Three ways to run a model in the cloud without managing GPUs, and the billing model is the whole decision.
| Billing | Idle cost | Free credits | |
|---|---|---|---|
| Replicate | per second of prediction time | none — $0 when idle | pay-as-you-go |
| Modal | per second, scale-to-zero | none when scaled to zero | $30/mo compute |
| Hugging Face Inference Endpoints | per hour, per replica | full hourly rate 24/7 with min 1 replica | Pro $9/mo (not compute) |
Replicate is the simplest: push a model, get an API endpoint, pay only for the seconds it spends running. GPU rates run from ~$0.000225/sec (T4) to ~$0.003525/sec (A100 80GB). For bursty or low-volume inference — a feature that fires a few thousand times a day — this is the cheapest and least fiddly option. No platform fee.
Modal is Replicate for people who want to write the infrastructure code themselves: define your container and GPU in Python, and it handles scaling including down to zero. Per-second GPU pricing is similar (T4 ~$0.000164/sec, H100 ~$0.001097/sec), every account gets $30/month in free credits, and there are startup ($25k) and academic ($10k) grant programs. The Team plan is $250/month. Best when you need custom logic around the model — batching, multi-step pipelines, your own preprocessing.
Hugging Face Inference Endpoints bills by the hour, and an endpoint with a one-replica minimum runs the full hourly rate around the clock whether it serves one request or a million. Rates go from $0.50/hr to $10/hr for a single H100. This only makes economic sense at sustained high utilization — if your endpoint is busy most of the time, hourly beats per-second; if it's spiky, you're paying for idle GPUs.
Bursty or low-volume: Replicate. Need custom pipeline logic: Modal. Sustained high-traffic inference: Hugging Face.