NVIDIA Kyber NVL144 rack delayed to 2028 amid PCB midplane manufacturability challenges

NVIDIA's next-generation Kyber NVL144 rack architecture, designed to pack 144 of its most powerful chips into a single cabinet for training and running advanced AI models, has been delayed by more than 12 months to 2028, according to research firm SemiAnalysis. The setback stems from difficulties manufacturing a key circuit board at the heart of the system, and it opens a rare opportunity for rivals like AMD and Google at the high end of the AI hardware market.

What happened

SemiAnalysis reported that the Kyber NVL144 rack architecture has been delayed to 2028 because the PCB midplane remains challenging from a manufacturability standpoint. The PCB midplane is a specialized, multi-layer printed circuit board that connects electronic modules within the system. The NVL576, a larger system linking eight Kyber racks via optical connections, is also likely delayed or limited to small volumes.

Adding to the strain, a backup plan to bolt two current-generation racks together for similar power has been scrapped. Cloud service providers and hyperscalers pushed back against the design, calling it awkward and costly to operate. SemiAnalysis stated the backup was cancelled due to heavy pushback from CSPs and hyperscalers over its odd design and heavy operational burden.

Why AI builders should care

The delay underscores a broader risk: NVIDIA's annual release cadence may be outpacing manufacturing capabilities. For teams building large-scale AI training or inference clusters, this means the expected density and latency improvements from Rubin Ultra may not arrive until 2028. In the near term, current Rubin systems are in full production and begin shipping this fall to eight cloud partners, including AWS, Microsoft Azure, and Google Cloud. SemiAnalysis still projects NVIDIA's data-center compute revenue will run 20% above Wall Street consensus in the second half of fiscal 2027.

Practical implications

SemiAnalysis predicts the delay could give rivals Advanced Micro Devices and Google (with their in-house chips already winning business from top AI labs) a rare technical opening at the high end of the market. With no proven solution to expand the scale-up world size for Rubin Ultra, NVIDIA faces a gap in its high-density roadmap.

For AI builders, the takeaway is to monitor AMD's MI-series and Google's TPU roadmaps for alternative high-density compute options. If your training workloads depend on NVLink-scale-up domain, plan for the possibility that the highest-density NVIDIA racks may not be available until 2028.

Caveats

These timing details rely on SemiAnalysis summaries and CNBC reporting. NVIDIA did not respond to CNBC's request for comment. Specifics about NVL144 and NVL576 deployment volumes and exact production timelines could change as NVIDIA or its suppliers provide updated guidance. The original Rubin Ultra itself was also reportedly cancelled as a four-die design and replaced by a dual-die version, compounding the uncertainty.

FAQs

What is NVIDIA Kyber NVL144 and how does it differ from current NVL682/GB200 systems?

Kyber NVL144 is a rack-scale architecture that packs 144 GPUs into a single cabinet using vertical compute trays to boost density and reduce latency. It was slated to debut with Vera Rubin Ultra in 2027 but is now delayed to 2028.

Why has NVIDIA's Kyber NVL144 rack been delayed to 2028?

The delay is attributed to manufacturability challenges with the PCB midplane at the heart of the system, according to SemiAnalysis.

What impact does the delay have on Rubin Ultra chips and data-center deployments?

The Kyber NVL144 delay pushes the high-density rack infrastructure for Rubin Ultra chips to 2028. Current Rubin systems are still shipping this fall to eight cloud partners, so near-term deployments continue on existing rack designs.

Could competitors gain market share due to this delay and how might NVIDIA respond?

Analysts cited by CNBC and SemiAnalysis indicate that rivals like AMD and Google could gain opportunities in the high-end AI hardware market if NVIDIA cannot scale Rubin Ultra quickly enough. NVIDIA has not provided updated guidance on its response.