DeepSeek published a number this week that deserves more attention than most model scores.
Three million sandboxes a day.
Its new DeepSeek Elastic Compute system, DSec, runs across roughly 160 nodes in a production deployment. DeepSeek says it can support more than 380,000 concurrent sandboxes and create over 5,000 new ones every second.
Those figures tell us something about the next stage of the AI race.
Frontier labs no longer need GPUs alone. Agent training needs millions of places where models can open files, run programs, compile code, call tools and get things wrong without taking the rest of the cluster down.
DeepSeek has built that layer at industrial scale.
Agents need somewhere to make mistakes
Traditional language-model training is comparatively tidy.
Feed data into accelerators. Calculate loss. Update weights. Repeat.
Agent training introduces a whole external world.
A coding agent may install packages, launch a compiler, change files and run a test suite. Another agent may need a browser or a full virtual machine. Some tasks take seconds. Others leave state behind for much longer.
DSec wraps several kinds of execution environment behind one system, including function calls, containers, microVMs and full virtual machines. It is designed to work directly with DeepSeek's reinforcement-learning infrastructure and keep stateful agent environments separate from the GPU workers training the model.
That separation sounds technical because it is.
It is also economically important.
An expensive accelerator should not sit idle because a test suite is compiling on a CPU somewhere else. Nor should an agent lose half an hour of work because the GPU worker attached to it gets pre-empted.
DeepSeek is solving those scheduling problems because agent training has become large enough for them to hurt.
This is what makes the three million figure interesting.
Nobody builds infrastructure for millions of daily execution environments because they expect agents to remain a side feature.
The next model is already visible in the infrastructure
DeepSeek released V4.1-Flash on 10 September. The company calls it the smallest member of a new architecture family and says the design is intended to scale to larger models. It uses a 552 billion parameter mixture-of-experts architecture, new pretraining techniques and larger-scale reinforcement learning.
DeepSeek has not publicly laid out every detail of an undisclosed successor's training run.
It does not need to for the direction to be obvious.
A new architecture family is being built to scale. A production sandbox system can generate millions of agent environments a day. Around 150 senior backend engineers are being recruited as DeepSeek rewrites infrastructure strained by growth in training jobs, containers, data volume and agent workloads.
Read together, those developments look less like maintenance and more like preparation.
DeepSeek is building the systems required to train and serve a more ambitious generation of models.
That supports the reports and industry expectation that the company is already pushing into its next frontier-model cycle, even where the exact checkpoint schedule remains private.
This is where DeepSeek may have a harder-to-copy advantage
AI models are unusually easy to study from the outside.
A competitor can read the paper. Engineers can inspect open weights. Architecture ideas spread quickly. Training recipes leak into the general research culture.
Production infrastructure is harder to copy.
DSec has to schedule huge numbers of stateful environments, move images efficiently, reclaim memory, isolate hostile code and coordinate all of it with an RL system running on expensive accelerators. DeepSeek's paper also describes mechanisms for dealing with misbehaving agents and reward-hacking behaviour.
None of those engineering problems looks glamorous in a model announcement.
Together, they determine how many useful experiments a lab can afford.
Suppose one laboratory can run an agent task for half the infrastructure cost of another. It can generate more trajectories from the same budget. It can test more reward schemes. It can discard more failures. It can train on harder environments without tying up GPUs unnecessarily.
That advantage compounds if the resulting model becomes good enough to generate more useful training experience for its successor.
DeepSeek's reputation was built partly on doing more with constrained compute. DSec suggests that efficiency work is moving beyond the model architecture into the factory around it.
That is harder for rivals to neutralise with a paper replication.
Three million bad trajectories would still be bad
There is an obvious limit to this argument.
Infrastructure does not create intelligence by itself.
Three million sandboxes can generate three million useless experiences. Poor tasks remain poor tasks at scale. Weak reward signals can train undesirable behaviour more efficiently.
DeepSeek still has to prove that this machinery produces better models.
But capacity changes what can be attempted.
A small research group may have an excellent reinforcement-learning idea and still be unable to test it across millions of isolated, long-running environments. DeepSeek can.
The same principle made giant GPU clusters strategically important. The hardware never guaranteed a good model. It allowed experiments that smaller competitors could not run.
DSec gives DeepSeek that sort of leverage in agent training.
And agent training is becoming far more important.
The frontier is moving away from models that merely answer questions and towards systems expected to use software, modify code and continue working for long periods without constant human intervention.
Those systems need practice. Practice needs environments. At DeepSeek's scale, the environments themselves become infrastructure worth engineering.
That is why DSec may tell us more about DeepSeek's next generation than another leaderboard would.
A company that can spin up millions of isolated worlds every day can let its models fail millions of times, score those failures and feed the useful experience back into training.
DeepSeek has spent years being described as unusually efficient at building models.
DSec shows it is now trying to become unusually efficient at building the worlds those models learn inside.