Beam

Beam is a developer-first serverless platform purpose-built for AI workloads. The pitch is simple: import a Python function, decorate it, and Beam runs it on a GPU somewhere with the right model weights cached, scaling to thousands of concurrent invocations and back to zero — with cold starts measured in single-digit seconds rather than minutes. It is one of a small cluster of next-generation Modal/RunPod competitors but with a strong emphasis on AI primitives: prebuilt images for vLLM, ComfyUI, and PyTorch; persistent volumes for model weights; task queues for background generation jobs; sandboxes for safely executing untrusted/agent-generated code; and HTTP endpoints for inference APIs. Pricing is fully usage-based with per-second billing on a wide GPU menu — A10G around $0.42/hr up to H100 around $3.93/hr — and a generous $30 free credit on signup. A $25/seat Pro tier adds team features. Beam is popular with AI startups that have outgrown Replicate but do not want to manage Kubernetes, and with agent builders who need a fast sandbox for running model-generated code. Logs, secrets, and deploys all live in a single CLI/SDK.