Modal

Modal is a Python native serverless cloud built for heavy AI workloads, covering inference, training, batch processing, and sandboxes with sub second cold starts, instant autoscaling, and a developer experience that feels local. Instead of Docker files or Kubernetes manifests, developers wrap functions with decorators that specify GPU type, container image, memory, and concurrency, and Modal handles the container build, scheduling, autoscaling, and per second billing. The company was founded by engineers with a strong open source pedigree, serves customers including Suno, Substack, Ramp, Notion, and Lovable along with roughly a thousand paying teams, and raised an eighty seven million dollar Series B in September 2025 at a valuation around one and one tenth billion dollars.

For agent builders, Modal Sandboxes are the execution layer: isolated, ephemeral environments for running untrusted code, spun up programmatically with custom images and any dependency, scaling to hundreds of thousands of concurrent runs. The same stack natively hosts training infrastructure, so interactive coding agents and long running reinforcement learning rollouts share one platform. Composable primitives cover persistent volumes, key value dictionaries, task queues, web endpoints, and scheduled functions, with integrated logging and full visibility into every function, sandbox, and container.

Modal routes workloads across clouds and regions in real time, offering nine GPU types from T4 to B200 on demand with no commitments or capacity planning, and scales back to zero when idle. Pricing is pure per second usage across GPU, CPU, and memory, the finest billing granularity among major compute products, with a free Starter tier that includes thirty dollars a month in credits, a Team plan at two hundred fifty dollars a month, and a sales led Enterprise tier with VPC deployment and marketplace billing. Regional and non preemption multipliers apply, and sandbox workloads run non preemptible at a premium, so agents that keep sandboxes alive should model those costs carefully.