Vast.ai Review: GPU Cloud Pricing

Vast.ai is a GPU cloud marketplace where prices move with supply and demand. You pay by the second, from $0.09 an hour for an RTX 3090 to $1.77 an H100.

Quick answer: Vast.ai is a marketplace, not a classic cloud. Independent hosts and data centers list GPUs, and the market sets the hourly rate, so the same card can cost very different amounts from one offer to the next. You add credit (from $5), pick an offer and launch a container or VM from the console, CLI, Python SDK or API, and you pay per second. Prices below are the live rates on Vast.ai’s pricing page on November 14, 2026, and they change hourly.

What Vast.ai does

• GPU Cloud: on-demand instances on 20,000+ GPUs in 40+ data centers, with 68+ GPU types from the RTX 3060 to the B200 and GB200.
• Serverless: deploy a model as an endpoint that autoscales to zero, so you only pay for compute time.
• Clusters: dedicated multi-node GPU clusters with InfiniBand for large training runs.
• Model Library: ready templates for popular open models, such as Qwen and MiniMax models.
• Built for automation: every GPU is provisioned through code, so scripts and AI agents can search, rent and release machines.

Vast.ai Pricing (2026)

GPU	VRAM	Lowest offer	Median
RTX 3090	24GB	$0.09 / hr	$0.21 / hr
RTX 4090	24GB	$0.13 / hr	$0.49 / hr
RTX 5090	32GB	$0.39 / hr	$0.67 / hr
H100 SXM	80GB	$1.77 / hr	$2.30 / hr
H200	141GB	$3.55 / hr	$4.94 / hr
B200	192GB	$6.67 / hr	$11.25 / hr

Three ways to pay: on-demand instances bill per second with no interruptions. Interruptible instances are 50%+ cheaper but can be reclaimed, so they suit batch jobs that save checkpoints. Reserved capacity on 1, 3 or 6 month terms is up to 50% off. The gap between the lowest offer and the median is the real story: the cheapest H100 was about 23% below the median, so filtering offers by reliability and location before you rent matters as much as the GPU you pick.

If you are setting up a local or containerised stack first, our guide to alternatives to Docker covers the tooling. For hosted inference without renting GPUs, see our WaveSpeed AI review, and for coding help, our list of the best AI agents for coding.

Who Vast.ai is for

Developers, researchers and small AI teams who are comfortable with containers and want the cheapest GPU hours for fine-tuning, image and video generation, or inference. Companies that need one vendor, fixed pricing and hands-off support may prefer a traditional cloud or a reserved contract.

Limitations

• Prices move: the same GPU can cost more tomorrow, so budgets are estimates.
• Hosts vary: reliability, network speed and location differ between offers.
• Interruptible means interruptible: cheap instances can stop mid-job.
• Technical setup: you work with Docker images, SSH or the CLI rather than a one-click app.

Vast.ai is one of the cheapest ways to rent serious GPU power if you can handle containers and check each offer.

Start with $5 on an on-demand card for a test run, then move batch work to interruptible instances once your jobs save checkpoints.