In a recent post on social media platform X, SemiAnalysis noted that the tech giant’s Kyber NVL144 has now been delayed until 2028. “Kyber NVL144 rack architecture has been postponed to 2028 as the PCB midplane remains difficult from a manufacturability perspective,” the firm stated, alluding to a specialized, multi-layer printed circuit board that connects electronic modules within a system.
The NVL576, a larger system that interlinks eight racks via optical connections, is also expected to face limitations in volume or possible delays, SemiAnalysis elaborated.
MASSIVE DELAY: Just 3 months after Jensen demoed Kyber NVL144 at GTC, it has encountered significant setbacks and has been postponed by over 12 months, now slated for 2028. Below, we explain why Kyber has met with substantial delays and why NVIDIA’s NVL72x2 back-to-back rack architecture was… pic.twitter.com/VYduxnu01B
— SemiAnalysis (@SemiAnalysis_) July 5, 2026
What is Nvidia’s Kyber NVL 144 rack?
The server rack system is intended to accommodate Nvidia’s 2027 Rubin Ultra chips, according to CNBC.
Kyber integrates 144 of Nvidia’s most powerful chips within a single unit, forming a massive computer that provides AI companies the bandwidth necessary for training and deploying advanced models.
It arranges graphics processing units (GPUs) in vertically-oriented compute trays instead of the traditional horizontal layout, minimizing latency and enhancing density.
The design was previously set to launch in 2027 with Vera Rubin Ultra, the company’s next-generation rack-scale system.
How does this impact Nvidia?
This delay is yet another in a series of recent challenges facing the firm led by Jensen Huang. The company’s alternative plan for the Kyber rack—connecting two of its current-generation racks for similar performance—has been scrapped after cloud customers deemed the design inefficient and costly to manage.
The increasing pressure on the company’s product lines raises concerns that Nvidia’s rapid annual product releases are clashing with manufacturing capabilities.
Nvidia currently has “no proven solution to expand the scale-up world size for Rubin Ultra,” pointed out SemiAnalysis.
This situation presents an opportunity for competitors like AMD and Google to gain a significant market share.
Asian technology stocks dipped following the report, according to Bloomberg.
Japan’s Ibiden Co., a PCB manufacturer that regards Nvidia as its largest customer, dropped 10%. Related suppliers such as Kingboard Laminates Holdings Ltd. fell 18% in Hong Kong, while Samsung Electro-Mechanics Co. declined by 11% in South Korea.
A setback in Kyber’s launch does not necessarily indicate a decline in AI capital expenditure, but “likely implies that Nvidia’s most ambitious next-gen system architecture may require more time for deployment, which isn’t entirely surprising,” noted Gary Tan, a portfolio manager at Allspring Global Investments, as reported by Bloomberg.
Nvidia’s current-generation Rubin systems are currently in production, with plans to start shipping this fall to eight cloud partners, including Microsoft Azure, Amazon Web Services, and Google Cloud.
