Beam

Beam is an open source serverless cloud for AI workloads, centered on secure sandboxes and GPU inference that spin up in milliseconds and bill only for active runtime. Its engine, beta9, is open source under the AGPL license, so the same runtime can run fully self hosted or bring your own compute on AWS, Google Cloud, Azure, or Hetzner using credits or hardware a team already has, then burst to Beam's managed cloud when a workload needs more capacity. That combination makes the managed price a ceiling rather than a floor and keeps sensitive code and data inside a team's own infrastructure when required.

Beam is built for production agents that need persistent state and compute intensity together. Sandboxes use gVisor isolation, support GPUs directly inside the sandbox, and offer stateful snapshots that restore a running environment into thousands of concurrent isolated runs, with no hard session cap that would cut off a long horizon task. The platform does not bill for cold start or image pull time, which matters for the high churn, many spin up pattern that agents create. Composable Python primitives cover autoscaling endpoints, task queues with retries and event based callbacks, distributed storage, and memory snapshotting for fast container boots. Adjacent open source products include Airstore, a virtual filesystem that turns sources like Gmail, GitHub, and Linear into files an agent can read.

Modal and Beam compete closely, but Beam differentiates on stateful, GPU capable sandboxes with bring your own compute. GPU rates are aggressive, with an H100 around one to three and a half dollars an hour depending on configuration and an A100 well below the managed market, billed per second or millisecond with scale to zero. Pricing is a free Developer tier with thirty dollars a month in credits, a Team plan at eighty nine dollars a month for more seats and concurrency, and the open source core free to self host, with storage included at no charge.