Provider review / verified 2026-09-05

Vast.aiGPU review

Vast.ai is a peer-to-peer GPU marketplace. Independent hosts list machines, you bid or take on-demand, and the platform is the matching layer, not the datacenter. That is the feature and the bug.

Models listed9
From /GPU/hr$0.29
Regions in sample1
Business modelMarketplace

Our reading

Who this provider is for

I rent Vast.ai when the job can die and come back. Fine-tunes with checkpoints, batch inference, anything I can restart without apology. I do not put a production API on an unverified community host and hope the reliability score was in a good mood. Read the listing like a used-server ad: disk, interconnect, datacenter-verified or not, then the dollar. The dollar is often a typo. The host is not.

Rent it if Researchers and indie labs who can checkpoint, who want consumer cards or a cheap H100, and who will actually open the host page before they paste a card.

Skip it if The run is the product, a live endpoint, a multi-day pretrain you cannot snapshot, or a compliance story that needs a named datacenter and an SLA.

How we compare hours

Strengths

  • Often the lowest published hour on consumer GPUs and on unverified H100/A100 inventory
  • Huge catalog: dozens of GPU types, interruptible and on-demand on the same board
  • Per-second billing and a reliability score on the host, if you bother to read it
  • Datacenter-verified listings exist when you are willing to pay up from the floor

Limits

  • Vast.ai is not a cloud with an SLA. It is a marketplace of other people’s machines
  • Unverified hosts are why the H100 looks like $0.90/hr. Treat that number as a bid, not a contract
  • Network, disk, and CUDA stack vary by listing. Your Dockerfile is not enough if the machine is a surprise
  • No native serverless. Idle still exists. You just pay a different person for it

Same standard, every cloud

What we measured on Vast.ai

Four checks we apply to every provider. This is not a score, and it is not a paid ranking.

01

Catalog

We index Vast.ai as a marketplace: many models, many hosts, on-demand and interruptible side by side. The interesting filter is verified vs community, not the homepage hero price.

02

Price signal

Floor prices on Vast.ai regularly undercut managed clouds on RTX 4090/5090 and on unverified datacenter cards. Verified H100 listings sit closer to RunPod Secure than to the meme screenshot.

03

Product shape

You rent a machine, not a platform. SSH, templates, and a host score. Interruptible is cheaper and can vanish. That is not on-demand with a discount sticker.

04

Still verify

Open the host. Check reliability, max duration, interconnect, and whether the listing is datacenter-verified. Then confirm the live rate. Our table is a map, not a reservation.

Model
Marketplace
Headquarters
Austin, US
Setup
SSH / templates on a host VM
Billing grain
Per second
On-demand
Yes, plus interruptible and reserved
Serverless
No native scale-to-zero
Interconnect
Host-dependent (PCIe to NVLink/IB)
Best known for
Consumer GPUs and floor-price H100

Catalog listings

Vast.ai catalog sample

Published on-demand hours first. Spot and waitlists are labeled. Confirm the live rate before you provision.

GPUsRegionBilling typeInterconnect/GPU/hr
RTX 4090In stockUS-EastOn-demandPCIe$0.29Open GPU
RTX 5090In stockUS-EastOn-demandPCIe$0.34Open GPU
L4LimitedUS-EastOn-demandPCIe$0.41Open GPU
A10In stockUS-EastOn-demandPCIe$0.49Open GPU
RTX PRO 6000In stockUS-EastOn-demandPCIe$0.72Open GPU
L40SIn stockUS-EastOn-demandPCIe$0.85Open GPU
A100In stockUS-EastOn-demandPCIe$1.19Open GPU
H100In stockUS-EastOn-demandPCIe$1.73Open GPU
H200LimitedUS-EastOn-demandPCIe$3.19Open GPU

Vast.ai is not a cloud. Say that first.

Most GPU blogs compare Vast.ai to RunPod as if they sold the same object. They do not. RunPod operates capacity. Vast.ai operates a market. The person behind an unverified listing might be a datacenter, a miner with leftover 4090s, or a lab dumping nights. You are underwriting that person.

That is why the hour can look like a bug in the page. It is not. It is residual capacity priced by someone who does not have a sales team. I use it. I also lose jobs on it. Checkpoint or leave.

When the cheap H100 is the wrong H100

An unverified H100 at under a dollar is a research tool. A verified datacenter H100 at one-fifty to two-something is closer to a small cloud. If your trainer cannot resume, you are not saving money. You are buying lottery tickets denominated in sunk steps.

For LoRA and full-parameter fine-tunes that snapshot every N steps, I still start on Vast.ai. For anything a customer will notice going down, I leave.

Consumer cards are the actual product

The internet argues about H100. Vast.ai’s real edge is 4090s and 5090s by the hour. If the adapter fits in 24–32 GB, paying Hopper tax is a skill issue. I would rather have two 4090s I can afford to leave on than one H100 I am scared to idle.

Watch VRAM, not the brand on the die. Watch PCIe vs SXM when the job leaves one GPU. Marketplace listings love to bury that.

How I rent it without getting burned

Filter on-demand. Prefer datacenter-verified if the run is long. Read reliability and the number of unique rentals. Avoid mystery disk. Pull a template you have used before. Checkpoint from minute one. Interruptible only when the work is embarrassingly restartable.

Then I compare the same GPU on RunPod Community and Secure. If Vast.ai is 40% cheaper after I account for a probable restart, I take Vast.ai. If it is 10% cheaper, I take RunPod and go do something else.

How I actually rent it

  1. 01

    Pick the GPU, not the screenshot

    Search the model and VRAM you need. Ignore the homepage floor until the listing is open.

  2. 02

    Vet the host like a used server

    Verified or not, reliability, duration, interconnect, disk. On-demand unless you can die.

  3. 03

    Checkpoint, then confirm the live rate

    Provision, snapshot early, and do not trust a catalog cell more than the host UI.

FAQ

Vast.ai questions

Is Vast.ai cheaper than RunPod?+

Often on raw on-demand for consumer GPUs and unverified datacenter cards. Not always on verified H100, and not if a killed job costs more than the delta. Compare the same GPU, same VRAM, on-demand only.

Is Vast.ai safe for production inference?+

I would not put a customer-facing endpoint on an unverified host. Datacenter-verified plus your own health checks is a maybe. RunPod Secure, Lambda, or a neocloud with an actual SLA is the adult path.

What is interruptible on Vast.ai?+

A cheaper bid the host can reclaim. It is spot with a marketplace accent. Use it for batch. Do not use it for a run you cannot resume.

Does Vast.ai have H100 and H200?+

Yes, as listings, not as a guaranteed pool. Stock and price move with hosts. Open the model page and sort on-demand, in stock.

Who is Vast.ai for?+

People who can read a host page and checkpoint. Students, indie fine-tunes, scrapers of cheap 4090 hours. Not teams who need a PDF that says uptime.

Sources

Where these numbers come from

Prices and SKUs are from the GPUBeacon catalog sample. Ranking rules are on the methodology page. Outbound links may be affiliate. They never change the hour we display.

GPUs

Keep comparing

Other providers to compare

Same fields, different product. Compare the GPU hour before you lock a brand.