Modal

Modal is a serverless GPU infrastructure platform for running AI workloads—inference, training, batch processing, and sandboxes—without managing servers or planning capacity. You define your entire cloud environment in Python using composable primitives that specify logic and hardware in one place. It's a 'cloud environment in code' approach that feels local but scales globally, with sub-second cold starts and instant autoscaling from 0 to 1000+ GPUs. Built for developers who want speed, whether deploying LLM APIs, fine-tuning open-source models, or building autonomous agents with isolated execution. For inference, Modal supports any model or engine on a variety of Nvidia GPUs including B300, B200, H200, H100, A100, and more. Online inference offers sub-10ms overhead latency globally, with built-in token streaming, WebRTC, and WebSocket. Multi-modal inference covers image, video, audio, and embeddings, plus batch processing for evals and dataset generation across thousands of GPUs in parallel—no job orchestration needed. Training spans single-GPU fine-tuning to multi-node runs on up to 128 B200s with Infiniband, plus hyperparameter sweeps and reinforcement learning with thousands of concurrent rollouts. Sandboxes provide programmatic, secure ephemeral environments for untrusted code—ideal for coding agents, background agents, and RL rollouts, with on-demand GPU acceleration. Pricing is per-second, so you pay only for compute used, with no idle costs—a key advantage for spiky workloads. The Starter tier includes $30/month free compute; Team tier offers $100/month free credits, with volume discounts and credit grants for startups and academics. Modal competes with traditional cloud AI services like AWS SageMaker and other GPU clouds, but its serverless, Python-first model removes capacity planning. It's a strong fit for teams that value speed and elasticity over reserved capacity. However, as of 2026, a security incident involving an AI agent hacking a Modal