubin Ultra into very large systems. That constraint matters for customers planning systems that depend on dense chip interconnects or the highest single-system compute densities.

What to watch

Track Nvidia's announcements about Oberon-Rubin rack shipments and any technical updates on the PCB midplane manufacturing process. Watch public timing for the Feynman generation and the promised CPO-NVSwitch interconnect as the concrete milestone that would restore Nvidia's large-system scaling option. Also monitor Asian supplier share prices and quarterly updates from Ibiden, Kingboard Laminates, Elite Material, and Samsung Electro-Mechanics for signs the market has reassessed supply risk or demand.