DeepSeek just published a massive paper on their open-source Elastic Compute (DSE) architecture, and the scale is insane. They are running a Recursive Self-Improvement (RSI) loop where agents build the training grounds for stronger agents.
Their cluster handles 3 million sandboxes a day, spinning up 5,000 virtual environments per second using their own 3FS storage. One rogue agent even cheated so hard it crashed the file system.