Nebius cut agent rollout collection time by 66.9%

Nebius reduced fixed-weight collection time for a 256-trajectory batch by 66.9% on a Terminal-Bench-based workload through changes to routing, sandbox execution and batch assembly. VP of AI Training Di Jin advises measuring time to a trainable batch rather than tokens per second alone.

• 256-trajectory batch collection time fell 66.9% on Terminal-Bench workload
• Measurement excluded trainer forward, backward and optimizer work
• Key levers were cache-aware routing, sandboxes and batch assembly
• CPU and sandbox capacity should scale independently of GPUs