According to SemiAnalysis’s latest report, Rubin Ultra’s HBM has been reduced to 192GB. In memory-sensitive scenarios such as agentic inference, long-context processing, and MoE large models, 192GB will be more constrained than 288GB, meaning its actual effective performance in some cases may approach or even fall below that of the standard Vera Rubin. If the upgraded Ultra Rubin performs worse than the standard Vera Rubin, deploying the same model may require more GPUs, increasing inter-GPU communication, model partitioning, and data movement overhead, reducing per-GPU effective utilization, and raising the cost per generated token. In other words, although GPU output has increased, if each task requires more GPUs, the system-level computational cost may not decrease proportionally—undermining the logic behind the Ultra version. Should we then simply skip the Ultra version and continue selling Vera Rubin through late 2027, then move directly to the next-generation Feynman architecture in 2028? 🤔