NVIDIA Kyber NVL144 Delayed to 2028: Production Issues Slow Down AI Roadmap
NVIDIA's journey towards the next generation of ultra-dense AI infrastructure has hit a snag. According to research firm SemiAnalysis, the rack-scale system Kyber NVL144 has been delayed by more than a year, with its debut now expected in 2028 instead of 2027.
Kyber is one of NVIDIA's most ambitious projects: a server cabinet capable of integrating up to 144 GPUs based on the Rubin Ultra architecture, designed to transform an entire rack into a single coherent computing unit for training and inference of next-generation AI models. The goal is to drastically increase computational density and reduce latency between processing units.
MASSIVE DELAY: Just 3 months after Jensen demoed Kyber NVL144 at GTC, it has faced major setbacks and has been delayed by more than 12 months, pushing it back to 2028. Below, we explain why Kyber has faced massive delays and why NVIDIA’s NVL72x2 back-to-back rack architecture was…
The delay is reportedly linked to production issues with the midplane PCB, a multi-layer printed circuit board that manages interconnections between the various system modules. SemiAnalysis highlighted how production complexity is slowing down the project's scalability.
In addition to Kyber NVL144, the broader NVL576 - which envisions interconnecting multiple racks through optical links - may also experience delays or be limited to reduced volumes. In the meantime, NVIDIA has reportedly scrapped a backup plan that involved aggregating two current racks, a solution that faced resistance from cloud providers due to costs and operational complexity. On the competitive front, this scenario might open up opportunities for competitors, particularly AMD and Google.
The Rubin GPUs are still in full production and are expected to start shipping this fall to cloud partners like AWS, Microsoft Azure, and Google Cloud. At this time, NVIDIA has not released official comments on the situation.