Nvidia Delays Kyber Server System by a Year, Opening Door for AMD and Google

Analytics firm SemiAnalysis reports that Nvidia is delaying the rollout of its Kyber NVL144 server system by more than a year, pushing it from 2027 to 2028. The cause is said to be manufacturing problems with the extraordinarily complex circuit board underpinning the entire architecture.

Kyber NVL144 is a server rack designed to house 144 of Nvidia's most advanced chips in a single enclosure, delivering unified AI computing at large scale. A bigger variant, the NVL576, which links eight such racks via optical interconnects, may also be delayed or limited to a small production run.

What is causing the delay

According to SemiAnalysis, the problem lies in the 78-layer midplane PCB, one of the most complex boards ever designed for commercial computing hardware. The firm points to difficulties spanning signal integrity, power delivery, thermal design, and simply whether manufacturing that many layers is feasible in practice.

Nvidia had considered a fallback plan of pairing two current-generation racks to approximate the same computing power. That idea was scrapped, however, after cloud customers rejected it as impractical and costly to operate. That leaves Nvidia without a proven way to scale Rubin Ultra in the near term.

Market reaction

Despite a delay of more than twelve months, the stock market's reaction was relatively muted. Nvidia shares fell 1.4 percent to $194.83 on Monday, following a 5 percent decline over the past month and trading roughly 18 percent below their 52-week high of $236.54. Asian PCB makers tied to Nvidia's supply chain also declined.

The relatively calm investor response stems partly from the fact that the current generation of Rubin systems remains in full production and is expected to begin mass shipping this fall. SemiAnalysis also notes that Nvidia's data center revenue in the second half of fiscal 2027 is expected to beat Wall Street consensus by 20 percent, which softens the picture.

An opening for rivals

The biggest beneficiaries of the delay could be AMD and Google, with its own TPU chips. Nvidia's lack of a proven path to scale Rubin Ultra means cloud providers planning purchasing budgets for the coming years may redirect part of their orders to rivals rather than wait for Kyber to become available.

It is a rare situation in a market where Nvidia has for years set the pace for the entire industry. For data center operators and public cloud providers, including those in Europe, it means having to revise hardware procurement plans for the next generation of AI models, which could push back deployments planned for 2027 and 2028.

The episode also shows how heavily the entire AI industry now depends on single, highly specialized links in the manufacturing chain, such as multilayer PCBs, where even minor engineering problems translate into months-long delays for billions of dollars worth of computing infrastructure.

Sources: MLQ News (mlq.ai), TIKR (tikr.com), Seeking Alpha (seekingalpha.com)