Nvidia Kyber rack-scale system reportedly pushed back to 2028 as manufacturing hurdles surface

Nvidia is facing fresh questions about how quickly it can turn new data-center AI platforms into shipments, after multiple outlets reported that its next-generation “Kyber” rack-scale system has been delayed to 2028.

The reports, drawing on an analysis and discussion from SemiAnalysis, point to Kyber NVL144, a rack-scale architecture Nvidia has been positioning as the next step for scaling training and inference workloads beyond the company’s current Oberon-based setup. Kyber systems are designed to bundle GPUs, networking, and supporting infrastructure into a cabinet-like unit so customers can scale large AI clusters in a more repeatable way.

According to the accounts, the delay would move a cabinet intended for scaling “Rubin Ultra” GPUs, which are associated with Nvidia’s upcoming next-generation compute, out of 2027 and into the following year. The time gap matters for buyers because data-center operators typically plan procurement around predictable deployment windows for large training runs and for power and cooling capacity.

Several reports also describe a specific technical bottleneck. SemiAnalysis-linked commentary cited “PCB midplane” manufacturing challenges, a reference to printed circuit board technology used in internal backplane-style connections that help link multiple rack units together through Nvidia’s orthogonal backplane approach. In this context, the midplane is part of the infrastructure that coordinates how compute trays and switching fabric connect at cabinet scale.

The reporting further says Nvidia’s attempted stopgap approach ran into difficulty. One outlet described that a rack meant to tide customers over, referred to as NVL72x2, was reportedly axed after customer pushback. Other reporting echoed the theme that the company may not have an immediately available alternative that can widen Rubin Ultra scale-up in 2027 on the same validated path.

Even so, at least one Yahoo Finance markets item said Nvidia maintained that its overall AI roadmap remained on track following the report. The same reporting framed the Kyber NVL144 slip as a manufacturing-driven issue rather than a broader change to the company’s product strategy.

The episode arrives as the industry increasingly treats “rack-scale” platforms as critical plumbing for AI clusters, not just as a higher-end packaging of chips. For Nvidia, timing is central because its competitive position and customer retention depend on meeting tight deployment schedules for major cloud operators and AI-focused enterprise buyers.

What Nvidia has not disclosed in public statements tied to this specific set of reports is the extent of schedule changes for individual customers, any revised delivery commitments by quarter, or the exact manufacturing recovery timeline. Until Nvidia or its customers provide more detail, the gap between “planned release” and “shipped systems” is likely to remain the key uncertainty for investors and for data-center planners. Next, the market will watch for any confirmation or denial from Nvidia, plus signs in supplier communications and customer procurement updates that would indicate whether 2028 is a hard milestone or a moving target.