RunPod GPU Pricing 2026: Per-Hour Cost Guide

RunPod GPU pricing 2026 per-hour: H100 PCIe Community $1.99 / Secure $2.89; H100 SXM Community $2.69 / Secure $3.29; B200 Community $5.98 / Secure $6.79; A100 SXM Community $1.39 / Secure $1.59. Per-second billing, self-serve setup in minutes. Community is host-managed; Secure is platform-managed.

Last verified: Sep 16, 2026.

At a glance

• H100 PCIe Community: $1.99/GPU-hr | Secure: $2.89/GPU-hr
• H100 SXM Community: $2.69/GPU-hr | Secure: $3.29/GPU-hr
• H100 NVL Community: $2.59/GPU-hr | Secure: $3.19/GPU-hr
• H200 Community: $3.59/GPU-hr | Secure: $4.59/GPU-hr
• B200 Community: $5.98/GPU-hr | Secure: $6.79/GPU-hr
• B300 Community: $6.94/GPU-hr | Secure: $7.89/GPU-hr
• A100 SXM 80GB Community: $1.39/GPU-hr | Secure: $1.59/GPU-hr
• L40S Community: $0.79/GPU-hr | Secure: $0.99/GPU-hr
• Cheapest: RTX A6000 Community $0.33/GPU-hr
• Billing: Per-second; 1-min minimum on Pods

Why RunPod is the practical sweet spot for indie developers and small teams

RunPod delivers per-second billing on dedicated GPU pods at the lowest published neocloud rates — with two distinct reliability tiers.

RunPod's pricing philosophy is 'GPU in a few minutes with an API that does not need a solutions architect.' Per-second billing ties cost to active compute; the API is self-serve; no quota approval is needed for 1-GPU jobs. RunPod bills in 1-second increments with a 1-minute minimum on Pods, which makes it well-suited for prototyping, fine-tuning runs, and bursty inference.

The Community vs Secure tier split is RunPod's key differentiator. Community Cloud hosts are third-party operators — individuals and small data centers renting spare GPU capacity on the RunPod marketplace. Community rates are 10-20% cheaper but the host can pull the machine offline for any reason. Secure Cloud is RunPod's own tier-3 datacenter infrastructure with platform-managed reliability — the right choice for production workloads that cannot tolerate random eviction (runpod.io/pricing, July 2026).

H100 pricing on RunPod

RunPod's H100 rates sit between Vast.ai marketplace ($1.49/GPU-hr interruptible) and Lambda on-demand ($3.29/GPU-hr 1x PCIe).

Verified July 2026 rates (runpod.io/pricing):

• H100 PCIe Community: $1.99/GPU-hr — third-party host, evictable
• H100 PCIe Secure: $2.89/GPU-hr — RunPod datacenter, platform-managed
• H100 SXM Community: $2.69/GPU-hr — NVLink-capable SXM form factor
• H100 SXM Secure: $3.29/GPU-hr — same silicon, RunPod datacenter
• H100 NVL Community: $2.59/GPU-hr — newer dual-card form factor
• H100 NVL Secure: $3.19/GPU-hr — same NVL on Secure Cloud

For a single-GPU fine-tuning job, RunPod Community H100 PCIe at $1.99/GPU-hr is the cheapest dedicated-pod rate in the category — below Lambda's $3.29/GPU-hr 1x PCIe and below CoreWeave's $6.16/GPU-hr 8x HGX (which requires buying 8 GPUs even if you only need 1). For workloads that benefit from NVLink gradient synchronization, H100 SXM Community at $2.69/GPU-hr is the right starting point.

B200 and B300 Blackwell pricing

RunPod's Blackwell rates are the cheapest published on-demand rates in the neocloud market in 2026.

Verified July 2026 (runpod.io/pricing):

• B200 Community: $5.98/GPU-hr
• B200 Secure: $6.79/GPU-hr
• B300 Community: $6.94/GPU-hr
• B300 Secure: $7.89/GPU-hr

Compared with Lambda's $6.69/GPU-hr 8x B200 SXM and CoreWeave's $8.60/GPU-hr HGX B200 8x on-demand, RunPod is $0.71-$2.60/GPU-hr cheaper. The price gap reflects the absence of forced 8-GPU minimums on RunPod — you can rent a single B200 pod and only pay for what you use. For Blackwell-capacity-constrained workloads in 2026, RunPod is the practical path to single-GPU Blackwell access without negotiating a 256-GPU reserved contract.

A100, L40S, and older generation pricing

For workloads that don't need the newest chip, RunPod's older-generation rates are the cheapest production-grade rates in the category.

Verified July 2026 (runpod.io/pricing):

• A100 PCIe 40GB Community: $1.19/GPU-hr | Secure: $1.39/GPU-hr
• A100 SXM 80GB Community: $1.39/GPU-hr | Secure: $1.59/GPU-hr
• L40S Community: $0.79/GPU-hr | Secure: $0.99/GPU-hr
• RTX A6000 Community: $0.33/GPU-hr (cheapest published)

A100 SXM 80GB at $1.39/GPU-hr Community is the cheapest published A100 rate for production training workloads. The L40S at $0.79/GPU-hr is well-suited to inference and 3D rendering workloads. The RTX A6000 at $0.33/GPU-hr is the floor for hobbyist fine-tuning (Markaicode, August 2026).

RunPod Serverless pricing

For spiky inference workloads, RunPod Serverless offers auto-scaling workers billed per-second with no idle cost.

Verified August 2026 (markaicode.com):

• RunPod Serverless A100 80GB Flex: ~$2.72/hr equivalent per-second
• RunPod Serverless H100 worker tier: $4.55/hr equivalent
• RunPod Active workers (always-on): Discounted rate per GPU type, displayed at deploy time

Serverless workers are auto-scaling — they spin up on demand and shut down when traffic stops. Flex workers are billed per-second across the full container lifecycle (including cold starts); Active workers are always-on but discounted. The choice between Flex and Active depends on traffic patterns: Flex wins for spiky traffic, Active wins for sustained traffic above 60% utilization.

The break-even math: a RunPod Secure Cloud A100 SXM Pod at $1.59/hr continuously provisioned costs $1.59/hr regardless of utilization. A Serverless Flex A100 80GB at $2.72/hr equivalent only bills while serving requests. At 50% utilization, Serverless costs $1.36/hr — 14% cheaper than the dedicated Pod. Below 50% utilization, Serverless dominates; above 60%, dedicated Pod wins.

RunPod Pod vs Serverless comparison

How to choose the right RunPod tier in 2026

The decision hinges on reliability tolerance, workload duration, and traffic patterns.

1. Short fine-tuning runs (under 1 week): Community Cloud H100 PCIe at $1.99/GPU-hr. Accept host-level reliability; checkpoint every 30 minutes.
2. Multi-week training you can't afford to lose: Secure Cloud H100 SXM at $3.29/GPU-hr. Platform-managed reliability; same per-second billing model.
3. Production inference with compliance requirements: Secure Cloud Pod with auto-scaling. Use Serverless Flex for spiky traffic below 60% utilization; switch to dedicated Pods above 60% utilization.
4. Blackwell frontier workloads: Secure Cloud B200 at $6.79/GPU-hr or Community Cloud B300 at $6.94/GPU-hr. Avoid 8-GPU minimums by going to RunPod before Lambda or CoreWeave.
5. Budget-constrained indie developers: Community Cloud A100 PCIe 40GB at $1.19/GPU-hr or L40S at $0.79/GPU-hr. Production-grade hardware at commodity pricing.

What enterprise buyers should do next

Three actions for organizations evaluating RunPod in 2026.

1. Run a 2-week pilot on both Community and Secure tiers. Same workload, same hyperparameters. Compare cost, throughput, and reliability variance. The price gap is real but so is the reliability gap — quantify both before committing to a multi-month reserved contract.
2. Model storage and registry costs separately. RunPod bills attached storage per GB-month and charges for private Docker image storage beyond the free tier. For data-intensive workloads (10+ TB datasets), storage can add 5-10% to the headline GPU rate.
3. Negotiate volume discounts if committing to 1,000+ GPU-hours/month. RunPod sales engages on volume — typically 15-25% off the published rate card for committed monthly spend. Reserved contracts at scale can match hyperscaler 3-year reservation discounts.

What to watch next

Three near-term datapoints. First, RunPod's B300-class pricing — current rates are published but availability is intermittent; expect rate-card adjustments as supply normalizes by Q4 2026. Second, RunPod Serverless FlashBoot reliability — the marketed sub-200ms cold start diverges sharply from real-world 7-120s boots for large model containers. If RunPod addresses cold-start variance, Serverless Flex becomes a stronger competitor to Modal ($3.95/hr H100) and Replicate ($5.49/hr H100). Third, RunPod's reserved-capacity offerings — RunPod has historically been pay-as-you-go only; if they introduce a 1-Click Cluster equivalent, RunPod pricing enters the Lambda 1-Click and CoreWeave reserved tier at the lower end of the range (runpod.io/pricing, July 2026).