Nvidia stock in focus as next AI super-rack faces manufacturing snag

Nvidia stock NASDAQ:NVDA remained in focus on Monday after a SemiAnalysis report revealed that its next-generation Kyber NVL144 rack-scale architecture has reportedly been delayed by more than 12 months.

The research firm said that the product is now expected to launch in 2028 instead of its previously anticipated 2027 timeline.

SemiAnalysis shared the update in a post on X, stating that the delay comes roughly three months after Nvidia Chief Executive Jensen Huang demonstrated Kyber NVL144 during GTC.

According to Odaily, the research firm also said Nvidia's NVL72x2 back-to-back rack architecture has been cancelled, a move it believes limits the scaling capability of Rubin Ultra.

Manufacturing challenges behind the delay

SemiAnalysis attributed the reported delay to manufacturing difficulties involving a key component within the system.

"Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint," the firm said.

The PCB midplane refers to a specialised multi-layer printed circuit board that connects electronic modules within the system.

According to SemiAnalysis, manufacturing this component has proved difficult, resulting in the reported delay.

Kyber is designed as a rack-scale server cabinet capable of housing 144 of Nvidia's most advanced chips within a single system.

The architecture enables the processors to function collectively as one large computing platform, providing the computing power required to train and operate advanced artificial intelligence models.

The system uses vertically mounted graphics processing units arranged in compute trays instead of a traditional horizontal layout.

The design aims to improve computing density while reducing latency.

Kyber had been expected to debut alongside Nvidia's Vera Rubin Ultra rack-scale platform in 2027.

Vera Rubin platform

The Nvidia Vera Rubin platform combines Nvidia Rubin GPUs and Nvidia Vera CPUs connected through Nvidia NVLink-C2C, Nvidia ConnectX-9 SuperNICs, and Nvidia BlueField-4 DPUs within a direct liquid-cooled architecture.

According to the company, the platform is intended for scientific computing workloads by offering native FP64 capabilities for high-accuracy simulations alongside AI performance for surrogate models, scientific foundation models and AI-assisted analysis.

The platform is designed to allow researchers to run numerical simulations, train and deploy AI models, stream data from scientific instruments, and perform real-time analytics on a single system.

Additional architecture changes

SemiAnalysis also reported that Nvidia's NVL72x2 back-to-back rack architecture has been cancelled.

According to the research firm, the alternative design involved combining two of NVIDIA's current-generation racks to deliver computing performance similar to the delayed Kyber platform.

However, the proposal reportedly failed to gain support from major cloud customers.

"It has since been cancelled due to heavy pushback from CSPs and hyperscalers over its odd design and heavy operational burden," SemiAnalysis said.

The research firm added that the cancellation leaves Nvidia with "no proven solution to expand the scale-up world size for Rubin Ultra."

SemiAnalysis also said NVL576, a larger system designed to connect eight racks using optical interconnects, is likely to face delays as well, or may be produced only in limited quantities.

Potential competitive impact

According to SemiAnalysis, the reported setbacks could create an opportunity for competitors.

The firm said the lack of a proven large-scale expansion solution for Rubin Ultra could provide rivals such as AMD and Google, whose in-house AI chips have already secured business from leading AI laboratories, with an opening in the high-end AI infrastructure market.

Despite highlighting the reported delays, SemiAnalysis maintained a positive outlook for Nvidia's business performance.

The research firm said Nvidia's current-generation Rubin systems are already in full production and are scheduled to begin shipping this fall to eight cloud partners, including Amazon Web Services, Microsoft Azure, and Google Cloud.

It also projected that Nvidia's data-centre compute revenue in the second half of fiscal 2027 would exceed Wall Street consensus by approximately 20%.

Nvidia stock barely moved in the premarket trading on Monday and were up 0.19% at the time of writing the report.