Modal

Modal's pitch is that backend infrastructure should feel like writing local code. You annotate a Python function to say it needs a GPU, a specific container image, or access to certain storage, and Modal handles provisioning and execution without you ever writing a Dockerfile or managing a cluster. Scaling happens automatically — from zero up to hundreds of parallel workers — and you only pay for actual compute time, not idle time.

For AI builders, Modal is popular for inference deployment, fine-tuning runs, eval pipelines, and running sandboxed code (isolated execution environments) for agent tools. The DX (developer experience) is frequently compared to what Vercel did for frontend deployment: it removed the friction of deploying backend services without sacrificing control for teams that need it.

Modal sits at a different level than GPU clouds. A GPU cloud gives you a server you manage. Modal gives you a serverless function runtime where the server is fully abstracted away. Teams often use Modal for workloads with bursty or unpredictable demand — a spike in eval jobs during a training run, or occasional fine-tune requests that do not justify keeping a server warm 24 hours a day.