DeepSeek founder Liang Wenfeng has co-authored a new paper detailing the technical architecture of DSec (DeepSeek Elastic Compute), a sandbox platform designed for large-scale AI agent training. The publication, which lists over 130 authors including Liang, marks the first systematic release of technical specifications for the company's agent infrastructure. The DSec platform operates at significant scale, with a single production unit comprising approximately 160 CPU nodes, 30,000 cores, and 250TB of memory to host petabyte-level images. Performance metrics indicate the system serves roughly 3 million sandboxes daily with peak concurrency exceeding 380,000, supports creation rates above 5,000 per second, and can launch up to 32,000 sandboxes simultaneously for individual training tasks.