DeepSeek's DSec: Inside the Agent Training Infrastructure

Episode 1 · DeepSeek's DSec Paper Reveals Massive Agent Training Sandbox Infrastructure (2026-09-23, 8 posts)

On September 23, DeepSeek published a paper on arXiv titled "DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale," revealing its production sandbox infrastructure for large-scale agentic training and evaluation. The paper has over 100 co-authors (some reports say more than 130), including founder Liang Wenfeng. It is DeepSeek's first public disclosure of the scale of its agent-training compute architecture.

Confirmed

• A single production DSec unit spans 160 nodes with 30K CPU cores and 250TB of memory
• Designed to run 3 million AI agent environments per day, creating 5,000+ sandboxes per second
• Supports over 380K peak concurrent sandboxes in production
• Unifies multiple execution forms including FnCall, containers, and microVMs
• Per @jiqizhixin, the platform underpinned RL training and evaluation from V3.2 through V4.1
• Per @Polymarket and @zhidxw, the system includes built-in guardrails against improper agent behavior
• The paper lists over 100 authors (over 130 per @zhidxw), including Liang Wenfeng

Why it matters

• The paper offers a rare look at a top LLM lab's agent-training infrastructure scale, providing a reference for estimating compute requirements for agentic RL
• 5,000+ sandboxes per second and 380K peak concurrency indicate massive-throughput agent training and evaluation, signaling DeepSeek's strong commitment to agents
• The unified FnCall/container/microVM design is directly useful to engineering teams building similar sandbox infrastructure

• DeepSeek Elastic Compute Paper Reveals 160-Node Units Serving 3M Sandboxes Daily — zephyr_z9 · 2026-09-23
• DeepSeek Unveils DSec: 3M Agent Environments Daily, 380k Concurrent Sandboxes — jiqizhixin · 2026-09-23
• DeepSeek unveils DSec sandbox infra running 3M agent environments per day — jiqizhixin · 2026-09-23
• DeepSeek paper details DSec: 3M sandboxes a day for Agent training, and the agents fight back — 量子位 · 2026-09-23
• DeepSeek paper introduces DSec, elastic sandbox infra for agentic training at scale — pstAsiatech · 2026-09-23
• DeepSeek's 31-page paper reveals DSec, the sandbox infra behind its agentic RL training — 机器之心 · 2026-09-23
• DeepSeek paper unveils DSec: 3M sandboxes a day powering agent RL training — 智东西 · 2026-09-23
• DeepSeek unveils agent training system running up to 380,000 sandboxes in parallel — Polymarket · 2026-09-24

Episode 2 · DeepSeek Sandbox Cluster Details Emerge, Estimated $20M Per Unit (2026-09-23, 3 posts)

Analysts estimate DeepSeek's DSec sandbox cluster costs under $20M in materials per unit. Its scale was corrected to 320 AMD EPYC 9655 CPUs (30,720 cores) across 160 nodes.

• DeepSeek's DSec sandbox fleet: ~384 EPYC 9655 CPUs powering 380k concurrent sandboxes at ~30% utilization — teortaxesTex · 2026-09-23
• DeepSeek DSec cluster BoM estimated at ≤$20M; $1B could buy 50 clusters and 19M concurrent sandboxes — teortaxesTex · 2026-09-23
• DeepSeek's DSec Sandbox Fleet Is ~320 EPYC 9655 Chips, 30,720 Cores — teortaxesTex · 2026-09-23

Episode 3 · DeepSeek Reveals DSec Elastic Compute Platform Behind Its RL Training (2026-10-02, 2 posts)

DeepSeek published a 31-page paper on DSec, its elastic compute sandbox infrastructure for large-scale agent RL training. The platform runs about 3 million sandboxes daily with up to 380,000 concurrent sandboxes, supporting models from V3.2 to V4.1.

• DeepSeek's DSec paper: 3M sandboxes/day, 380K concurrent, powering V3.2-V4.1 RL — jiqizhixin · 2026-10-02
• DeepSeek details DSec: 380K concurrent sandboxes and 50x oversubscription powering Agent training — 量子位 · 2026-10-03