Cette section n'est pas encore traduite. Le contenu ci-dessous est affiché en anglais, tel qu'il a été publié. Lire la version anglaise.
Last reviewed · How we check sources
NVIDIA Kyber rack-scale systems, explained
Kyber is NVIDIA's next rack architecture, the successor to the Oberon rack used for today's NVL72 systems. It is designed around 800 V DC power and liquid cooling and is meant for the densest AI systems on NVIDIA's roadmap.
What changes in the rack
Kyber turns compute blades vertically, like books on a shelf, with up to 18 compute blades per chassis. NVLink switch blades sit at the back and connect through a cable-free midplane. [1]
How many GPUs
NVIDIA's descriptions have changed. In October 2025 it said Kyber would house 576 Rubin Ultra GPUs by 2027. Its March 2026 roadmap describes Kyber as a rack with 144 GPUs, doubling the NVLink domain per rack, first launched with Vera Rubin Ultra as a standalone NVL144 system. The 576-GPU Vera Rubin Ultra NVL576 is now described as eight separate MGX racks of 72 GPUs in one NVLink domain, and Kyber is meant to scale to an NVL1152 system for the Feynman generation. [1][2]
Power: 800 V DC
Kyber is designed for 800 V DC power delivered to each compute node, where a 64:1 converter steps it down to 12 V next to the GPU. NVIDIA says full-scale production of 800 VDC data centres will coincide with Kyber in 2027, and that 54 V in-rack power would need up to 200 kg of copper busbar for a 1 MW rack. [1][3][4]
What it means for facilities
NVIDIA frames Kyber as part of the move to megawatt-class racks, with liquid cooling and DC power planned from the building in. For operators that means grid, power distribution and cooling decisions made years before the hardware ships. Dates and configurations are NVIDIA roadmap statements and may change. [1][3]
Sources
- NVIDIA, partners drive next-gen efficient gigawatt AI factories in buildup for Vera Rubin (October 2025) (opens in a new tab) NVIDIA Blog
- NVIDIA Vera Rubin POD: seven chips, five rack-scale systems, one AI supercomputer (16 March 2026) (opens in a new tab) NVIDIA Technical Blog
- NVIDIA 800 VDC architecture will power the next generation of AI factories (20 May 2025) (opens in a new tab) NVIDIA Technical Blog
- Building the 800 VDC ecosystem for efficient, scalable AI factories (13 October 2025) (opens in a new tab) NVIDIA Technical Blog
Written from official sources only. Figures are as published by each source and are not independently verified. Company figures are targets unless stated. Report a correction.