Crusoe trained FLUX.1 in 23.63 minutes on 72 Blackwell Ultra GPUs, just under NVIDIA's own 72-GPU time of 23.68 minutes, Crusoe said.
Virtualization did not cost speed.
All 10 results ran on virtualized Crusoe Cloud instances that customers rent today, Crusoe said, using KVM with 1:1 GPU passthrough. The full GB300 NVL72 rack ties 18 nodes of four GPUs into one NVLink domain, Crusoe said, so gradients stay off the network fabric. The comparison pool is large. MLCommons lists 24 submitters in Training v6.1, including CoreWeave, Lambda and Nebius, per MLCommons, which makes parity with the reference system the number to watch.
Crusoe put GPT-OSS 20B at 22.94 minutes on 64 B300 GPUs and 15.41 minutes on 128 GPUs, Crusoe said. Crusoe described the 128-GPU run as the largest HGX B300 configuration submitted for that benchmark in the round. The 64-GPU time matched bare metal on premises systems at the same scale, Crusoe said.
Llama 2 70B LoRA moved from 6.82 minutes on 8 B300 GPUs to 3.93 minutes on 16 GPUs and 2.55 minutes on 32 GPUs, Crusoe said. That is 2.7 times faster on four times the GPUs over Quantum-X800 InfiniBand, Crusoe said.
What the times do not show is cost, power or variance across repeats. Published pricing or energy per run would change the read for buyers who rent by the hour.
Can a customer reproduce NVIDIA reference speed on a rented rack without benchmark tuning?





