Nvidia’s Kyber May Slip to 2028, Forcing a New Repricing Across the AI Supply Chain
A new SemiAnalysis report says Nvidia’s Kyber NVL144 rack, unveiled by Jensen Huang at GTC 2026, may be delayed by more than 12 months to 2028, while the companion NVL72x2 back-to-back rack architecture has reportedly been canceled. The change would compress Rubin Ultra’s NVLink expansion domain and marks the latest in a series of supply-chain setbacks tied to Nvidia’s next-generation AI infrastructure roadmap. Over the past 30 days, third-party research has pointed to delays in native 800VDC power deployment, potential slippage in co-packaged optics commercialization, and the cancellation of the four-die Rubin Ultra configuration in favor of a two-die version with roughly half the original compute and memory bandwidth. These reports have already triggered sector-level repricing across optical communications, PCB manufacturing, and power infrastructure names. The immediate losers appear to be advanced PCB makers and parts of the CPO ecosystem, while copper cable, connector vendors, and some upstream material suppliers may benefit from a longer lifecycle for the mature Oberon/NVL72 architecture. Nvidia has not officially confirmed the reported delays, leaving investors focused on upcoming earnings commentary and order guidance from Taiwanese ODMs and PCB suppliers for confirmation.



