Serverless GPU pricing
Serverless GPU = per-second billing that scales to zero between requests: you pay for compute seconds, not idle hours. For bursty inference it beats renting an instance; for sustained load an on-demand instance is usually cheaper. Rates below are $/hr equivalents of per-second prices.
| GPU | $/hr equivalent | $/second | Provider | Verified | |
|---|---|---|---|---|---|
| L4 | $0.80 | $0.000222 | Modal | 2026-07-08 | Try → |
| L40S | $1.95 | $0.000542 | Modal | 2026-07-08 | Try → |
| A100 PCIe 40GB | $2.10 | $0.000583 | Modal | 2026-07-08 | Try → |
| A100 (unspecified) | $2.50 | $0.000694 | Modal | 2026-07-08 | Try → |
| RTX PRO 6000 | $2.99 | $0.000831 | fal | 2026-07-08 | Try → |
| RTX PRO 6000 | $3.03 | $0.000842 | Modal | 2026-07-08 | Try → |
| H100 (unspecified) | $3.95 | $0.001097 | Modal | 2026-07-08 | Try → |
| H100 (unspecified) | $3.99 | $0.001108 | fal | 2026-07-08 | Try → |
| H200 (SXM) | $4.50 | $0.001250 | fal | 2026-07-08 | Try → |
| H200 (SXM) | $4.54 | $0.001261 | Modal | 2026-07-08 | Try → |
| B200 | $6.25 | $0.001736 | Modal | 2026-07-08 | Try → |
| B200 | $6.25 | $0.001736 | fal | 2026-07-08 | Try → |
Serverless bills only active compute seconds but adds cold-start latency and per-request overhead. Sustained workloads above roughly 50-60% utilization are cheaper on-demand: run your numbers in the calculator.
When serverless wins
Pick serverless when
- Traffic is bursty or unpredictable
- You serve a model occasionally (demos, side projects, internal tools)
- You want zero idle cost and no capacity management
Pick an instance when
- Utilization is sustained (training, batch jobs, steady APIs)
- Cold starts are unacceptable
- You need full control of the environment or multi-GPU nodes
Modal includes a recurring monthly free credit tier; see free credits. RunPod also offers serverless endpoints priced separately from the pod rates shown on the RunPod page.