If we go by the latest report from SemiAnalysis, the HBM on the Rubin Ultra has dropped to 192GB. In memory-sensitive scenarios like Agentic inference, long-context processing, and MoE large models, 192GB will be more constrained compared to 288GB. This means its actual effective performance in certain scenarios might be close to or even lower than the standard version of Vera Rubin. The upgraded Ultra Rubin performs worse than the standard Vera Rubin. The same model might require more GPUs to complete deployment, which would increase inter-GPU communication, model partitioning, and data transfer overhead, reducing the effective utilization rate of a single GPU and driving up the cost per token generated. In other words, even though GPU production has increased, if each task requires more GPUs, the system-level compute cost might not decrease accordingly. The logic behind the upgrade just doesn’t hold up. So, maybe there’s no need to push the Ultra version. Just keep selling the Vera Rubin version in the second half of 2027 and wait until 2028 to roll out the next-gen Feynman architecture.