NVIDIA's Kyber rack system for Rubin Ultra slips to 2028

A rare NVIDIA roadmap wobble: SemiAnalysis reports that Kyber, the rack-scale system built to house 144 of NVIDIA's 2027 Rubin Ultra chips, has slipped more than twelve months to 2028 due to manufacturing difficulties, per reporting picked up by CNBC. Rack-scale integration — power, cooling, interconnect and packaging at the whole-rack level — has become as hard as the silicon itself, and Kyber's delay suggests the bottleneck is increasingly systems engineering, not just chips.

The timing is awkward given NVIDIA's simultaneous message of overwhelming demand: the company said AI infrastructure demand has already exceeded its prior $500 billion forecast and unveiled financing with Sharvin AI and Firmus to help emerging cloud providers deploy up to 210,000 Blackwell GPUs via revenue-sharing. H200 and Blackwell lead times remain stretched at 12-20 weeks for large orders. Strong demand plus a slipping next-gen roadmap is a tension worth watching.

Competitively, the delay lands as customers hedge. DeepSeek is developing its own inference chip, and Perplexity CEO Aravind Srinivas said the company is running its sandbox infrastructure on NVIDIA's new Vera CPUs — a reminder that even as buyers diversify, NVIDIA's CPU and networking businesses are expanding. NVIDIA also pushed physical-AI narratives this week with Isaac GR00T humanoid workflows and a Nemotron industrial alarm-triage agent, and touted 100M Nemotron downloads.

Watch whether the Kyber slip opens a window for AMD, custom silicon, or Chinese alternatives to gain share in the 2027 timeframe, and whether NVIDIA reframes its roadmap at its next event.