DeepSeek Expands Elastic-Computing Team, Details DSec Agent Sandbox Infrastructure

DeepSeek is hiring heavily for its elastic-computing team, with a particular need for senior engineers, according to a report by QbitAI. Rather than attaching a conventional job description, the company pointed to a technical write-up titled "DeepSeek Elastic Computing (DSec): Sandbox Infrastructure for Large-Scale Agent Training."

The numbers in that write-up are substantial. One DSec scale-out shard runs on roughly 160 servers with 30,000 CPU cores and 250TB of memory, per the report. At peak more than 380,000 sandboxes are online at once and more than 5,000 are created every second, and DeepSeek says it has deployed several such shards in production, capable of supporting millions of sandboxes simultaneously.

DSec provides the sandbox infrastructure for agent training. According to the write-up, it underpins the entire training, evaluation and data-preprocessing pipeline for DeepSeek's flagship models; every sandbox workload generated by agent training, evaluation and data prep runs on DSec.

That workload looks different from ordinary cloud computing. During agent training, a model repeatedly enters real environments to perform tasks — reading code, editing files, installing dependencies, running tests and launching services. Each step mutates the environment, and the next turn must continue from the previous state. Sandboxes therefore have to be created quickly and at scale, yet cannot be torn down and rebuilt after every step.

To cut the cost of provisioning environments, DSec splits the runtime into three layers — a base image for the OS and base software, a workspace for task code and dependencies, and a toolkit for tools — managing each version separately and combining them when a sandbox is created. It also places image data in DeepSeek's own 3FS distributed file system, keeping only metadata locally and reading data on demand. In a test creating 8,192 containers, on-demand loading cut the task from more than 60 minutes to about 35, while disk writes fell about 57%.

On resource efficiency, DSec found that agent sandboxes are remarkably sparse: 90% use on average less than 5% of their requested CPU, because CPUs sit idle while agents wait for the model to generate the next action, even though files and processes must be kept alive in memory. DeepSeek says its production oversubscription rate now exceeds 50x.

Together, the details sketch a clear bet. As agents are asked to operate ever more complex real software and systems, training them requires being able to build and supply runtime environments cheaply and at scale. Whoever makes this "invisible infrastructure" more efficient is better positioned to push agents from demos into production.

What to watch: DeepSeek says it plans to scale the number and variety of agent environments by orders of magnitude, support four execution backends — FnCall, Container, MicroVM and Full VM — and even use agents to build the environments that agents run in. Whether that path works will shape the ceiling of its next-generation models' agentic abilities.