--- title: Nebius cut rollout collection time by 66.9% — here’s... description: In agentic reinforcement learning training, expensive GPUs often sit idle waiting on slow rollouts and cheap containers, says Nebius VP of AI Training Di Jin.... canonical: https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Nebius cut rollout collection time by 66.9% — here’s what changed | daily.dev og:description: In agentic reinforcement learning training, expensive GPUs often sit idle waiting on slow rollouts and cheap containers, says Nebius VP of AI Training Di Jin.... og:url: https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3 og:image: https://api.daily.dev/og/posts/RwDQqYWL3.png og:image:alt: Nebius cut rollout collection time by 66.9% — here’s what changed og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Nebius cut rollout collection time by 66.9% — here’s what changed **[The New Stack](https://daily.dev/sources/newstack)** · 8 min read · 0 upvotes · 0 comments ## Summary In agentic reinforcement learning training, expensive GPUs often sit idle waiting on slow rollouts and cheap containers, says Nebius VP of AI Training Di Jin. In a Nebius experiment on a Terminal-Bench-based workload, changes to routing, sandbox execution, and batch assembly cut fixed-weight collection time for a 256-trajectory batch by 66.9%. That figure measured batch collection only, not full training speed. Jin recommends tracking time to a trainable batch, managing policy staleness through divergence measurement rather than a fixed age limit, scaling CPU and sandbox capacity independently of GPUs, and aligning numerics between the trainer and rollout engine. He calls building environments 'the new data curation.' ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much did Nebius reduce rollout collection time in their agentic RL training experiment? Nebius reduced fixed-weight collection time for a 256-trajectory batch by 66.9% in a Terminal-Bench-based workload. The gains came from changes to routing, sandbox execution, and batch assembly. This measurement covered batch collection only, excluding the trainer's forward, backward, and optimizer work, so it does not reflect complete training speed. _Teams tuning RL training pipelines can track infrastructure benchmarks like this one on daily.dev._ ### Should I worry about using stale policy data in asynchronous reinforcement learning training? Staleness itself is not inherently harmful, according to Nebius VP of AI Training Di Jin; unmeasured staleness is the problem. Letting generation use weights that are one or a few updates old lets generation and training overlap, improving GPU utilization. The key is measuring divergence between the generating policy and the current policy rather than applying a universal age limit, and applying importance-sampling corrections or dropping samples beyond a threshold. _Developers designing async RL pipelines can weigh these staleness tradeoffs via daily.dev._ ### Why are my GPUs idle during agentic reinforcement learning training? Expensive training GPUs often wait on cheap, under-provisioned infrastructure like sandboxes, containers, and browsers rather than on compute itself, per Nebius VP of AI Training Di Jin. Rollout generation and trajectory stragglers in synchronous batches dominate wall-clock time. Scaling CPU and sandbox capacity independently and generously, keeping sandboxes on the same network fabric as GPUs, and tracking GPU-hours spent waiting on environments can reduce this idle time. _Engineers debugging GPU utilization in RL pipelines can dig into bottleneck patterns like this on daily.dev._ ## Similar posts on daily.dev - [Deployment Isn't the Bottleneck: Inference Economics Is](https://daily.dev/posts/deployment-isn-t-the-bottleneck-inference-economics-is-joy2kikzs) · DigitalOcean Community · 0 upvotes · 0 comments - [What Actually Mattered in AI — August 2026](https://daily.dev/posts/what-actually-mattered-in-ai-august-2026-ybkm1vfpo) · Medium · 1 upvotes · 0 comments - [Assess the entire inference stack — not just open model benchmarks](https://daily.dev/posts/assess-the-entire-inference-stack-not-just-open-model-benchmarks-dqd19ns0g) · The New Stack · 0 upvotes · 0 comments - [Building a reliable cloud native foundation for distributed AI training](https://daily.dev/posts/building-a-reliable-cloud-native-foundation-for-distributed-ai-training-pwun78uwc) · CNCF · 1 upvotes · 0 comments --- Tags: [#gpu](https://daily.dev/tags/gpu), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#agentic-ai](https://daily.dev/tags/agentic-ai) [View this post on daily.dev](https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Nebius cut rollout collection time by 66.9% — here’s what changed","url":"https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3"},"datePublished":"2026-10-27T15:46:44.815Z","dateModified":"2026-10-27T15:50:47.985Z","description":"In agentic reinforcement learning training, expensive GPUs often sit idle waiting on slow rollouts and cheap containers, says Nebius VP of AI Training Di Jin....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ff796635875fcb61c7b2e00af52cce73?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ff796635875fcb61c7b2e00af52cce73?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The New Stack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The New Stack","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/newstack","url":"https://daily.dev/sources/newstack"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"gpu,reinforcement-learning,agentic-ai","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The New Stack","item":"https://daily.dev/sources/newstack"},{"@type":"ListItem","position":3,"name":"Nebius cut rollout collection time by 66.9% — here’s what changed"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nebius-cut-rollout-collection-time-by-66-9-here-s-what-changed-rwdqqywl3#faq","mainEntity":[{"@type":"Question","name":"How much did Nebius reduce rollout collection time in their agentic RL training experiment?","acceptedAnswer":{"@type":"Answer","text":"Nebius reduced fixed-weight collection time for a 256-trajectory batch by 66.9% in a Terminal-Bench-based workload. The gains came from changes to routing, sandbox execution, and batch assembly. This measurement covered batch collection only, excluding the trainer's forward, backward, and optimizer work, so it does not reflect complete training speed. Teams tuning RL training pipelines can track infrastructure benchmarks like this one on daily.dev."}},{"@type":"Question","name":"Should I worry about using stale policy data in asynchronous reinforcement learning training?","acceptedAnswer":{"@type":"Answer","text":"Staleness itself is not inherently harmful, according to Nebius VP of AI Training Di Jin; unmeasured staleness is the problem. Letting generation use weights that are one or a few updates old lets generation and training overlap, improving GPU utilization. The key is measuring divergence between the generating policy and the current policy rather than applying a universal age limit, and applying importance-sampling corrections or dropping samples beyond a threshold. Developers designing async RL pipelines can weigh these staleness tradeoffs via daily.dev."}},{"@type":"Question","name":"Why are my GPUs idle during agentic reinforcement learning training?","acceptedAnswer":{"@type":"Answer","text":"Expensive training GPUs often wait on cheap, under-provisioned infrastructure like sandboxes, containers, and browsers rather than on compute itself, per Nebius VP of AI Training Di Jin. Rollout generation and trajectory stragglers in synchronous batches dominate wall-clock time. Scaling CPU and sandbox capacity independently and generously, keeping sandboxes on the same network fabric as GPUs, and tracking GPU-hours spent waiting on environments can reduce this idle time. Engineers debugging GPU utilization in RL pipelines can dig into bottleneck patterns like this on daily.dev."}}]} ```