Modal

Modal is a serverless cloud computing platform for artificial intelligence workloads, founded in 2021 by Erik Bernhardsson, a former Spotify principal engineer and creator of the Luigi workflow engine. Modal lets developers run GPU inference, fine-tuning, batch processing and sandboxed agent code by decorating ordinary Python functions, without provisioning or managing servers.^[1]

History

Modal launched before the generative AI boom and found its first real traction with the release of Stable Diffusion in 2022, when demand for on-demand GPU inference outstripped the capacity most startups could buy. The company raised a $25 million Series A in April 2024, bringing total funding at the time to about $32 million, with more than 100 enterprise customers.^[2]

The platform then rode the AI infrastructure wave. In October 2026 Modal announced a $750 million Series D led by Accel at a $15.75 billion post-money valuation — more than tripling its valuation in five months — after crossing $1 billion in annualised run rate. Total disclosed funding reached roughly $1.22 billion.^[3][4]

Key Concepts and Technology

Modal's model rests on three ideas. First, serverless GPUs: compute is created on demand and billed by the second, so a job that needs eight H100s for ninety seconds costs ninety seconds of eight H100s.^[1] Second, Python-native deployment: functions are wrapped in decorators such as @app.function(), which specify GPU type, memory, timeout, secrets and dependencies; Modal builds and caches a container image automatically. Third, fast cold starts: containers spin up in seconds rather than minutes, making it practical to autoscale from zero traffic to hundreds of replicas.

Beyond training and inference, Modal provides Modal Sandboxes — short-lived, network-isolated containers used to run untrusted or AI-generated code, a common requirement for coding agents — plus persistent volumes for checkpoints, distributed job fan-out and scheduled workloads. It competes with Replicate, Together AI, Fireworks AI, Lambda and the hyperscaler GPU services.

Applications and Impact

Typical uses include serving image and video generation models, running batch document-processing pipelines, fine-tuning open-weight models, executing reinforcement-learning rollouts and powering agent sandboxes. For startups, the impact is economic: it removes the need to buy or multi-year-commit to GPUs, converting a capital expense into a per-second operating cost, and it lets small teams scale an inference service without a platform-engineering function.

>See Also

• MLOps
• Model Serving
• GPU Cluster
• Replicate
• AWS Malaysia Region
• National AI Compute Exchange