Runpod review

Runpod rents cloud GPUs for AI development and production. Its main options are GPU Pods for interactive work, training, and persistent compute; Serverless endpoints for inference; and Clusters for workloads spread across multiple GPUs. The homepage says Runpod offers more than 30 GPU SKUs across 31 regions, with Serverless endpoints that autoscale and scale to zero when idle.

This is infrastructure, not a managed model-training product: you choose compute and run your own workloads. That can suit developers who need a particular GPU or want to bring their own setup, but it also means accounting for storage and time spent configuring jobs. The homepage lists FlashBoot with sub-200ms cold starts and says network storage has no egress fees. Those claims describe the platform; they don’t guarantee that every GPU or region will be available when you need it.

The appeal is flexible GPU access without a monthly subscription. According to the pricing page, Pods have no commitment and are billed per GPU hour. Secure Cloud examples include an RTX A5000 at $0.27/hour, an A100 at $1.79/hour, and an H100 PCIe at $2.89/hour. Community Cloud rates differ: an H100 PCIe starts at $1.99/hour, while an H200 starts at $3.59/hour.

Serverless pricing is per worker hour, with listed rates from $0.58/hour for 16GB GPUs to $11.99/hour for a B300. The page lists an H100 at $5.39/hour and an RTX 4090 at $1.34/hour. For distributed jobs, Clusters are billed per GPU hour and can scale to 64 GPUs; listed examples include an A100 SXM at $2.15/hour and an H200 SXM at $6.35/hour.

• Pods fit jobs that need a dedicated instance; Serverless suits API inference that can scale to zero.
• Network storage is listed at $0.07/GB/month under 1TB or $0.05/GB/month over 1TB for Standard, and $0.14/GB/month for High-Performance.
• Container Disk costs $0.10/GB/month; Volume Disk costs $0.10/GB/month while running or $0.20/GB/month idle.

Runpod’s hourly rates are only part of the bill. If a job leaves files on a running Pod, you may keep paying GPU charges while retrieving them. In one Reddit thread, a user reported download speeds of 50–60kbps for a trained LoRA; commenters suggested moving files to network storage, which adds a separate storage charge. That is an anecdote, not evidence of a general download-speed problem, but it shows how a cheap GPU hour can turn into a longer bill.

Reliability complaints also appear in user discussions. A ComfyUI thread includes reports of Pods timing out, missing template images, lost network-storage data, and extended loss of access. These are user reports, not a measured failure rate. One commenter also complained about support and the lack of SLAs; the published pricing information does not publish an SLA.

There’s no listed free plan, and a free-trial offer is not published on the vendor's site. Some Cluster GPU options—L40S, H100 SXM, and B200—require contacting sales, as do Reserved Clusters. Reserved pricing is custom, with terms listed from 1 to 12+ months. Community and Secure Cloud rates differ, so check the actual GPU and location before estimating a workload.

Runpod charges for usage, not a monthly subscription, and Pods have no commitment. The pricing page lists Secure Cloud Pods from $0.27/hour for an RTX A5000 and $0.50/hour for an RTX 3090, rising to $1.79/hour for an A100, $2.89/hour for an H100 PCIe, and $8.99/hour for a B300. Community Cloud rates differ; an H100 PCIe starts at $1.99/hour.

Serverless ranges from $0.58 to $11.99 per worker hour. Listed examples include $5.39/hour for an H100 and $1.34/hour for an RTX 4090. Clusters list an A100 SXM at $2.15/hour and H200 SXM at $6.35/hour; some GPU prices require a sales quote. Reserved Clusters are custom-priced.

Budget for storage too: Network Storage Standard costs $0.07/GB/month under 1TB or $0.05/GB/month above it; High-Performance costs $0.14/GB/month. Volume Disk is $0.10/GB/month while running and $0.20/GB/month idle. These hourly rates can make costs hard to predict if workloads run longer than planned.

The migration burden depends on what you run. A workload using standard GPU instances should be simpler to move than one built around Runpod Serverless endpoints, Clusters, or its network storage. The published materials don’t document a one-click migration path, export workflow, or compatibility guarantees, so check how your model files, containers, endpoint configuration, and stored data will move before committing.

Runpod says its Network Storage has no egress fees, but that doesn’t establish how quickly or easily you can retrieve artifacts. The export process and any other transfer charges are not published on the vendor's site. Plan to copy important files out and verify them before shutting down resources.

Reddit users report reliability concerns; the review material specifically flags artifact downloads and support as weak points. Those are practical reasons to leave, especially if dependable downloads matter to production. For a move in, price a representative workload first: GPU hourly rates, storage, and idle Volume Disk charges can all affect the bill.