Vast.ai Pricing (2026): What an Instance Really Costs
Verified July 2026 Vast.ai rates: RTX 4090 from $0.13/h, median $0.36/h. The $5 minimum deposit, why the floor and the median differ by 7x, and interruptible vs on-demand.
Vast.ai is a marketplace, not a price list. Hosts set their own hourly rate, so the same RTX 4090 can be $0.13/h on one machine and $0.60/h on another in the same set of search results. Vast publishes a “from” low and a median for each card, and the median is the number worth planning against.
What each GPU costs
| GPU | VRAM | From | Median | Typical use |
|---|---|---|---|---|
| RTX 3090 | 24 GB | $0.05/h | $0.15/h | Experiments, Stable Diffusion |
| RTX 4090 | 24 GB | $0.13/h | $0.36/h | SDXL, fast inference |
| A100 SXM4 | 80 GB | $0.13/h | $0.90/h | 70B fine-tunes |
| H100 PCIe | 80 GB | $1.33/h | $2.00/h | Fastest single-GPU inference |
| H100 SXM | 80 GB | $1.73/h | $3.49/h | Multi-GPU training |
Verified against vast.ai/pricing on 29 July 2026. The catalogue floor across all cards is $0.03/h, which lands on older hardware that is not in this table.
The floor and the median are different products
Look at the A100 SXM4 row: $0.13/h at the low end against $0.90/h at the median. That is a 7x spread on identical silicon, and it is the single most misread number on the platform.
The cheap end is real but it is not representative. A host clearing idle capacity in a region with cheap power will price near the floor, and that machine may come with a slow disk, thin upstream bandwidth, or a reliability score that makes a 40-hour run a gamble. Build a budget on the “from” column and you will be out by a factor of several once you filter for hosts you would actually trust.
What it costs to start
The minimum deposit is $5, payable by credit card, BitPay or Crypto.com (checked on docs.vast.ai in August 2026). There is no free tier and no trial credit.
At the $0.36/h median rate for an RTX 4090, $5 buys roughly 14 hours. At the $0.13/h floor for the same card, closer to 38. When the balance reaches zero Vast stops your running instances automatically, so a long job needs either a bigger deposit or autobilling switched on in the Billing page.
Interruptible instances and the bid
The cheapest rates on Vast are interruptible. You set a maximum bid, and if someone bids higher for that machine they take it, with a notice window measured in minutes. On-demand instances cost more and carry a commitment that the host will not terminate the job without reason.
Interruptible suits checkpointed training, batch inference, and anything that can resume from disk. It is the wrong tier for a job holding user traffic. Write checkpoints on a fixed interval regardless of tier, because a peer host can go offline for its own reasons on either one.
How Vast.ai compares
| Provider | Entry rate | Best for |
|---|---|---|
| Vast.ai | $0.03/h | Cheapest marketplace rates, widest price variance |
| RunPod | $0.16/h | Balance of price, GPU selection and tooling |
| Massed Compute | $0.35/h | Mid-range cards without a marketplace |
| Paperspace | $0.45/h | Managed notebooks |
| AWS GPU (EC2) | $0.53/h | Existing AWS estates |
| Lambda Labs | $0.69/h | Multi-GPU training instances |
| Google Cloud GPU | $0.71/h | Existing GCP estates |
| CoreWeave | $1.25/h | Enterprise contracts, per-node billing |
Entry rates are verified provider floors as of July 2026 and refer to different GPUs. The full GPU cloud comparison matches them on identical hardware, and the Vast.ai review covers host vetting, the interface and the safety question in detail.
FAQ
Is Vast.ai actually cheaper than RunPod Community Cloud?
On consumer cards the two are close. An RTX 4090 is $0.36/h at the Vast median against $0.34/h on RunPod Community, which is a rounding error. The gap opens on datacenter parts: an A100 80GB runs $0.90/h median on Vast against $1.19/h on RunPod Community, about 24% less. What you trade away is RunPod’s template library and a single support desk in place of an individual host.
Why is my bill higher than the hourly GPU rate suggested?
The GPU rate is not the whole invoice. Vast lists storage and internet transfer as separate per-host line items, so a machine with a cheap GPU price and expensive bandwidth can cost more overall than a dearer-looking alternative. Both figures sit on the host’s card in the search results next to the GPU price, and they are worth reading before you filter on the hourly rate alone.