DeepSeek DSec: 160 Nodes Power 380,000 LLM Training Sandboxes — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- DeepSeek Elastic Compute (DSec) is a production sandbox platform that exposes FnCall, container, microVM, and full-VM backends through a unified SDK for large-scale agentic training and evaluation of LLMs.
- DSec is co-designed with DeepSeek's reinforcement learning framework, decoupling stateful rollout execution from preemptible GPU training and coordinating sandbox lifecycle with training to preserve rollout state while reclaiming idle resources.
- A single production-scale unit of DSec spans approximately 160 nodes and serves about 3 million sandboxes per day.
- DSec sustains over 380,000 concurrent sandboxes and more than 5,000 sandbox creations per second, combining memory sharing, reclamation, and CPU scheduling for high-density execution.
- DSec loads image data on demand from the Fire-Flyer File System (3FS), DeepSeek's cluster-wide distributed filesystem, and composes environments from independently versioned layers to reduce setup and image-distribution overhead.
- DSec is described in the paper as mitigating agent misbehavior such as reward hacking through its coordinated lifecycle management with the training loop.
Why it matters: Agentic AI training at scale demands millions of isolated, stateful execution environments per day, and DSec's reported numbers — 3 million daily sandboxes and 380,000 concurrent on roughly 160 nodes — show DeepSeek has engineered infrastructure density that materially lowers the cost ceiling for running RL rollouts against tools and repositories. The co-design with the RL framework and 3FS integration are concrete architectural choices that competitors replicating DeepSeek's agentic training pipeline will have to match.
Ask SkimNews
