Nebius Benchmark Finds Cosmos 3 Super Throughput Gains Come With Longer Latency

Nebius benchmarked NVIDIA Cosmos 3 Super text-to-video serving across four GPU layouts, finding that eight single-GPU replicas produced more valid video per node-hour on HGX B200 while taking substantially longer per request. The results show a practical tradeoff: maximize batch output or return individual clips sooner.