Nvidia Kyber Rack System Delayed to 2028 Amid Manufacturing Hurdles
Circuit board production challenges push back the 144-chip architecture designed for Rubin Ultra by more than a year, opening a window for AMD and Google.
Nvidia's ambitious Kyber rack-scale architecture won't arrive until 2028, more than a year later than planned, according to research firm SemiAnalysis. The delay stems from persistent manufacturing difficulties with a critical circuit board component at the heart of the system.
Kyber represents Nvidia's next-generation approach to packing 144 of its most powerful graphics processing units into a single server cabinet, where they function as one unified computing platform. The architecture was originally scheduled to debut in 2027 alongside Vera Rubin Ultra, Nvidia's forthcoming chip generation designed for the most demanding AI workloads.
Manufacturing roadblock
The holdup centers on what engineers call a PCB midplane—a specialized, multi-layer printed circuit board that connects electronic modules within the system. SemiAnalysis reported Monday that this component "remains challenging from a manufacturability standpoint," forcing Nvidia to push the timeline back to 2028.
The Kyber design mounts GPUs in vertical compute trays rather than the traditional horizontal orientation, a configuration intended to increase chip density and reduce communication delays between processors. A larger variant called NVL576, which would link eight Kyber racks through optical connections, faces similar delays or will be available only in limited quantities, according to the research firm.
Backup plan scrapped
Nvidia explored an interim solution—connecting two current-generation racks to achieve comparable performance—but abandoned the approach after cloud service providers rejected it. SemiAnalysis noted that hyperscalers and cloud platforms pushed back against "its odd design and heavy operational burden," leading to cancellation of the stopgap architecture.
That leaves Nvidia without a clear path to scale up computing power for Rubin Ultra in the near term, the research firm said. The gap could create an opening for competitors Advanced Micro Devices and Google, both of which have already secured contracts with leading AI labs for their own chip designs.
Why it matters
The delay raises questions about whether Nvidia's aggressive annual product cadence is sustainable given the physical constraints of cutting-edge manufacturing. While the company maintains dominance in AI accelerators, any extended gap in its ability to deliver larger-scale systems could allow rivals to gain ground in the high-end market where AI labs train frontier models. The setback also highlights how even industry leaders face hard limits when pushing the boundaries of chip density and interconnect technology.
Nvidia's current Rubin systems remain on track, with production underway and shipments beginning this fall to eight cloud partners including Amazon Web Services, Microsoft Azure, and Google Cloud. SemiAnalysis projects Nvidia's data center compute revenue will exceed Wall Street expectations by 20% in the second half of fiscal 2027.
Nvidia did not respond to requests for comment. The company's shares were nearly flat in premarket trading following the report.
These details were first reported by CNBC, citing SemiAnalysis research.
This is an original analysis by the Omega editorial team. Source reporting: AI Watch.
Want systems like this working for your business?
Book a Call
