If the work can be interrupted and the data is not sensitive, stay on Vast.ai. Low prices are the reason to be there, and the risks are ones you can manage with checkpoints. If either condition fails, the question becomes what the step up costs, and the live table on this page answers it for four common GPUs.
One middle option is easy to miss. Vast.ai has a datacenter tier of its own, so the first alternative to a Vast.ai consumer host is a different Vast.ai listing.
What you are actually renting on Vast.ai
Vast.ai does not own the GPUs. Its docs, read on 21 September 2026, describe it as "a cloud computing, matchmaking, and aggregation service" whose software "allows anyone to easily become a host by renting out their hardware". Hosts "download and run our management software, list their machines, configure prices". Vast.ai says it adds "no markup layered on top of host-set prices". That is why there is no price list to quote: the price is whatever a host asks today.
What you get is a container. Instances are "Docker containers that give you exclusive GPU access". The GPU is "never shared between users", while CPU, memory and storage are a share of the host machine. Linux only, no remote desktop.
Two kinds of host
The listings mix two very different things. Vast.ai's security FAQ puts it this way: "Tier 4 datacenters have extensive physical and operational security" and "Individual hosts may have less formal security measures". A consumer host can be anyone who installed the hosting software.
The filter that separates them is "Secure Cloud (Only Trusted Datacenters)", which shows vetted partners "current with ISO 27001 and/or Tier 3/4 standards". The docs say to "look for the blue datacenter label". Vast.ai's about page also states a SOC 2 Type I report dated 2025. That covers Vast.ai the company. It does not cover each host.
The word that misleads people is "verified". Vast.ai says host verification is "fully automated and powered by proprietary algorithms", based on reliability and configuration, with "no manual intervention". So a verified badge means the machine behaves. It does not mean anyone audited the person who owns it.
What that means for your data
Vast.ai documents isolation between renters: "Clients are isolated in unprivileged Docker containers". The pages retrieved give no guarantee about what a host can or cannot see on a machine they physically control. We treat that as unknown. The practical reading: model weights you downloaded from a public hub are fine on any host. Customer data, private weights and long-lived API keys belong on the datacenter tier or on another provider.
What that means for reliability
There are three rental types, and they fail differently.
- On-demand has a "fixed price set by the host" and "cannot be interrupted". It does have an end date. The host sets a maximum duration, and "when the rental contract ends, hosts may extend the rental or stop the instance". Expired instances "may be deleted 48 hours after expiration".
- Interruptible is a bid. "You set a bid price", and the instance "can be stopped by higher bids". Vast.ai says this "saves 50-80% on costs". When you are outbid, "your instance is stopped (killing running processes)" and it "may wait long to resume". On-demand renters always have priority. The instance is paused, not deleted, and there is no 24-hour cap of the kind Google's old preemptible instances had.
- Reserved is a prepaid discount of "up to 50%" by commitment length, with "credits locked to the specific instance". You are committing to one host's machine, not to Vast.ai in general.
For comparison, a hyperscaler spot instance comes with a stated warning: two minutes on AWS and 30 seconds on Azure, per their docs read on 21 September 2026. Vast.ai's docs state no notice period. Write checkpoints as if there is none.
Vast.ai's pricing guide is candid about the trade: "Higher reliability scores typically correlate with higher prices". The cheapest listing and the dependable listing are rarely the same row.
Two billing details catch people. "Storage charges continue even when instances are stopped", so stopping is not the same as ending the bill. And bandwidth is priced per host: "Data transfer costs vary by host and include both upload and download traffic." RunPod and Lambda both say they charge nothing for ingress or egress. Our egress reference has the rest.
The managed premium, live
This is the central number. Each cell is the cheapest current on-demand price from that provider for that GPU.
| Provider | RTX 4090 $/GPU-hr | RTX 5090 $/GPU-hr | A100 $/GPU-hr | H100 $/GPU-hr |
|---|---|---|---|---|
| Vast.ai | $0.45 | $0.53 | none in stock | $3.07 |
| RunPod | $0.74 | $0.99 | $1.59 | $3.19 |
| Massed Compute | none in stock | none in stock | $1.35 | $2.73 |
| Hyperstack | none in stock | none in stock | none in stock | $2.50 |
| Lambda Labs | none in stock | none in stock | $1.99 | none in stock |
| VERDA | none in stock | none in stock | $1.27 | $3.37 |
| Vultr | none in stock | none in stock | $2.80 | none in stock |
Read the Vast.ai column with care, because of how this site treats marketplaces.
- For Vast.ai we collect only datacenter-hosted, verified, high-reliability listings. Consumer-host listings are not in our data at all. So the gap you see is the premium of a managed cloud over Vast.ai's safer tier. Against a consumer host the gap is probably wider, since Vast.ai's own pricing guide ties lower prices to lower reliability scores. Vast.ai's search page is the only place to check that floor.
- Offers from peer-to-peer hosts on any provider are tagged P2P. They appear on provider pages, separately, and they never set the "current price" headline on a rent page such as RTX 4090 rental. That is why a marketplace's cheapest price and our headline price can differ.
- Salad is a marketplace, and every offer on its provider page is tagged P2P in our data, so it has no column here. RunPod's column is Secure Cloud only, for the same reason.
- The table shows only in-stock, on-demand offers seen in the last 15 minutes. Stock is rechecked about every minute.
To turn the gap into a budget, multiply the hourly difference by the hours you will run. A fine-tune that takes a few hours on an RTX 4090 hides almost any premium. A card held around the clock for a month does not.
The second view is one GPU across pricing types. It shows the cheapest on-demand, spot and reserved listings we hold for the RTX 4090, again without peer-to-peer hosts.
If the spot figure is close to the on-demand figure, interruption is buying you little and on-demand is the better deal. If it is far below, checkpointing pays for itself quickly.
Alternatives by what you need
Secure datacenter GPUs
Start with the Secure Cloud filter on Vast.ai itself. You keep the account, the tooling and per-second billing.
The closest outside match is RunPod Secure Cloud, which RunPod says "operates in T3/T4 data centers". It is also container based, billed by the second, with no ingress or egress fees. RunPod's cheaper Community Cloud is a peer-to-peer pool like Vast.ai's consumer hosts, and RunPod says it is no longer accepting new hosts for it. RunPod vs Vast.ai compares the two.
Past that are conventional GPU clouds: Hyperstack, Lambda, Verda and Vultr. SemiAnalysis rates GPU providers for cluster workloads in its ClusterMAX system. That rating says little about renting a single GPU, and our neocloud explainer covers it. Lambda vs Vast.ai and Hyperstack vs Vast.ai show current prices.
Consumer GPUs from a company, not a host
If you came to Vast.ai for GeForce cards, the step up is harder, because not every managed cloud lists them. The RTX 4090 has 24 GB of memory and the RTX 5090 has 32 GB. The spec table puts them next to an A100.
RunPod's price list on 21 September 2026 included both the RTX 4090 and the RTX 5090 on Secure Cloud. Lambda's instance list on the same date had no GeForce cards. For the rest, including Massed Compute, trust the live table above over any list in an article.
Salad is a sideways move, not a step up. It is a company with one bill, but every Salad machine in our data is a peer-to-peer host, so the data question is the same one you have on Vast.ai.
Long-running training
Vast.ai's docs say: "For long-running jobs (days/weeks), verify host reliability scores". Add the host's maximum rental duration to that check, since the contract can end before your run does.
If you would rather not manage that, the alternatives differ in what they promise. RunPod describes on-demand pods as resources that "cannot be displaced by other users", and sells savings plans paid three or six months upfront. Lambda has no interruptible tier on its price list. Its instances are first-come and cannot be stopped, only terminated, and billing runs by the minute until you do. Vultr rents whole bare metal GPU servers, per its public plan API on 21 September 2026.
One hardware limit applies wherever you rent. The RTX 4090 and RTX 5090 connect over PCIe only, with no NVLink, so in multi-GPU training all traffic between cards goes over PCIe. SXM versions of the A100 and H100 have NVLink for that. For multi-node work, see GPU cluster pricing.
Before you put data on a marketplace host
- Decide which tier the data allows. Public weights and public datasets: any host. Anything you would not publish: turn on the Secure Cloud filter and check for the blue datacenter label.
- Read "verified" as a machine health check. It is automated and it is not an audit.
- Check the reliability score and the maximum rental duration before you rent. An expired instance may be deleted 48 hours later.
- Check the host's bandwidth price in both directions before you pull a large dataset. It varies by host and is charged per byte.
- Keep long-lived secrets out of the image and the environment. Use credentials you can revoke the moment the job ends.
- Checkpoint to storage outside the host. An interruptible instance is stopped without a stated warning, and its processes are killed.
- Delete the instance when you finish. A stopped instance still bills for storage.
- Keep the prepaid balance small. Vast.ai's billing docs say: "After spending credits, there are absolutely no refunds." At zero balance instances are stopped, and without a saved card "your instances and data will be scheduled for deletion".
The rule
Stay on Vast.ai consumer hosts when the job can restart and nothing on the disk is private. Move to Vast.ai's datacenter tier or RunPod Secure Cloud when the data is private but the budget is tight. The table above shows which is cheaper today. Move to Lambda, Hyperstack, Verda or Vultr when the job must run for days without your attention, and accept the premium as the price of not checking on it. If the live premium for your GPU is small, take the managed option, because your time costs more than the difference.
Sources
Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.
All accessed 21 September 2026.
- Vast.ai docs: General FAQ
- Vast.ai docs: Security FAQ
- Vast.ai docs: Understanding verification
- Vast.ai docs: Pricing
- Vast.ai docs: Billing
- Vast.ai docs: Rental types
- Vast.ai docs: Instance types
- Vast.ai docs: Instances overview
- Vast.ai pricing
- Vast.ai about page
- RunPod pricing
- RunPod docs: Pods overview
- RunPod docs: Choose a Pod
- RunPod docs: Pod pricing
- Lambda pricing
- Lambda docs: Billing
- Lambda docs: Creating and managing instances
- Vultr bare metal plan API
- AWS docs: Spot Instance interruption notices
- Microsoft docs: Azure Spot Virtual Machines