Modal is cloud infrastructure for teams that build, train and serve AI applications. Developers write Python functions and Modal runs them on its own container runtime, autoscaling up with traffic and down to zero, so there is no charge for idle machines. It serves language, image, video and audio models for inference, with OpenAI-compatible endpoints, and its training side supports supervised fine-tuning, LoRA, full fine-tunes and reinforcement learning on GPUs with fast start times.
Other products include Sandboxes, which are isolated environments for coding agents, reinforcement learning rollouts and untrusted code, with the company claiming it can create millions of concurrent sandboxes in under a minute. Multi-node GPU clusters with RDMA are listed as now available to all customers, and there are Notebooks, a Batch product and web tools such as a GPU glossary and a serverless GPU calculator. The site lists DoorDash, Cognition and Ramp as customers, runs across more than 20 clouds, and is SOC 2 Type II compliant and HIPAA ready.
The Starter plan has no monthly fee and includes $30 of compute a month. On the US pricing page GPU time is billed per second, for example $0.001097 a second for an H100 SXM5, $0.000306 for an A10 and $0.000222 for an L4, with CPU and memory billed separately. Usage beyond the included compute is charged, and the page says Shared Endpoint tokens are billed from the first request, a change that took effect on 1 September 2026. Credits for startups and up to $10,000 for academic researchers are available by application.
