AIBRIEF
A cloud infrastructure vendor shows off a rebuilt control plane for running millions of short lived, isolated compute environments (the kind AI agents and training jobs consume) at once.
Modal, a cloud-infrastructure vendor for AI agents, says it "created all 1 million in under a minute." The line comes from a company blog post describing a self-run demonstration on Modal's own platform, not an independent benchmark. The 1,000,000 number is the company's claim, and Modal rebuilt its sandbox control plane to make it possible.
A sandbox, in Modal's vocabulary, is a short-lived, isolated compute environment used to run an AI agent or a training job. The pre-rebuild ceiling was roughly 50,000 concurrent sandboxes per customer. The new architecture pushes container creation directly onto load balancers in the worker fleet, removing the centralized scheduler and etcd writes that the post names as Kubernetes' bottlenecks at agent scale.
The demand-side signal is more useful than the demo: a separate Amplify Partners writeup says an unnamed major AI lab is already running about 100,000 concurrent sandboxes for reinforcement-learning workloads, with a stated goal of 1,000,000. The lab is identified only through a VC observer, not on the record. Modal is positioning itself against E2B; a Hacker News commenter noted Firecracker microVMs and SQLite on bare metal can also scale fast, suggesting the design space is still open.