DeepSeek details an AI training system built for scale
The move
DeepSeek published a paper on September 23 describing DSec, an internal system for managing sandboxes used to train AI agents. A single production unit spans about 160 nodes, supports roughly 3 million sandboxes a day and can create more than 5,000 sandboxes per second. Founder Liang Wenfeng is listed among the paper's more than 130 authors.
Why it matters
It shows DeepSeek is putting serious engineering effort into the infrastructure behind agentic AI, not just model releases, as it competes with far more heavily funded US labs.
The bear case
Publishing a research paper isn't the same as shipping a product; there's no indication yet that DSec powers any consumer-facing DeepSeek release.
What to watch
Whether DeepSeek ties DSec to a new agent product launch.