Modal

Deploying a GPU workload on Modal means writing ordinary Python: decorate a function, declare its container image and GPU type in code, and the platform builds, ships, and autoscales it from zero to over 1,000 GPUs with sub-second container cold starts. The service targets AI workloads specifically, offering H100, A100, B200, and A10G GPUs for LLM inference, image and video generation, fine-tuning, reinforcement learning, batch processing, and sandboxed code execution for agents. Billing is per-second with no reserved capacity or commitments, and the free tier includes 30 dollars of compute credits per month. The client SDK is open source on GitHub under modal-labs, but the platform itself is proprietary, so nothing runs without Modal's cloud. Teams commonly use it to serve vLLM and diffusion models without managing Kubernetes or bare GPU instances, trading self-hosting control for developer speed. Documentation, examples, and pricing live at modal.com, and the workflow stays entirely in Python from local iteration through production deployment.