Nvidia's Kyber rack for Rubin Ultra reportedly delayed to 2028, stopgap solution also axed due to customer pushback

NVIDIA's Kyber NVL144 rack delivery has been pushed to 2028, a delay of over 12 months that impacts the planned rollout of Rubin Ultra GPUs in 2027. The setback stems from manufacturing challenges wit... Read original →

NVIDIA’s Kyber rack delay to 2028 reveals a systemic bottleneck in AI hardware integration. The all-copper NVLink midplane—designed to replace cables with a rigid PCB—suffers from signal integrity collapse and thermal runaway at high layer counts under 3nm process constraints, forcing Rubin Ultra to abandon its quad-die for a dual-die configuration. This not only derails 2027 AI cluster rollouts but also triggered customer rejection of stopgap solutions like the NVL72x2 dual-cabinet design, highlighting hyperscalers’ intolerance for operational complexity. Technically, co-packaged optics for Feynman-era NVSwitch are now delayed, locking NVIDIA into copper interconnects longer than planned and exposing the physical limits of rack-scale scaling. Geopolitically, reliance on advanced PCB and EUV capacity concentrated in Taiwan, China and South Korea amplifies supply chain fragility. AMD and Intel may exploit this gap with modular MI300X or Gaudi 4 deployments in 2027. Over the next 18 months, the industry will recalibrate toward engineering feasibility over peak performance density.