Nvidia's Kyber Server Racks for Rubin Ultra Delayed, Suppliers' Shares Drop Over 10 Percent

Nvidia's next-generation server rack architecture, designed to house next year's Rubin Ultra chips, will slip by more than a year to 2028, according to research firm SemiAnalysis. Nvidia has firmly denied the report, but markets reacted nervously, especially among Asian component suppliers.

At the heart of the dispute is a technically complex component: the midplane board connecting compute modules inside the Kyber rack. It is a 78-layer PCB design that, according to SemiAnalysis, has proven harder to manufacture at scale than expected. The issue centers on the so-called orthogonal backplane, the element responsible for moving enormous volumes of data between processors packed densely into the rack.

Nvidia's response

An Nvidia spokesperson dismissed the SemiAnalysis report with a brief statement, cited by Tom's Hardware and Benzinga. The company offered no additional technical details and did not rebut the analysts' findings point by point, sticking instead to a general assurance that its plans remain on track.

Our roadmap remains unchanged - Nvidia spokesperson, quoted by Tom's Hardware

Market analysts read the statement in different ways. Daniel Newman of Futurum Group noted that the wording leaves room for interpretation, but stressed that fiduciary duties to shareholders would likely rule out making such an assurance carelessly. Patrick Moorhead of Moor Insights & Strategy called Nvidia's response broadly reassuring. CNBC commentator Jim Cramer went further, suggesting it was a signal to buy the stock.

The scrapped backup plan

The situation is further complicated by the fate of Nvidia's plan B. The company had reportedly prepared a bridge solution, a rack called NVL72x2, built by combining two of the current Vera Rubin generation designs. It was meant to deliver similar computing power until Kyber was ready. According to SemiAnalysis, cloud customers rejected the design as impractical and costly to operate, and Nvidia subsequently abandoned it. That leaves the company without a proven alternative to fill the Rubin Ultra scaling gap in 2027.

Market reaction

Nvidia itself reacted to the news relatively calmly, with its shares falling about 1.4 percent on Monday, a move investors read as a temporary setback rather than a serious crisis given the company's roughly $4.7 trillion market cap. Component suppliers in Asia's PCB manufacturing chain fared far worse: Japan's Ibiden, Hong Kong's Kingboard Laminates, and Taiwan's Taiwan Union Technology each lost more than 10 percent of their value in a single trading day. These are the companies that produce the specialized multilayer printed circuit boards underpinning the entire Kyber architecture.

What it means for the industry

The current generation of Rubin systems remains in full production and is set to reach eight cloud partners this fall, including Amazon Web Services, Microsoft Azure, and Google Cloud - the delay applies only to the next step, Rubin Ultra and its accompanying Kyber architecture. Even so, industry commentators, including FourWeekMBA, note that even a one-year slip in the most advanced, densely packed AI rack segment could give rivals such as AMD or Google a rare chance to close the gap at the highest technological tier.

For Polish companies relying on public clouds, the direct impact of this delay should be limited in the near term, since availability of the current Rubin generation is not at risk. What matters more is the pricing and competitive outlook: if Nvidia genuinely loses a year of lead time in the densest server segment, it could slow the pace at which AI computing costs fall, costs that underpin cloud services used in Poland as well.

Sources: Tom's Hardware (tomshardware.com), Benzinga (benzinga.com), 24/7 Wall St (247wallst.com), CNBC (cnbc.com)