Nvidia's next-generation AI server rack, the Kyber NVL144, faces a major delay of more than a year, now expected in 2028 instead of 2026, according to analyst firm SemiAnalysis. Circuit board manufacturing problems drive the postponement, forcing the company to resolve fundamental production issues before scaling production.
The delay triggered immediate market reaction. Asian suppliers with exposure to Nvidia's supply chain lost up to double-digit percentages in market value as investors reassessed their revenue forecasts. The company also canceled the Rubin Ultra variant, a more powerful version that was part of its roadmap.
The setback matters because Nvidia dominates the AI infrastructure market. Data center operators and cloud providers depend on its latest server racks to train and deploy large language models. A two-year delay opens competitive windows for AMD and Google, both pushing alternative AI accelerator solutions. AMD's MI300X and upcoming MI325X chips give customers another option. Google's custom TPU chips appeal to internal workloads and select partners.
Kyber NVL144 was designed to succeed the current Blackwell-based systems with improved performance density and power efficiency. The server rack targets the next wave of AI model scaling, where performance gains become incremental but costly. Manufacturing delays in advanced circuit boards, likely related to substrate complexity or yield issues, suggest Nvidia hit technical walls in production.
The broader implication cuts deeper than one product delay. It signals that even Nvidia faces manufacturing constraints as AI infrastructure demands accelerate. Circuit board production represents a traditional bottleneck that money cannot easily solve. Fabs and substrate manufacturers operate at full capacity.
For Nvidia, the delay risks losing contractual commitments from major cloud providers planning 2026-2027 deployments. Customers may place orders elsewhere to meet timelines. For competitors, the window to capture market share grows