RunPod vs Lambda Cloud vs Vast.ai: Real H100 GPU Rental Cost per 10,000 GPU Hours in 2026
Prices checked: October 2, 2026.
Short answer
For 10,000 H100 GPU hours (for example, one 8-GPU node running about 52 days):
- RunPod Secure Cloud H100 SXM: $34,900. H100 PCIe is $28,900 and H100 NVL $31,900, all at listed on-demand rates.
- Lambda Cloud on-demand H100 SXM: $39,900. That's $5,000 more than RunPod SXM, but Lambda has no egress fees and includes large local NVMe.
- Vast.ai: roughly $8,900 to $61,800, depending on which host you rent. Interruptible listings have been seen as low as $0.89 per GPU-hour. Verified on-demand H100 SXM hosts are reported at about $2.70–$6.18 per GPU-hour, with a median around $4.44.
Vast.ai can be the cheapest by far, or more expensive than both, because it's a marketplace and not a price list. RunPod and Lambda give you a predictable number.
Native billing units
All three bill per GPU-hour, per second or minute of runtime. They differ in what surrounds the GPU:
| Provider | GPU billing | Storage | Egress | Label |
|---|---|---|---|---|
| RunPod Secure Cloud | Per GPU-hour, on-demand | Container disk $0.10/GB-mo; volume disk $0.10/GB-mo running, $0.20/GB-mo idle; network storage $0.07/GB-mo (<1 TB), $0.05/GB-mo (>1 TB) | Not listed on pricing page | OFFICIAL |
| Lambda Cloud | Per GPU-hour, on-demand | Local SSD included in instance; separate filesystem pricing not shown on pricing page | "No egress fees" | OFFICIAL |
| Vast.ai | Per GPU-hour, set by each host; per-second billing | Set by each host | Set by each host | OFFICIAL (model); prices REPORTED |
H100 price per GPU-hour
| Offer | $/GPU-hour | Label |
|---|---|---|
| RunPod Secure Cloud H100 SXM 80 GB | $3.49 | OFFICIAL |
| RunPod Secure Cloud H100 NVL 94 GB | $3.19 | OFFICIAL |
| RunPod Secure Cloud H100 PCIe 80 GB | $2.89 | OFFICIAL |
| RunPod Serverless H100 80 GB | $4.79 | OFFICIAL |
| RunPod Community Cloud H100 (third-party price index, Oct 2, 2026) | $1.99 | REPORTED |
| Lambda on-demand H100 SXM 80 GB | $3.99 | OFFICIAL |
| Lambda reserved / 1-Click Cluster H100 | Contact sales | QUOTE-ONLY |
| Vast.ai H100 SXM, cheapest interruptible listing (Oct 2, 2026 index) | $0.89 | REPORTED |
| Vast.ai H100 SXM, verified US/CA on-demand hosts | $2.70–$6.18 (median $4.44) | REPORTED |
Cost for 10,000 GPU hours
| Option | Cost for 10,000 GPU-hours |
|---|---|
| Vast.ai interruptible, cheapest listing | $8,900 |
| RunPod Community Cloud (reported) | $19,900 |
| Vast.ai verified host, low end | $27,000 |
| RunPod Secure H100 PCIe | $28,900 |
| RunPod Secure H100 NVL | $31,900 |
| RunPod Secure H100 SXM | $34,900 |
| Lambda on-demand H100 SXM | $39,900 |
| Vast.ai verified host, median | $44,400 |
| RunPod Serverless H100 | $47,900 |
| Vast.ai verified host, high end | $61,800 |
What 10,000 GPU-hours means in practice:
- An 8× H100 node for 1,250 hours (about 52 days)
- 32 GPUs for about 13 days
- A single GPU for about 14 months
The costs around the GPU
Storage on RunPod. A 2 TB network volume for datasets and checkpoints costs $100 a month at the over-1 TB rate. Idle volume disk costs double ($0.20/GB-month). A 1 TB volume left attached to a stopped pod costs $200 a month.
Egress. Lambda says there are no egress fees. RunPod's pricing page doesn't list egress charges. On Vast.ai, each host sets its own bandwidth price, so moving a 1 TB checkpoint can cost anything from nothing to a meaningful amount. Read the listing before you rent.
Interruption cost on Vast.ai. An $0.89 interruptible GPU is only cheap if your job survives preemption. Assume 10% of the work gets redone after interruptions. 10,000 useful hours then need about 11,111 paid hours, and the real cost of that listing rises to about $9,900. That's still far cheaper, but only for checkpoint-friendly training or batch inference.
Multi-GPU fabric. For multi-node training, the NVLink/InfiniBand fabric matters more than the hourly rate. Lambda's large-cluster pricing and RunPod's reserved capacity are quote-only. Vast.ai rarely offers matched multi-node InfiniBand clusters.
Which one fits
- Predictable budget for production training or inference: RunPod Secure Cloud. It's $5,000 cheaper than Lambda per 10,000 SXM hours at list price.
- Heavy data movement or large local disk: Lambda. The no-egress-fee policy and bundled NVMe can close the $0.50/hour gap if you move many terabytes out.
- Interruptible, checkpointed workloads on a tight budget: Vast.ai, using reliability filters and hosts with good uptime history.
- Bursty inference with idle gaps: RunPod Serverless at $4.79/hour. You only pay while workers run, but sustained use costs 37% more than an always-on SXM pod.
Assumptions
- 10,000 GPU-hours of on-demand runtime, US regions, October 2, 2026 list prices
- Excludes taxes, storage and egress unless stated
- Vast.ai figures come from third-party price indexes, because live Vast.ai prices change by the hour
- RunPod Community Cloud price is reported by a third-party index, not RunPod's pricing page
Pricing sources
- RunPod pricing: https://www.runpod.io/pricing
- Lambda pricing: https://lambda.ai/pricing
- Vast.ai pricing (live marketplace): https://vast.ai/pricing
- Daily GPU price index (RunPod/Vast.ai): https://www.madebyagents.com/hardware/gpu-rental-prices
- H100 price tracker (Vast.ai verified hosts): https://www.thundercompute.com/blog/nvidia-h100-pricing