NVIDIA's Kyber rack delayed over a year to 2028 amid Rubin manufacturing snags

NVIDIA's Kyber rack-scale architecture—intended to house the 2027 Rubin Ultra generation—has reportedly slipped more than 12 months to 2028 due to manufacturing challenges, per SemiAnalysis. Rack-scale systems like Kyber (successor to the GB200/GB300 NVL72 designs) are how NVIDIA packages GPUs, networking and cooling into deployable AI supercomputers, so a year-plus delay ripples into hyperscaler capacity planning.

Mechanically, the snags reportedly trace to Taiwan manufacturing and the complexity of the Rubin platform; rack-scale integration multiplies the number of things that can go wrong versus shipping discrete chips. A slip here doesn't stop current GB300 shipments (Grok 4.5 trained on GB300 this week) but pushes out the next big capacity step.

Competitively the timing is uncomfortable: it lands as DeepSeek builds its own inference silicon and ZML pushes cross-chip inference to break NVIDIA lock-in—both signs of a market actively hedging against NVIDIA dependence. NVIDIA simultaneously deepened its software moat, partnering with Hugging Face on the LeRobot robotics platform and touting Nemotron 3 Ultra's cost wins. Skeptics will watch whether the delay opens a window for AMD or custom silicon. Watch NVIDIA's official roadmap commentary and whether hyperscalers adjust orders in response.