Nvidia delays Kyber AI rack to 2028 over manufacturing issues

Nvidia's next-generation Kyber rack-scale architecture, a cornerstone of its future AI infrastructure plans, has been delayed by more than a year to 2028 due to significant manufacturing challenges. The setback, reported by semiconductor research firm SemiAnalysis, impacts the system designed to house Nvidia's upcoming Rubin Ultra processors and raises questions about the company's aggressive product roadmap.

The delay stems from difficulties manufacturing a complex printed circuit board (PCB) midplane, a critical component within the Kyber NVL144 system. This specialized board, which reportedly consists of 78 layers and is prone to defects, is essential for enabling the dense, high-speed interconnects that allow the 144 advanced GPUs within a single Kyber cabinet to function as a unified supercomputer. The larger NVL576 system, which would link eight such racks via optical connections, is also likely to be delayed or limited in production volume as a result.

Compounding the issue, Nvidia has scrapped a contingency plan known as NVL72x2, which aimed to link two of its current-generation racks to approximate Kyber's power. Major cloud providers and hyperscale customers reportedly...