RunPod
RunPod rents GPUs by the second for AI training and inference. H100 pods start near $1.99/hr, with serverless and Instant Clusters added in 2026.
About RunPod
What is RunPod (2026 update)
RunPod is a cloud GPU rental platform built for AI development, model training, fine-tuning, and inference at scale. You rent GPU compute either as a persistent Pod for interactive work or as a Serverless endpoint that spins workers up and down automatically based on incoming requests, billed per second either way.
What is new in 2026
RunPod shipped a general availability REST API v2 on August 18, 2026, which moved to a new base URL, standardized request and response shapes across the platform, and added catalog endpoints, pod log streaming, and Serverless observability. The platform added a beta ECR integration that lets you pull private AWS container images straight into Pods and Serverless endpoints without migrating to a different registry or juggling separate credentials. Instant Cluster pod scaling arrived in beta too, letting teams boost GPU capacity on an existing cluster without tearing it down and rebuilding, with new pods joining the cluster’s private network automatically. On the inference side, RunPod added Batch Jobs, so you can submit a large set of inference requests to a Serverless endpoint as one managed unit and poll for progress, plus worker affinity for load balancer endpoints so follow-up requests land on the same worker.
Key features
- On-demand GPU Pods across Community Cloud (cheaper, marketplace-style) and Secure Cloud (SOC2, dedicated resources)
- Serverless endpoints that autoscale AI inference workers up and down per request
- Per-second billing on both Pods and Serverless
- Public Endpoints with pre-built models across video, image, text, and audio
- Instant Cluster scaling and beta ECR integration for private container images
- Vercel AI SDK integration and expanding developer tooling
Pricing in 2026
RunPod prices vary by GPU model and whether you pick Community Cloud or Secure Cloud, with Secure Cloud costing more for SOC2 compliance and dedicated resources. Storage runs about $0.10/GB/month for container and running volumes, and roughly $0.07/GB/month for network volumes under 1TB.
| Plan | Price | What you get |
|---|---|---|
| RTX 4090 Pod | $0.34 to $0.74/hour | 24GB GPU, dedicated vCPU and RAM, Community or Secure Cloud |
| A100 80GB Pod | $1.19 to $1.59/hour | 80GB GPU for training and larger inference workloads |
| H100 Pod | $1.99 to $3.49/hour | 80GB GPU, PCIe or SXM variant, higher throughput for large models |
| Serverless (per GPU class) | From about $1.10/hour equivalent (RTX 4090) up to $5.93/hour (H200) | Autoscaling inference workers billed per second of actual usage |
| Enterprise | Custom | Long-term reservation discounts, contact sales |
Who should use it
RunPod fits ML engineers, AI startups, and researchers who need GPU compute on demand without committing to a long-term contract with a hyperscaler. It works well both for interactive Pod-based training and fine-tuning, and for production inference where Serverless autoscaling keeps costs down between bursts of traffic. The Community Cloud tier is aimed at cost-sensitive teams that can tolerate marketplace-style availability, while Secure Cloud suits teams that need SOC2 compliance and dedicated resources. It is a weaker fit for teams that need a single fully managed MLOps platform with built-in experiment tracking and data pipelines, since RunPod is focused on raw compute rather than a full ML lifecycle product.
Limitations and alternatives
Community Cloud pricing and availability can behave more like a marketplace than a guaranteed reservation, so capacity for a specific GPU model is not always instantly available. Exact per-hour prices also shift over time and by region, so check the official pricing page before budgeting a large training run. For comparison, Lambda Labs offers a similar on-demand GPU cloud with a narrower GPU catalog and a reputation for straightforward pricing. Vast.ai runs a more explicit marketplace model where prices float with demand, often undercutting fixed-price providers for spare capacity. Coreweave targets larger enterprise AI workloads with committed-use contracts and a broader managed Kubernetes layer on top of GPU infrastructure.
Frequently asked questions
How is RunPod billed?
RunPod bills per second of actual usage on both GPU Pods and Serverless endpoints, so you only pay for the compute time you actually use.
What is the difference between Community Cloud and Secure Cloud?
Community Cloud offers more GPU variety at lower marketplace-style prices, while Secure Cloud costs more per GPU but adds SOC2 compliance, dedicated resources, and enhanced support.
Does RunPod support Serverless inference?
Yes, RunPod Serverless autoscales inference workers up and down based on incoming requests and bills per second, with 2026 additions including Batch Jobs and worker affinity for load balancer endpoints.
How much does an H100 GPU cost on RunPod?
As of September 2026, H100 pods range from about $1.99/hour on Community Cloud PCIe instances up to around $3.49/hour for Secure Cloud SXM instances, with Serverless H100 usage priced separately.
Information last verified: September 2026.
Related Tools
