[NVIDIA's New Rubin Platform Performance Expectations Rise, but Kyber Cabinets Reportedly Delayed to 2028] BlockBeats News, September 15: In its latest morning report, TMT Breakout cited a SemiAnalysis report stating that the unit power token throughput of NVIDIA's Vera Rubin platform in pre-release testing could reach up to approximately 7 times that of Blackwell. Even though the hourly rental cost per chip for Rubin is higher than Blackwell Ultra, in inference scenarios with 80 TPS and P90 latency, the token output corresponding to its three-year rental total cost may still be about 62% higher, with advantages in high-interaction workloads potentially expanding further. However, alongside rising performance expectations, new variables have emerged in the system-level delivery timeline for Rubin Ultra. The Kyber NVL144 cabinet, originally planned to house 144 Rubin Ultra GPUs, may face delays of over 12 months from the originally scheduled 2027 timeline due to PCB mid-board manufacturing challenges, pushing it to 2028. The alternative NVL72x2 solution is also reportedly canceled. This risk will primarily impact high-density cabinets, NVLink interconnects, and large-scale system integration. Even if Rubin chips themselves progress on schedule, the ability of cloud providers to achieve 144-card scale expansion by 2027 may still be constrained, making AI infrastructure investments more reliant on existing architectures like Oberon as alternative solutions. NVIDIA responded to related reports by stating that 'the product roadmap remains unchanged,' but has not confirmed the Kyber delay or any adjustments to the overall Rubin Ultra timeline.