NVIDIA's post-Rubin roadmap announced at GTC 2026 includes Kyber (Vera Rubin Ultra with vertical 144-GPU racks, shipping H2 2027), Feynman (next-next-gen platform, expected 2028), and CPO switches entering volume production for Rubin Ultra (NVIDIA newsroom, 2026).
Data last verified September 9, 2026 from NVIDIA newsroom and the GTC 2026 keynote.
Why a post-Rubin roadmap matters
Vera Rubin is shipping in Q3 2026. But the AI compute demand curve is steep — inference workloads are doubling every 9-12 months, and physical AI (robotics, autonomous vehicles, digital twins) adds a new growth vector. To stay ahead, NVIDIA needs to announce the next two generations publicly so customers can plan multi-year infrastructure investments (NVIDIA, 2026).
Kyber — Vera Rubin Ultra (2027)
Kyber is NVIDIA's next-rack architecture. Key changes vs Vera Rubin:
- 144 GPUs per rack (vs 72 in Vera Rubin NVL72) — doubles compute density.
- Vertical compute trays — shorter cable length reduces latency by ~40%.
- CPO switches — co-packaged optics integrated into the switch ASIC, replacing pluggable transceivers.
- Liquid-cooled fully integrated — direct-die cooling on every GPU.
Expected shipping: H2 2027. Pricing TBA (NVIDIA, 2026).






