Nvidia's most advanced AI server system was meant to be its clearest statement of dominance in 2027. It is now delayed to 2028, and several related plans have been scrapped entirely.
The system in question, Kyber NVL144, is a server rack that packs 144 of Nvidia's most powerful chips into a single unit that behaves like one enormous computer. The bottleneck is a central circuit board inside it, a 78-layer board that routes signals between all those chips. That board has proven extremely difficult to manufacture without defects. Nvidia's CEO Jensen Huang had stood on stage at the company's GTC conference just three months ago to show it off.
The problems go beyond a single delay. A backup design called NVL72x2, which would have connected two existing server racks back to back as a stopgap, has been canceled. Large cloud operators pushed back against it: the format was unusual and expensive to run. Separately, the more powerful version of the next chip generation, called Rubin Ultra, has been scaled back from a four-chip to a two-chip design. The reason is that connecting four large chips in a single package causes the base material to warp and bend under heat, breaking electrical connections. The revised version delivers roughly half the computing power Nvidia originally announced.
A key technology that would allow Rubin Ultra to scale to very large systems, called co-packaged optics, where optical signal components are built directly into the chip package, will not arrive until the generation after next. That means Nvidia's highest-end offering will have real limits on how large a cluster it can build, at least for now.
The market reaction was sharp and immediate. Circuit board suppliers across Asia sold off hard. Ibiden of Japan, whose largest customer is Nvidia, fell 10 percent. Kingboard Laminates in Hong Kong fell 18 percent. Elite Material in Taiwan dropped 10 percent. Samsung Electro-Mechanics in South Korea slid 11 percent. The losses look dramatic in isolation. They look even more so against the backdrop: Samsung Electro-Mechanics had gained over 600 percent this year before this drop, and Kingboard Laminates over 470 percent. These stocks were priced for a future where Nvidia kept hitting every deadline on its aggressive annual release schedule. That pricing is now being reassessed.
Nvidia has not confirmed any of this. The company did not respond to CNBC's request for comment. The source is SemiAnalysis, a well-regarded semiconductor research firm, and the reporting has been picked up by CNBC and Bloomberg. That combination gives it weight, but the official picture remains unconfirmed.
For anyone who buys or plans AI services, the practical picture does not change much today. Nvidia's existing systems continue to ship. The Rubin generation, in its scaled-back form, is still targeted for 2027. What changes is the competitive picture for 2027 and 2028.
AMD's next server chip series is aimed squarely at the same 2027 window. Google is now selling its custom AI chips, called TPUs, to external customers for the first time, with its eighth generation targeting the same period. Major AI companies including Anthropic and Meta are actively building capacity on non-Nvidia hardware. Nvidia still holds roughly 80 percent of the market for AI server chips. But a delay at the high end, combined with a downsized chip and a missing scaling technology, hands competitors a window that did not exist six months ago.
The broader pattern here is worth noting. Nvidia has been operating on an annual hardware release schedule, something no chip company had attempted at this scale before. The Kyber delay suggests that schedule has run into the physical limits of what can actually be manufactured. A 78-layer circuit board connecting 144 high-performance chips inside a liquid-cooled rack is not a product that bends to a press release timeline. Physics moves slower than roadmaps.