NVIDIA Delays Kyber NVL144 AI Racks to 2028 and Cancels Rubin Ultra Chip Option
NVIDIA's Kyber NVL144 rack architecture has been delayed until 2028, postponing delivery by more than 12 months, according to research from SemiAnalysis. The delay comes just three months after CEO Jensen Huang demonstrated the system at the company’s GTC technology conference and stems from manufacturability challenges with its printed circuit board midplane. The setbacks also force NVIDIA to cancel its alternative NVL72x2 back-to-back rack design, which hyperscale cloud providers rejected over concerns about operational complexity.
The production delays extend to NVIDIA’s upcoming AI accelerators, forcing the company to scrap a 4-chip configuration of its Rubin Ultra chip and leave only a 2-chip variant that delivers roughly half the real-world performance. With the higher-end Kyber rack and advanced accelerator plans scaled back, NVIDIA plans to prioritize shipments of its current-generation Oberon Rubin systems. The timeline leaves a gap in NVIDIA’s ability to scale large language models this year, potentially allowing rival hardware like AMD’s MI500X to capture more hyperscaler orders.
From the sources (10 posts)
@jukan05@JulienTechInvst AMD, yes. But I’m not so sure about TPUs.
@scaling01this surely also pushes Feynman back and with that 100T models
@jukan05@StreetSignal__ I agree. It also makes me wonder whether AMD might struggle to meet its own specs as well.
@jukan051. Bullish for the NPO supply chain. 2. The spec downgrade of NVIDIA Rubin Ultra signals the erosion of NVIDIA’s performance moat. 3. Bearish for the CPO supply chain. 4. Big winners: AMD and the TPU ecosystem?
@semianalysis_As the NVIDIA roadmap indicates, CPO NVSwitch will not be available until Feynman. As a result, NVIDIA currently has no proven solution to expand the scale-up world size for Rubin Ultra, leaving a gap for competitors like AMD MI500X or TPUv
@semianalysis_This news also comes as the 4-compute-die Rubin Ultra has been cancelled, leaving only the smaller 2-compute-die Rubin Ultra, which will deliver roughly half the real-world performance of the 4-die Rubin Ultra. 5/6🧵
@semianalysis_NVIDIA will sell significantly more Oberon Rubin racks and Oberon Rubin “Ultra” racks to make up for this shortfall. We discuss the implications of these mass NVIDIA delays and cancellations for the memory, PCB, and ODM supply chains in ou
@semianalysis_MASSIVE DELAY: Just 3 months after Jensen demoed Kyber NVL144 at GTC, it has faced major setbacks and has been delayed by more than 12 months, pushing it back to 2028. Below, we explain why Kyber has faced massive delays and why NVIDIA’s NV
@semianalysis_Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint. NVL576, which connects 8x Oberon racks over CPO between the NVSwitches, is also likely delayed or restricte
@semianalysis_NVL72x2 back-to-back rack architecture was the new proposed architecture NVIDIA was developing as an alternative to Kyber. It was designed to increase the pure-copper NVLink scale-up world size by placing two Oberon racks back-to-back. Howe