Nebius and NVIDIA Benchmark Cosmos 3 Super: Serving World Models Reveals GPU Topology Tradeoffs

NVIDIA and Nebius have published benchmark results demonstrating the latency-versus-throughput tradeoff when serving the Cosmos 3 Super world model, showing that GPU topology choices can improve throughput by 42.5% at the cost of substantially higher latency. The open-weight, 64.6-billion-parameter model, designed for generating physics-consistent video for robotics and autonomous-vehicle simulation, demands a different serving strategy than traditional large language models.