Nvidia's next-generation Kyber NVL144 AI server rack delayed to 2028 due to PCB midplane manufacturing defects; Rubin Ultra variant canceled.

Nvidia will not deliver its Kyber NVL144 rack until 2028, a delay exceeding 12 months that postpones the cabinet designed for 2027's Rubin Ultra GPUs into the following year, according to analysis from SemiAnalysis. The delay stems from manufacturing challenges with a PCB midplane that interconnects eight Oberon racks via NVSwitches—what Nvidia calls the orthogonal backplane. Compounding the setback, Nvidia has discontinued the NVL72x2, a stopgap solution intended to bridge customers until Kyber availability, leaving no proven alternative to scale Rubin Ultra deployments in 2027.

The orthogonal backplane is central to Kyber's design. It sits between vertically mounted compute trays and switch trays, replacing earlier racks' cable harnesses with a rigid board carrying an all-copper NVLink fabric. Kyber employs liquid cooling by default and accommodates 144 Rubin Ultra packages—double the 72 in a current Oberon NVL72 rack. Every GPU-to-GPU link within the cabinet traverses this board, but copper traces face signal-integrity degradation as layer counts increase, alongside challenges in power delivery and thermal management.

The backplane represents a sophisticated engineering feat. Trade analyses describe three 26-layer sections laminated into a 78-layer stack approximately one square meter in area, with trace spacing at or below 25 micrometers and impedance held within 5% tolerance to maintain 448 Gb/s-class signaling integrity. A cabled equivalent would demand upward of 20,000 discrete cables, motivating Nvidia's shift to a single passive board.

The cancelled NVL72x2 would have coupled two Oberon racks back-to-back to achieve Kyber-class density over copper NVLink. Nvidia scrapped the design after major customers refused to operate two linked cabinets as a single unit. Separately, NVL576—a configuration tying eight racks together through co-packaged optics—is likely to experience delays or limited availability until that optical technology matures.

These cancellations leave Nvidia without a proven means to expand Rubin Ultra's scale-up capabilities, meaning the largest single Rubin Ultra domain in 2027 could match, but not surpass, what Oberon already delivers. The situation is further constrained by Nvidia's recent decision to abandon a quad-chiplet Rubin Ultra GPU design in favor of a dual-chiplet part, halving the accelerator's per-package compute. SemiAnalysis projects that a fully production-ready co-packaged optics NVSwitch will not arrive before the Feynman generation succeeding Rubin, leaving copper as the only near-term solution for linking Rubin Ultra at rack scale.

The delay affects only Rubin Ultra and its Kyber rack. Nvidia's 2026 Rubin GPUs, which reuse the current Oberon rack, are not subject to this reported delay.