x.com/chatgpt21/status/2042135575390572654

This sentence understates Rubin's published FP4 training performance by about 2×.

NVIDIA's January 2026 technical blog lists Rubin GPU NVFP4 training = 35 PFLOPS per GPU for dense compute, and NVFP4 inference = 50 PFLOPS per GPU. Separately, NVIDIA's HGX Rubin NVL8 specs list 280 PFLOPS NVFP4 training for an 8-GPU system, which is again 35 PFLOPS per GPU. That directly contradicts the post's claim of 17.5 PFLOPS per GPU.

Because the post's training-time estimate is based on that per-GPU throughput, the quoted 4-day / 40-day figures are also off by about 2× under the same assumptions. Using the standard dense-transformer training-compute estimate the post is implicitly using (~6 × parameters × tokens):

• 100T parameters × 1T tokens ≈ 6 × 10^14 × 10^12 = 6 × 10^26 FLOPs
• 100,000 Rubin GPUs at 35 PFLOPS/GPU provide 100,000 × 35 × 10^15 = 3.5 × 10^21 FLOPs/s
• 6 × 10^26 / 3.5 × 10^21 ≈ 171,429 seconds ≈ 1.98 days

For 10T tokens, that scales linearly to about 19.8 days, not 40 days.

So the numerical claim is incorrect on two levels: the Rubin per-GPU FP4 training figure is wrong, and the derived time estimates are therefore about double what NVIDIA's published specs imply.