Where can I rent a GPU cheaply to run or fine-tune AI models?
Last verified 2026-11-17Paid placement never changes our picks.
Short answer
As of November 2026, start with RunPod if you want per-second GPU pods in community or secure tiers, Modal if you want serverless code with no VM to manage, and Google Colab if you only need a free notebook. Interruptible or community capacity can be paused mid-job, and listed prices change often.
Top picks
Per-second GPU pods with a cheaper community tier
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
A100 80GB PCIe is $2.09 an hour on Secure Cloud and from $1.19 on Community Cloud. Community Cloud has no uptime warranty.
Serverless GPU functions with nothing to manage
- Free tier
- $30/month free credits on Starter, up to 3 seats; compute billed beyond credits
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
Billed per second with nothing charged while idle. The free Starter plan includes $30 of monthly credits.
Marketplace GPUs with a cheaper interruptible tier
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
Per-second billing. Interruptible instances may be paused, and storage bills until deleted. Rates load live, so none are shown.
Managed fine-tuning and preemptible H100 clusters
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-10-12
H100 is $1.99 an hour preemptible and $3.99 on demand. Fine-tuning is billed per token with per-model minimums.
Free notebooks for small experiments
- Free tier
- Free GPUs/TPUs heavily restricted; notebooks run at most 12 hours
- Starting price
- See pricing page
- Platforms
- Web
- Last verified
- 2026-11-17
Google calls free GPU access heavily restricted and publishes no limits; runtimes are capped at 12 hours. Paid plan prices did not load.
On-demand GPU instances and clusters
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
A100 80GB is listed at $3.19 per GPU hour and H100 at $4.59, before tax. No free credits stated.
Running hosted models by API without managing GPUs
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
A100 80GB is $5.04 an hour, billed per second. Replicate's blog says it is joining Cloudflare.
Free notebooks for ML practice
- Free tier
- See pricing page
- Starting price
- See pricing page
- Platforms
- Web
- Last verified
- 2026-11-17
Documents a free weekly GPU and TPU quota; the exact hours were not readable for us.
GPU VMs if you already use Google Cloud
- Free tier
- $300 free credits for new customers; terms not confirmed
- Starting price
- See pricing page
- Platforms
- Web, API
- Last verified
- 2026-11-17
An A100 80GB VM is listed at about $5.07 an hour on demand; prices vary by region. New customers get a $300 credit.
Prices come from each tool's fact block. "See pricing page" means we have not confirmed the number on the official pricing page in the last 30 days, so we do not show one.
How to choose
- If a job can resume after an interruption → use the cheaper tiers: Vast.ai interruptible instances (may be paused) or RunPod Community Cloud (no uptime warranty).
- If you do not want to manage a server → Modal or Replicate; both bill per second, and Modal charges nothing while idle.
- If you only need a small experiment → Google Colab or Kaggle free notebooks; neither publishes guaranteed limits.
- If you want managed fine-tuning → Together AI bills fine-tuning per token with per-model minimums, instead of renting a raw GPU by the hour.
- Check storage and egress before you commit: Vast.ai storage keeps billing until you delete the instance and bandwidth varies by host; Modal's Starter plan includes 1 TiB of egress a month.
How we checked
- 9 of 9 picks have a Toolsfine fact block.
- 7 of 9 picks had their pricing checked on the official site in the last 30 days.
- 1 of 1 models also suggested RunPod.Panel: Claude (2026-11-17).
- Verified on 2026-11-17.
Listed in other directories
| Tool | theresanaiforthat.com | futurepedia.io | toolify.ai |
|---|---|---|---|
| RunPod | Not checked | Not checked | Not checked |
| Modal | Not checked | Not checked | Not checked |
| Vast.ai | Not checked | Not checked | Not checked |
| Together AI | Not checked | Not checked | Not checked |
| Google Colab | Not checked | Not checked | Not checked |
"Not checked" means we could not open that directory's page; it does not mean the tool is missing there.
Where sources disagree
- Prices are snapshots: GPU prices change often. Vast.ai's rates load live and could not be read, Colab's plan prices did not load and Kaggle's quota was not readable, so none are shown.
- Hourly figures are not like for like: RunPod, Lambda, Replicate and Google Cloud list A100 80GB variants; Together lists H100 only. Compare the same GPU and billing unit before you decide.
- Replicate and Cloudflare: Replicate's own blog says it is joining Cloudflare, so its prices and terms may change.
- Compliance claims: Vendor states SOC 2 Type II on Modal, RunPod and Lambda pages; Vast.ai lists SOC 2 as a claim. We did not read audit reports.
Related answers
Changelog
- 2026-11-17 · Picks, prices and sources re-checked
- 2026-11-17 · Page updated
- 2026-11-17 · First published
Last verified 2026-11-17 · next re-check 2027-02-15. Report an error
Researched with AI assistance and checked by our editors. AI can make mistakes and vendors change plans often — confirm prices and terms on the vendor's site before you buy.
Paid placement never changes our picks. Data: JSON