NVIDIA's Kyber NVL144, demoed just three months after GTC, has been delayed by more than 12 months, pushing its release back to 2028. The company also cancelled its NVL72x2 back-to-back rack architecture, which limits the scale-up domain for Rubin Ultra.
NVIDIA's next-gen compute scale-up is now significantly constrained by delays and architectural cancellations, impacting future large-scale AI cluster deployments.