Modal

Modal has carved out a strong reputation among AI developers by making serverless GPU compute feel like ordinary Python rather than a DevOps project. Defining hardware requirements and application logic directly in a Python SDK, then deploying with sub-second cold starts and autoscaling from zero to over a thousand GPUs, removes a huge amount of the infrastructure work that used to be table stakes for running inference or training at scale. It isn't a fine-tuning framework itself, but its tight compatibility with tools like Axolotl and Unsloth makes it a popular backend for LoRA and full fine-tuning jobs, alongside solid support for LLM, audio, image, and video inference and sandboxed code execution. A GPU fleet spanning T4 through B200 covers most workloads, and per-second billing (H100 at roughly $3.95/hour) is competitive and transparent compared to reserved-instance cloud pricing. The free tier's $30/month in credits and generous concurrent GPU limit make it easy to prototype before committing. The tradeoff is that Modal is a developer-first, code-centric platform with no low-code UI, so it suits engineering teams far more than non-technical users. A genuinely well-built option for teams that want GPU infrastructure without managing it themselves.