Modal

Modal runs Python code on cloud GPUs without you managing any servers: you add a decorator to a function, and it executes remotely with the hardware and dependencies you specified.

Pricing model: you pay per second of actual compute, billed only while your function runs — containers scale to zero when idle, so an endpoint that receives no traffic costs nothing. This is the key difference from renting a GPU instance by the hour, and it is what makes it practical for bursty workloads.

Best for model inference endpoints, batch jobs, fine-tuning runs and scheduled data pipelines. Not for non-developers — this is an infrastructure tool and assumes you write Python. Serious GPU work still costs real money, so set spending limits before your first long training run.