NVIDIA’s plans for its next-generation AI computing architecture have hit an unexpected snag, with the company’s Kyber rack-scale system, designed to power its 2027 Rubin Ultra chips, now delayed by more than a year to 2028, according to research firm SemiAnalysis.
Highlighting the growing complexities of scaling advanced AI hardware at unprecedented levels.
Kyber represents a significant leap in system design, packing 144 of NVIDIA’s most powerful chips into a single server cabinet to function as one massive computing unit. This configuration is intended to deliver the immense processing power required for training and running the largest AI models. The architecture features vertically mounted graphics processing units in compute trays, an innovation aimed at maximizing density and minimizing latency compared to traditional horizontal layouts.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The delay stems from challenges in manufacturing a critical multi-layer printed circuit board, known as the PCB midplane, that serves as the system’s central nervous system, connecting various electronic modules, SemiAnalysis reported on Monday.
“Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint,” the firm said.
A larger related system, NVL576, which would link eight racks through optical connections, is also likely to face delays or limited initial availability, according to the research firm.
This setback adds to a series of reported challenges across NVIDIA’s product development pipeline, raising questions about whether the company’s aggressive annual release schedule is beginning to strain manufacturing capabilities and supply chain partners. A backup approach, combining two existing-generation racks to approximate Kyber’s capabilities, has also been abandoned after cloud customers pushed back against what they viewed as an awkward and operationally burdensome design.
“It has since been cancelled due to heavy pushback from CSPs [cloud service providers] and hyperscalers over its odd design and heavy operational burden,” SemiAnalysis noted.
As a result, NVIDIA currently lacks a proven solution for expanding scale-up capabilities for its Rubin Ultra platform, potentially creating an opening for competitors like Advanced Micro Devices and Google, whose in-house chips have already secured business from major AI laboratories.
Despite the Kyber delay, NVIDIA’s core business remains exceptionally strong. Its current-generation Rubin systems are in full production and scheduled to begin shipping this fall to eight major cloud partners, including Amazon Web Services, Microsoft Azure, and Google Cloud.
SemiAnalysis projects that NVIDIA’s data center compute revenue will exceed Wall Street consensus estimates by 20% in the second half of fiscal 2027. Shares of NVIDIA showed little movement in premarket trading, last down less than 0.1% at $194.79.
The Challenges of Scaling AI Infrastructure
The Kyber delay underscores a fundamental reality of the AI boom: while demand for computing power continues to grow exponentially, the physical and engineering challenges of building ever-larger systems are becoming more pronounced. Creating a rack that can efficiently house and interconnect 144 high-performance chips requires overcoming significant hurdles in thermal management, power delivery, and signal integrity — challenges that appear to have proven more difficult than anticipated for the specialized circuit board at Kyber’s core.
The AI hardware ecosystem is deeply mired in complexities. As models grow larger and training requirements expand, the supporting infrastructure must evolve in tandem. Manufacturing specialized components at the necessary scale and precision is pushing the limits of current production capabilities, even for established leaders like NVIDIA.
The rejection of the interim rack-combining solution by cloud providers further illustrates the practical considerations that go beyond raw performance. In data center environments, operational efficiency, manageability, and cost-effectiveness are critical factors. Designs that create additional complexity or operational burden are likely to face resistance, regardless of their theoretical capabilities.
Analysts believe the delay could provide a window of opportunity for NVIDIA’s competitors. This is because Advanced Micro Devices has been gaining traction with its MI series accelerators, while Google’s custom tensor processing units have secured significant internal usage and external customers. If NVIDIA cannot deliver large-scale solutions on its original timeline, some AI developers may explore alternatives more aggressively.
However, experts warn that it would be premature to view this as a fundamental threat to NVIDIA’s market position. The company continues to dominate the AI accelerator space, with robust demand for its current-generation products and strong revenue projections. Its ecosystem of software tools, developer support, and established customer relationships provides a significant moat that competitors will find difficult to overcome quickly.
Looking ahead, it is believed that NVIDIA’s ability to maintain its leadership will depend on how effectively it addresses the Kyber manufacturing challenges. The company has a strong track record of resolving technical hurdles, and its substantial resources and expertise position it well to overcome current obstacles.
The delay, while notable, appears contained and does not impact near-term product shipments. For customers, the postponement may require adjustments to deployment timelines for the most ambitious AI training initiatives. However, NVIDIA’s existing product lines and incremental improvements are likely sufficient for the majority of applications in the near term.