Short answer. On the median on-demand offers checked 11 November 2026, Vast.ai was cheaper for A100 SXM ($0.72/hr vs RunPod $1.39 Community / $1.79 Secure), H100 SXM ($2.67 vs $2.69 / $3.99) and RTX 4090 ($0.45 vs $0.34 / $0.89). RunPod Secure is the more defensible production option; Vast is often cheaper for experimentation and interruptible work.
On-demand rates
| GPU | Vast.ai median | RunPod Community | RunPod Secure |
|---|---|---|---|
| RTX 3090 24GB | $0.21 | $0.22 | $0.50 |
| RTX 4090 24GB | $0.45 | $0.34 | $0.89 |
| RTX 5090 32GB | $0.62 | $0.69 | $1.19 |
| A100 PCIe 80GB | $0.67 | $1.19 | $1.79 |
| A100 SXM 80GB | $0.72 | $1.39 | $1.79 |
| H100 PCIe 80GB | $2.47 | $1.99 | $2.89 |
| H100 SXM 80GB | $2.67 | $2.69 | $3.99 |
Vast.ai: median of rentable on-demand single-GPU offers from its public offers API at 11:03 CET, 11 November 2026. RunPod: pricing page checked 20 November 2026. Marketplace prices move hourly; treat these as a snapshot.
100-hour cost example
RTX 4090, 100 hours:
Vast $45 · RunPod Community $34 · RunPod Secure $89
A100 SXM, 100 hours:
Vast $72 · RunPod Community $139 · RunPod Secure $179
H100 SXM, 100 hours:
Vast $267 · RunPod Community $269 · RunPod Secure $399
At 100 hours, the gap is large enough to test both. At 10 hours, the setup and migration time can cost more than the GPU saving.
Pricing model
| Dimension | RunPod | Vast.ai |
|---|---|---|
| Marketplace vs managed | Secure Cloud (T3/T4 data centers) and Community Cloud | Marketplace of independent hosts |
| Billing step | Per second for Pods | Per-second/minute, varies by host |
| Interruptible | Spot options | Interruptible offers are common |
| Serverless | Yes | Not the primary model |
| Storage | Network and container volumes | Host-dependent; check disk and bandwidth |
| Reliability | Higher on Secure Cloud | Varies by host and rating |
What the hourly rate leaves out
- Storage: model weights and datasets can cost more than the GPU for short jobs.
- Egress: downloading checkpoints or serving large outputs can add bandwidth fees.
- Idle time: a pod left running between experiments still bills.
- Setup time: driver, CUDA, container and model download time is real engineering cost.
- Interruptions: a cheaper interruptible offer can fail mid-job and erase the saving.
- Reliability: a host with poor disk or network can make a fast GPU feel slow.
When to choose Vast.ai
- You are experimenting and can tolerate host variability.
- You want the lowest median rate for A100 or H100 capacity.
- Your job is checkpointed and can resume after interruption.
- You can verify the host and benchmark before committing.
When to choose RunPod
- You need a data-center tier for production or compliance.
- You want managed storage, serverless and a predictable platform.
- You value support and a stable catalog over the lowest marketplace bid.
- You need per-second billing for short interactive jobs.
When neither is the right answer
If you do not need GPU operations, a managed model API is cheaper than renting a GPU that sits idle. Plugsky's Hobby is $8/month, Starter $20, Builder $60 and Scale $120; Builder and Scale have no monthly token cap, with fair-use concurrency and RPM. See pricing and the self-hosting break-even calculator.
FAQ
Is Vast.ai cheaper than RunPod?
For the median on-demand offers checked 11 November 2026, yes for A100 SXM, H100 SXM and RTX 4090 versus RunPod Secure; the RTX 4090 median was higher than RunPod Community. Marketplace prices vary by host.
What is the difference between RunPod Secure and Community Cloud?
Secure Cloud runs in T3/T4 data centers with higher reliability; Community Cloud uses vetted individual providers at lower prices. Both bill by the second.
Does Vast.ai guarantee host reliability?
No. It is a marketplace; check host rating, verification, interruptible terms and benchmarks before production.
Which is better for LLM inference?
Vast for cheap experimentation and checkpointed batch work; RunPod Secure for production-grade hosting. For a managed API without GPU operations, use a model API.
What hidden costs should I compare?
Storage, egress, idle time, setup time, interruptions and engineering hours. Cost per completed job matters more than cost per GPU-hour.
Reviewed 20 November 2026. Vast.ai medians are from its public offers API on 11 November 2026; RunPod rates are from its pricing page checked 20 November 2026. Marketplace prices change hourly; verify before purchase. This page is editorial and does not rank providers by payment.