Beam

Beam is a serverless GPU cloud built for AI workloads. You write a plain Python function, decorate it with @endpoint(gpu="A10G"), and Beam deploys it as an autoscaling inference endpoint — no Dockerfiles, no Kubernetes, no cluster babysitting. Its custom container runtime and scheduler claim sub-second cold starts, and workloads scale to zero when idle so you pay only for what runs.

Beyond endpoints, Beam offers sandboxes (isolated containers for running LLM-generated code — handy for AI agents), scheduled jobs, webhooks, fan-out across hundreds of containers, and persistent volumes for model weights. GPU options span 4090s to H100s, and you can also bring your own hardware. The underlying runtime, beta9, is open source (Apache-2.0).

The honest caveat: Beam is a smaller player next to Modal, RunPod, and the hyperscalers — capacity on the hottest GPUs can be constrained, and serverless GPU pricing rewards spiky workloads more than 24/7 ones. Great for inference endpoints and batch jobs with uneven traffic; compare per-hour costs before committing a constant load.

1. Create an account at beam.cloud and install the client: pip install beam-client.
2. Authenticate with beam configure (or set your API token).
3. Write a function decorated with @endpoint(...) or @function(...) specifying GPU/CPU/memory.
4. Deploy with beam deploy app.py:handler — you get an HTTPS endpoint URL back.