Cloud GPU Pricing for AI: Live Price Comparison (2026)
Live hourly prices for cloud GPUs that can run AI models — updated automatically several times a day. Pick how much VRAM your model needs (check with our LLM VRAM Calculator) and see the cheapest way to rent it.
| GPU | VRAM | Provider | Type | $ / hour |
|---|---|---|---|---|
| Loading live prices… | ||||
How to choose
Community/spot instances (RunPod Community, Vast marketplace) are the cheapest way to run inference or fine-tune for a few hours — they can be preempted, so checkpoint often. Secure/on-demand instances cost more but stay up. For a 70B model at Q4 you want 48 GB+ on one card or 2× 24 GB; for 8-14B models a 16-24 GB card is plenty — see the exact math in the VRAM calculator.
FAQ
What is the cheapest cloud GPU for LLM inference?
Community-cloud RTX 4090/3090 instances (24 GB) regularly go for $0.20-0.40/hour and run any model up to ~32B at Q4. For 70B-class models, community A100/H100 80GB instances are usually the cheapest workable option.
How much does it cost to run a 70B model in the cloud?
On a community-cloud 80 GB card at roughly $1-1.5/hour, casual use (an hour a day) is under $50/month — often cheaper than buying hardware unless you run it constantly.
Are community/spot GPUs safe to use?
They're fine for inference and experimentation: your data lives on the instance while it runs. For sensitive data or guaranteed uptime, choose secure/on-demand instances from an established provider.
Should I rent a cloud GPU or buy one?
Rule of thumb: if you'd use it under ~4 hours a day, renting wins. A $1,600 RTX 4090 equals roughly 5,000 hours of community-cloud 4090 time.
Prices from provider APIs (live) and public price lists (dated). Not affiliated with any provider. Companion tool: LLM VRAM Calculator.
