RunPod Review

RunPod is the AI-hosting platform we keep recommending to founders training their first models — and the GPU pricing platform we use to populate our calculator. Per-second billing across the NVIDIA catalogue means a half-finished fine-tune doesn't bankrupt you, and the on-demand → reserved gradient gives teams enough rope to right-size as workloads stabilise.

What we tested

We ran a representative inference + fine-tune workload against three GPU classes (A100 80GB, A40, and 3090) across two pods. We tracked spin-up latency, network throughput to S3-compatible storage, sustained GPU utilisation during a 30-minute fine-tune, and per-second billing accuracy at teardown. Performance landed at 8.5 / 10 — comfortably better than vast.ai for the same A100 SKU, slightly behind Lambda Cloud on raw throughput but ahead on flexibility.

What we liked

The per-second billing model is the headline. Combined with the wide catalogue, teams can swap a 3090 for an A100 mid-experiment without rebuilding. Documentation has improved meaningfully in the past quarter and the Discord community responds quickly.

What needs work

Cold-start times vary by region — EU pods spin up faster than US-east in our tests. Reserved-instance documentation could use a worked example for first-time buyers. Reliability scored 7.5 / 10, dragged by one network-storage hiccup during testing that resolved within minutes.