No minimum commitment · Per-second billing
What is RunPod?
RunPod is a GPU cloud marketplace founded in 2022, now one of the most popular platforms for AI developers. It operates on two tiers:
- Community Cloud — third-party hosted GPUs. Cheapest prices, no uptime guarantee. From $0.16/h. Best for fault-tolerant batch jobs.
- Secure Cloud — dedicated datacenter hardware. Reliable, slightly pricier. From $0.27/h. Best for inference and long runs.
RunPod also offers Serverless inference endpoints — your model scales to zero when idle, you pay per compute unit. This makes it genuinely competitive with Modal and Replicate for intermittent inference workloads. We rank it #1 in our serverless GPU comparison.
RunPod Pricing (July 2026)
| GPU | VRAM | Community Cloud | Secure Cloud | Best For |
|---|---|---|---|---|
| RTX A5000 | 24 GB | $0.16/h | $0.27/h | Fine-tuning 7B, SD XL |
| RTX 3090 | 24 GB | $0.22/h | $0.46/h | Fine-tuning 7B, SD |
| RTX 4090 | 24 GB | $0.34/h | $0.69/h | Fast inference, SD XL |
| RTX A6000 | 48 GB | $0.33/h | $0.49/h | Fine-tuning 7B, SD XL |
| L40S | 48 GB | $0.79/h | $0.99/h | Fast inference, SD XL |
| A100 80GB | 80 GB | $1.19/h | $1.39/h | Large training runs |
| H100 PCIe | 80 GB | $1.99/h | $2.89/h | Fastest inference |
| H100 SXM | 80 GB | $2.69/h | $2.99/h | Fastest inference |
| H200 | 141 GB | $3.59/h | $4.39/h | 70B models, training |
Both tiers read off runpod.io/pricing on 20 July 2026. Community Cloud rates depend on which hosts have capacity, so they move more than Secure Cloud.
Where RunPod Bills Surprise People: Storage
GPU charges stop the second you stop a pod. Storage does the opposite. A volume disk runs at $0.10/GB/month while the pod is active and $0.20/GB/month once it goes idle, so stopping a pod to save money doubles the rate on everything attached to it.
A 200 GB volume holding model weights and a dataset costs $20/month while you are working and $40/month after you stop. Leave it attached to a finished project for a year and that is $480 for storage nothing is reading. At Community RTX A5000 rates, $40 would have bought 250 GPU-hours.
Network storage avoids the penalty: $0.07/GB/month under 1 TB, $0.05/GB/month above it, and the rate does not change when nothing is running. The same 200 GB costs $14/month there, 2.9x cheaper than an idle volume disk. Keep anything that has to survive between sessions on network storage and let pod volumes stay disposable.
RunPod Pros & Cons
- Cheapest community GPUs from $0.16/h
- Massive GPU variety including H100
- Serverless endpoints for inference APIs
- Great UI and pod management
- Community cloud less reliable than dedicated
- Storage costs add up over time
- Support can be slow on free tier
Community Cloud vs Secure Cloud — Which Should You Choose?
Use Community Cloud when you have fault-tolerant batch jobs (training with checkpointing, Stable Diffusion renders, data processing). The price savings are massive and downtime is acceptable.
Use Secure Cloud when downtime costs more than the price delta: production inference APIs, multi-day fine-tuning runs, anything with SLA requirements. The hardware is dedicated and significantly more reliable.
RunPod Serverless — Is It Worth It?
RunPod Serverless is genuinely excellent for intermittent inference. You deploy a Docker image once, RunPod handles scaling. Cold starts are the main concern (~5-15s for large models), but warm workers can be pinned at additional cost. Compared to Modal and Replicate, RunPod Serverless is typically 30-50% cheaper at scale.
RunPod Alternatives
- Vast.ai — Even cheaper community cloud. Less UI polish, more raw savings.
- Lambda Labs — Better reliability, great for on-demand H100 access without the marketplace complexity.
- Paperspace — Better integrated notebook environment for research teams.
- CoreWeave — Enterprise Kubernetes-native clusters for foundation model training.
Verdict
RunPod is our #1 pick for most AI developers. The combination of price, GPU variety, and ease of use is unmatched. Secure Cloud is reliable enough for most production workloads. Serverless is genuinely competitive for inference APIs. The main weakness is support responsiveness on free/hobbyist plans. If you need bulletproof enterprise SLAs, look at Lambda Labs or CoreWeave.
RunPod FAQ
Is RunPod reliable?
RunPod's Secure Cloud uses dedicated datacenter hardware with strong uptime. Community Cloud depends on individual hosts and can be interrupted. For production inference or long training runs, use Secure Cloud.
Does RunPod have H100s?
Yes, RunPod offers H100 SXM and PCIe instances in both tiers. Availability fluctuates — H100s are the first to sell out. Check the real-time availability map on RunPod.io.
How does RunPod billing work?
RunPod bills per-second with a 1-minute minimum, and GPU charges stop when you stop a pod. Storage does not. A volume disk costs $0.10/GB/month while the pod runs and $0.20/GB/month once it is idle, so stopping a pod doubles the storage rate rather than ending it. Network storage is cheaper at $0.07/GB/month under 1 TB and $0.05/GB/month above it.
What is RunPod Serverless?
RunPod Serverless deploys inference endpoints that scale to zero. You pay only per request (compute unit), not per idle hour. Ideal for low-traffic inference APIs.
Can I use custom Docker images?
Yes — any public Docker image works. RunPod also has 50+ official templates (vLLM, Ollama, ComfyUI, Kohya, etc.) that launch pre-configured in seconds.