Vast.ai is the cheapest place to rent a GPU for AI in 2026. An RTX 4090 starts at about $0.39 an hour, an RTX 5090 at about $0.49 and an RTX 3090 at just $0.17.
It’s cheap because Vast.ai is a marketplace. Thousands of hosts, from home rigs to datacenters, list their machines and set their own prices. That also means quality varies, and there are a few billing traps that catch new users.
This guide walks you through everything step by step: creating an account, picking the right GPU, launching ComfyUI or Ollama from a template, how billing really works, how to keep your data safe, and how to fix common problems.
- Prices (14 Nov 2026 floors): RTX 3090 $0.17/hr, RTX 4090 $0.39, RTX 5090 $0.49, RTX PRO 6000 $1.40, H100 PCIe $2.85.
- Getting started: sign up, verify your email, add at least $5 credit, add an SSH key, pick a template, then rent.
- Templates: ComfyUI and Open WebUI (Ollama) launch in one click and open from the Instance Portal.
- Billing traps: disk size can’t be changed later, stopped instances still pay for storage, and some hosts charge a lot for downloads.
- Safety: you run on someone else’s machine, so use Secure Cloud (datacenter) hosts for anything private.
What Is Vast.ai and How Does It Work?
Vast.ai is a GPU cloud marketplace. Instead of owning all its hardware like AWS or Lambda, it connects you with hosts who rent out their GPUs. You pick a machine from a searchable list, choose a software template (a ready-made Docker image), and the machine starts with your software inside a container.
There are two kinds of hosts:
- Community hosts: individuals and small operators. They’re usually the cheapest, but reliability and security vary.
- Secure Cloud hosts: vetted datacenter partners. They cost a bit more but are more reliable and better for private data. You can filter for them in search.
Every listing shows a reliability score from 0 to 1, based on the host’s uptime history, and a DLPerf score, which is a measured deep learning performance figure you can use to compare machines.
Vast.ai GPU Prices in November 2026
Prices change all the time because hosts compete. These are the floor prices (the median of the five cheapest verified single-GPU offers) recorded by GPU Economy on 14 November 2026 at 23:40 UTC.
| GPU | VRAM | From ($/hr) | Good for |
|---|---|---|---|
| RTX 3090 | 24 GB | $0.17 | Cheapest 24 GB card; SDXL, Flux, 14B-32B LLMs |
| RTX 4090 | 24 GB | $0.39 | Faster image and video generation |
| RTX 5090 | 32 GB | $0.49 | Big image/video models, 32B LLMs at higher quants |
| A100 40GB | 40 GB | $0.43 | Fine-tuning small models, 70B at low quants (2 cards) |
| A100 80GB | 80 GB | $1.07 | 70B models, LoRA training |
| RTX PRO 6000 Blackwell | 96 GB | $1.40 | Large video models without quantizing, 70B+ LLMs |
| H100 PCIe | 80 GB | $2.85 | Serious training and fast serving |
| H100 SXM | 80 GB | $3.84 | Multi-GPU training |
| H200 | 141 GB | $4.34 | Very large models |
| B200 | 192 GB | $11.89 (single), $8.75 (4-GPU nodes) | Frontier-scale work |
H100 floors on verified hosts have risen since October, when Spheron reported about $0.90-$2.27/hr for H100 PCIe including unverified hosts. Always check live listings before you plan a budget.
To compare these prices with RunPod, Lambda and the big clouds, see my cheapest cloud GPU for AI guide. To see how much VRAM each model needs, check how much VRAM do I need for AI.
How Billing Works (and the Traps)
| Charge | How it’s billed | Watch out for |
|---|---|---|
| GPU | Per second while the instance runs | Stops when you stop the instance |
| Storage (disk) | Continuously while the instance exists, about $0.10-$0.15 per GB per month | Still charged when stopped; only deleting ends it |
| Bandwidth | Per byte, upload and download, set by each host | Some hosts charge $20+ per TB; big model downloads add up |
Three rules will save you money:
- Pick your disk size carefully. Vast.ai says the disk size you choose cannot be changed later. To get more space you have to create a new instance and copy your files across.
- Delete when you’re done. Stopping pauses the GPU charge but storage keeps billing. A stopped ComfyUI box with a big disk can cost around $1.30 a day.
- Check the download price. Use the bandwidth filter in search. A host with a cheap GPU rate but expensive bandwidth can cost more overall if you download large models.
When your balance reaches zero, Vast.ai stops your instances. If you have no card saved, your instances and data are scheduled for deletion unless you add credit quickly. Turn on auto-billing if you run long jobs.
On-Demand vs Interruptible vs Reserved
| Type | Price | Can it be stopped by others? | Best for |
|---|---|---|---|
| On-demand | Standard rate | No, high priority | Interactive work like ComfyUI and chat |
| Interruptible | Often over 50% cheaper | Yes, if outbid or an on-demand renter needs the machine | Batch jobs with checkpoints, bulk image generation |
| Reserved | Up to 50% off on-demand | No, high priority | Long projects; you prepay for a set period |
Interruptible instances work on bids. If someone bids higher, or an on-demand renter takes the machine, your instance pauses. Your data stays on disk and the instance resumes when it regains priority, but anything running is interrupted, so save your work often.
To get a reserved price, rent an on-demand instance first, then click the discount badge on its instance card to convert it. The prepaid credit is locked to that instance, and cancelling early gives only a partial refund.
How to Get Started Step by Step
- Create an account at vast.ai and verify your email. You can’t rent until it’s verified.
- Add credit. Go to Billing and click Add Credit. The minimum is $5, and you can pay by card, BitPay or Crypto.com. Turn on auto-billing to avoid surprise shutdowns.
- Add an SSH key (optional but recommended). Create a key pair on your PC with
ssh-keygen -t ed25519and paste the public key (the.pubfile) into the Keys page. - Pick a template. Open Templates and choose one such as ComfyUI, Open WebUI (Ollama) or PyTorch.
- Search for a machine. Filter by GPU type, number of GPUs, VRAM, reliability, location and price, and set the disk size with the slider.
- Click Rent. The instance appears under Instances. A fresh image can take anywhere from a couple of minutes to an hour to download, depending on the host’s connection.
- Click Open when it’s ready. This opens the Instance Portal with links to your apps, Jupyter and a terminal.
- Stop or delete the instance when you finish. Copy anything you want to keep first.
How to Choose the Right GPU and Host
| What you want to do | GPU to rent | Disk |
|---|---|---|
| SDXL, Flux and normal image work | RTX 3090 or 4090 (24 GB) | 100-150 GB |
| Video models (Wan, MiniMax H3, Hunyuan) | RTX 5090 (32 GB) or RTX PRO 6000 (96 GB) | 200-300 GB |
| Chat with 7B-32B LLMs | RTX 3090, 4090 or 5090 | 50-100 GB |
| 70B LLMs | 2x RTX 4090, A100 80GB or RTX PRO 6000 | 50-150 GB |
| LoRA training | RTX 4090/5090 for images; A100/H100 for LLMs | 100-200 GB |
When comparing hosts, I’d check these in order:
- Reliability: aim for 0.95 or higher for on-demand work and at least 0.90 for interruptible jobs.
- Download speed: 500 Mbps or more makes model downloads much faster.
- Bandwidth price: lower is better if you’ll pull large models.
- Secure Cloud: tick this filter for private data or client work.
- Location: a nearby host feels snappier in the ComfyUI interface.
Not sure which card you need? My RTX 5090 vs RTX 5080 for AI and best GPU for local AI guides explain what each card handles.
How to Run ComfyUI on Vast.ai
- Choose the official ComfyUI template from the Templates page.
- Set the disk to at least 100 GB, or 200 GB+ if you use video models. Remember you can’t change it later.
- Rent a 24 GB or larger GPU, such as an RTX 4090 or 5090, with good reliability and a reasonable bandwidth price.
- Wait for the Open button, click it, then click Launch Application next to ComfyUI in the Instance Portal.
- Download models. Use ComfyUI Manager, or open the Jupyter terminal and use
wgetinside the right folder, such asComfyUI/models/checkpoints,models/diffusion_models,models/vaeormodels/loras. - Press R in ComfyUI to refresh the model lists, then load your workflow.
- Download your outputs from
ComfyUI/outputthrough Jupyter before you delete the instance.
For gated Hugging Face models, add your Hugging Face token to the download command, for example wget --header="Authorization: Bearer YOUR_TOKEN" URL. Never paste tokens into a shared template.
New to ComfyUI? Start with my ComfyUI guide for beginners. If you prefer RunPod’s network volumes, which keep models between sessions, see my ComfyUI on RunPod tutorial.
How to Run Ollama and Open WebUI on Vast.ai
- Select the Open WebUI (Ollama) template.
- Set enough disk and VRAM. Vast.ai’s own example for
deepseek-r1:70buses a 50 GB disk and at least 44 GB of total GPU RAM, such as 2x RTX 4090. - Rent the machine and click Open, then click the Open WebUI link in the Instance Portal.
- Create the admin account straight away so nobody else can claim it.
- Pull a model from Admin Panel, Settings, Models, using the exact name from ollama.com, for example
qwen3:32b. - Use the API if you want. Ollama listens on port 11434 inside the container. Find the external port with the IP button on the instance card and send the instance token as a
Bearerheader.
To run the same setup on your own PC, follow my Open WebUI with Ollama setup guide.
Is Vast.ai Safe?
Vast.ai itself is a legitimate company, and the platform holds SOC 2 Type II certification. But individual community hosts are not audited, and your container runs on their hardware. In principle a host could see what’s on the disk.
- For hobby work like generating images or testing open models, community hosts are fine.
- For private data, client files or company code, use Secure Cloud hosts only.
- Don’t store long-lived secrets (API keys, tokens) on the instance. Use short-lived tokens and revoke them after.
- Protect web apps. Open WebUI and ComfyUI links include an access token. Don’t share those links.
Vast.ai vs RunPod vs Lambda
| Feature | Vast.ai | RunPod | Lambda |
|---|---|---|---|
| Model | Marketplace of hosts | Own cloud plus community | Own datacenters |
| Price | Usually the cheapest | Slightly higher, fixed prices | Higher, datacenter GPUs only |
| Consumer GPUs (3090, 4090, 5090) | Many | Yes | No |
| Persistent storage between instances | Per-instance disk | Network volumes | Persistent filesystems |
| Reliability | Varies by host | More consistent | Most consistent |
| Best for | Cheapest experiments, bulk jobs | Easy ComfyUI and serverless | Training and production |
My take: start on Vast.ai for the lowest prices, and switch to RunPod if you’re tired of re-downloading models every time you start a new instance.
Troubleshooting
The instance is stuck on “loading” for a long time
The host is downloading the Docker image. This can take up to an hour on slow hosts. If it’s still stuck, delete it and pick a host with faster download speed.
“Bad machine” or the instance won’t start
Delete the instance and rent a different one. Datacenter (Secure Cloud) hosts fail less often.
I ran out of disk space
You can’t resize a running instance. Save your outputs, create a new instance with a bigger disk, and copy files across with Vast.ai’s copy tools or rsync.
ComfyUI won’t load in the browser
Wait 60-90 seconds after the instance starts, then refresh. Open it from the Instance Portal rather than a saved link, because the token can change.
CUDA out of memory
Rent a card with more VRAM, lower the resolution or batch size, or use a quantized model. My CUDA out of memory fix guide has more options.
My bill is higher than expected
Check for stopped instances that still have storage, and look at bandwidth charges from large downloads. Delete anything you don’t need.
Frequently Asked Questions
How much does Vast.ai cost?
You pay per second for the GPU plus storage and bandwidth. On 14 November 2026, floor prices were about $0.17/hr for an RTX 3090, $0.39 for an RTX 4090, $0.49 for an RTX 5090 and $2.85 for an H100 PCIe.
What is the minimum deposit on Vast.ai?
$5. You can pay by card, BitPay or Crypto.com, and you must verify your email before renting.
Does Vast.ai charge when an instance is stopped?
Yes, for storage. Stopping ends the GPU charge, but the disk is billed until you delete the instance.
Is Vast.ai safe to use?
It’s a legitimate SOC 2 Type II certified platform, but community hosts aren’t audited. Use Secure Cloud datacenter hosts for private data.
What is an interruptible instance on Vast.ai?
A cheaper, bid-based rental that can be paused if someone outbids you or an on-demand renter needs the machine. Your data is kept and it resumes later.
Can I run ComfyUI on Vast.ai?
Yes. Use the official ComfyUI template, give it at least 100 GB of disk, and open it from the Instance Portal.
Is Vast.ai cheaper than RunPod?
Usually yes, because hosts compete on price. RunPod costs a little more but offers more consistent machines and network volumes for keeping models between sessions.
Can I change the disk size later?
No. Vast.ai’s docs say the disk size can’t be changed after you create the instance, so pick enough space up front.
Final Thoughts
Vast.ai is the cheapest way to get a powerful GPU for AI in 2026, often for less than the price of a coffee per hour. With an RTX 5090 at around $0.49 an hour, renting beats buying for most people, especially with GPU prices this high.
Just remember the three rules: choose your disk size up front, delete instances you no longer need, and check bandwidth prices before downloading big models. Do that, and Vast.ai is hard to beat.
Sources
- Vast.ai Docs: Quickstart
- Vast.ai Docs: Pricing
- Vast.ai Docs: Instance types (on-demand, interruptible, reserved)
- Vast.ai Docs: Ollama and Open WebUI
- GPU Economy: Vast.ai GPU pricing per hour (14 November 2026)
- Spheron: Vast.ai pricing 2026 (updated 19 October 2026)
- Prompting Pixels: How to run ComfyUI on Vast.ai
- AI Tool Discovery: Vast.ai review (updated November 2026)




