Nvidia’s next-generation Kyber NVL144 rack-scale AI system has reportedly been delayed by more than a year after manufacturing challenges emerged with a critical component at the center of the platform, slowing the company’s aggressive annual product rollout even as demand for AI computing infrastructure remains strong.
The reported setback comes at a time when Nvidia is navigating an increasingly complex operating environment. Alongside unprecedented demand for AI chips from cloud providers, the company has also been dealing with export restrictions affecting sales to China, while governments continue to prioritize domestic semiconductor production and supply chain resilience following years of geopolitical tensions.
According to research firm SemiAnalysis, the Kyber NVL144 rack architecture is now expected to launch in 2028 instead of its previously planned 2027 debut because the PCB midplane, a specialized multi-layer printed circuit board that links key modules inside the rack, has proved difficult to manufacture, CNBC reported.
Kyber is designed to house 144 of Nvidia’s highest-performance graphics processing units (GPUs) inside a single rack, allowing them to operate as one integrated computing system for training and running large artificial intelligence models. The system uses vertically mounted compute trays instead of the conventional horizontal arrangement to increase computing density while reducing latency. Nvidia unveiled the architecture earlier this year alongside its Vera Rubin Ultra platform during its GTC developer conference.
The manufacturing issue centers on what SemiAnalysis described as the PCB midplane remaining “challenging from a manufacturability standpoint,” according to the firm’s post. The component acts as the primary electrical connection between modules throughout the rack, making it essential to the system’s operation.
Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint. NVL576, which connects 8x Oberon racks over CPO between the NVSwitches, is also likely delayed or restricted to small volumes given the current… pic.twitter.com/rGtjAsi927
— SemiAnalysis (@SemiAnalysis_) July 5, 2026
The research firm also reported that Nvidia’s larger NVL576 system, which is designed to connect eight racks using optical interconnects, is also likely facing delays or may initially ship only in limited quantities.
The reported delay follows another change to Nvidia’s roadmap. SemiAnalysis said the company had also abandoned an interim NVL72x2 “back-to-back” rack design after cloud service providers and hyperscale customers objected to its complexity and operational demands. The proposal would have combined two current-generation racks to deliver similar computing capability while Kyber was being prepared for launch.
According to Seeking Alpha, citing CNBC’s reporting and the SemiAnalysis research, the cancellation of the backup architecture leaves Nvidia without an alternative large-scale rack design for Rubin Ultra while Kyber remains delayed. The publication also noted that SemiAnalysis expects Nvidia’s current-generation Rubin systems to begin shipping this fall to major cloud providers including Amazon Web Services, Microsoft Azure and Google Cloud.
Nvidia has accelerated its product cadence from multi-year GPU refreshes to annual platform introductions in response to surging AI infrastructure demand. The company has rapidly moved from Hopper to Blackwell and then to the Rubin generation as hyperscalers continue investing heavily in AI data centers.
Despite the reported delay, SemiAnalysis maintained a positive outlook for Nvidia’s near-term data center business. The research firm projects the company’s data center compute revenue in the second half of fiscal 2027 will exceed Wall Street expectations by roughly 20%. Nvidia’s Rubin systems are already in full production and are scheduled to begin shipping later this year to eight cloud partners, including Amazon Web Services, Microsoft Azure and Google Cloud.
