Nebius cut rollout collection time by 66.9% — here’s what changed
The New Stack Mark Chillingworth
Nebius says it cut rollout collection time by 66.9% in an agent training setup. The big win wasn’t bigger GPUs; it was fixing the stuff around them that kept making them wait.
Based on reporting by The New Stack, Mark Chillingworth — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
In agent training, the GPU is often not the thing slowing you down. Di Jin, Nebius’s VP of AI Training in Amsterdam, argues that the real choke point is usually the rollout system: the code that runs agents through tasks, collects their attempts, and feeds those results back into training. If that pipeline is badly designed, expensive hardware sits idle while the model waits for stragglers, sandboxes, browsers, or verifiers to finish.
That’s why Jin says teams should measure time to a trainable batch, not just tokens per second. In synchronous reinforcement learning, one slow trajectory can hold up the whole step. A batch may look fast on average, yet still stall because training cannot move on until the last agent attempt is done. His point is blunt: if no one owns inference inside the training project, inference becomes the bottleneck.
The other trap is stale data. Jin says some policy lag can help because generation and training can overlap, which improves utilization. But that only works if teams know how far the policy has drifted and how to handle the mismatch. His line is memorable for a reason: staleness is not the problem; unmeasured staleness is. The fix is to track divergence, use corrections where needed, and drop or down-weight samples once they go too far off the current policy.
He also wants teams thinking less about algorithms in the abstract and more about environments. For agentic RL, the environment is the data source, so reproducible containers, realistic websites, codebases, APIs, and good verifiers matter as much as the model itself. If the task set gets too easy, the learning signal dries up. If the verifier is sloppy, the model learns the loophole instead of the job.
Nebius tested that idea on a Terminal-Bench-based workload and says changes to routing, sandbox execution, and batch assembly cut fixed-weight collection time for a 256-trajectory batch by 66.9%. That figure covers batch collection only, not the full training loop. Still, it makes the same point the whole piece keeps making: in agent training, the cheapest container can end up deciding what the priciest GPU does all day.
My take — AI-written commentary, not fact-checked reporting
This is the kind of unglamorous infrastructure work that actually moves AI forward, which is bad news for people who prefer demo reels to plumbing. The industry keeps pretending model quality is mostly about more scale, but a lot of the pain is just bad scheduling and lazy ownership. The future belongs to teams that treat rollout systems like first-class software, not a side quest with a fancy budget.
Read more about this at: The New Stack