A single production-scale unit of DeepSeek Elastic Compute, known as DSec, manages 3 million sandboxes daily and sustains 5,000 creations per second. The architecture, which spans roughly 160 nodes and supports over 380,000 concurrent sandboxes, scales the compute necessary for extensive reinforcement learning agentic trajectories, according to a new research paper.

Estimates from Astra place the bill of materials for one DSec cluster at $20 million or less, noting that DRAM prices are a primary cost driver. A hypothetical $1 billion system deploying 20,000 EPYC CPUs and 20,000 GPUs could process 15 trillion output tokens every 24 hours using the V4.1 model. Current production units utilize approximately 384 EPYC 9655 chips, averaging 12.6 concurrent sandboxes per CPU core.