If you are leaving RunPod over price, a marketplace such as Vast.ai is the place to look. If you are leaving because a GPU was not there when you came back for it, look at a conventional GPU cloud such as Lambda, Hyperstack or Vultr, where no individual host sits between you and the machine. If your product runs on RunPod Serverless, moving costs more than the hourly rates suggest, and staying may be the right call.
This page ranks nobody. It sorts the options by the reason you are leaving and lets live prices do the comparing.
Same GPU, RunPod against the rest, right now
The table shows the cheapest current on-demand price per provider for three GPUs on RunPod's price list: the RTX 4090, the A100 and the H100.
| Provider | RTX 4090 $/GPU-hr | A100 $/GPU-hr | H100 $/GPU-hr |
|---|---|---|---|
| RunPod | $0.74 | $1.59 | $3.19 |
| Vast.ai | none in stock | $0.80 | none in stock |
| Lambda Labs | none in stock | $1.99 | $3.29 |
| Hyperstack | none in stock | $1.60 | $2.50 |
| VERDA | none in stock | none in stock | none in stock |
| Massed Compute | none in stock | $1.35 | $2.73 |
| Ori | none in stock | none in stock | $2.90 |
| Vultr | none in stock | $2.80 | none in stock |
Three things to know before you read it.
- RunPod's column is Secure Cloud. Community Cloud machines belong to individual hosts. This site tags those offers P2P and keeps them out of comparison prices, so the RunPod figure here can be higher than the cheapest pod you have seen in RunPod's console.
- Vast.ai's column is its safer tier. For Vast.ai this site collects only datacenter-hosted, verified, high-reliability listings. Consumer-host listings are not in the table, so Vast.ai's own search page may show lower prices than this column.
- Salad has no column. Salad is a marketplace, and every Salad offer in our data is tagged P2P, so none of them can set a price here. Its catalog is on the Salad provider page.
The table shows only in-stock, on-demand offers seen in the last 15 minutes, and stock is rechecked about every minute. So it compares datacenter capacity with datacenter capacity. The lowest current price for each GPU under the same rules, across every provider we track, is below.
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| RTX 4090 | $0.74 | RunPod | 2 |
| A100 | $0.68 | LeaderGPU | 10 |
| H100 | $2.50 | Hyperstack | 9 |
| H200 | $3.43 | QuantaCloud | 5 |
Why people leave RunPod
Most of the reasons come from how RunPod's products are built, so start with what those products are.
What you are leaving: RunPod in 2026
RunPod's pricing page, read on 21 September 2026, describes three products: "Pods for dedicated GPU instances, Serverless for API inference, and Clusters for multi-node jobs." A pod is a container with a GPU attached, billed by the second. RunPod's docs say there are "no fees for data ingress or egress".
Pods come from two pools. Secure Cloud "operates in T3/T4 data centers, providing high reliability and security". Community Cloud "connects individual compute providers to users through a vetted, secure peer-to-peer system", and RunPod's own comparison table rates its reliability as "Variable". On RunPod's price list of 21 September 2026, every GPU offered in both pools was cheaper on Community Cloud.
Three things changed that older reviews miss, all per RunPod's documentation on 21 September 2026.
- "Runpod is no longer accepting new hosts for Community Cloud. Existing Community Cloud resources remain available." The cheap pool is not closing, but no new hosts can join it.
- Serverless active-worker prices are no longer published. The docs describe them as "Discounts available through sales inquiry". Articles that quote a fixed active-worker discount are out of date.
- The pricing docs list two ways to pay for pods: on-demand and a savings plan paid "3 or 6 months upfront". Spot pods are not mentioned on the pages retrieved. That is not proof they are gone, so check the console if you depend on them.
The published serverless rate is the flex worker rate. RunPod's pricing page listed an 80 GB H100 flex worker at $4.79 per hour on 21 September 2026. Check any older figure you have seen against the page.
The full catalog we track is on the RunPod provider page.
Reliability, as far as it can be sourced
The reasons below are the ones RunPod's own documentation supports. Outage counts from review sites that state no method are left out. RunPod's status page on 21 September 2026 showed component uptime bars at 100% over 90 days, next to incidents that included one hour of API errors on 16 September 2026 and a network storage issue in its CA-MTL-3 datacenter in July 2026 that ran for 9 hours 40 minutes. That is a record of incidents, not a failure rate.
The GPU may be gone when you restart
A stopped pod stays tied to the physical machine it ran on. RunPod's troubleshooting docs say: "If another user rents that GPU while your Pod is stopped, the GPU will be occupied when you try to restart." RunPod offers a start with zero GPUs, which it calls "primarily a data recovery feature". If you stop pods at night and resume in the morning, this is the failure you will meet.
Storage is not meant to be permanent
RunPod's docs say plainly: "Runpod is not designed for long-term cloud storage. Back up critical data". A stopped pod's volume is billed at double the running rate. If your balance reaches zero, "Pods with a network volume are stopped and their data is preserved, while Pods without a network volume are terminated and their data cannot be recovered."
Network volumes have strings attached
Network volumes are "only available for Pods in the Secure Cloud". They "must be attached during Pod deployment" and cannot be added or removed later without deleting the pod. On Serverless, attaching one "constrains worker deployments to that volume's datacenter, which may limit GPU availability".
Serverless costs are harder to predict
Cold starts are billed. The docs say: "You're billed from when a worker starts until it fully stops, rounded up to the nearest second", and the billed phases include container start and model load, plus an idle timeout that defaults to 5 seconds. "Active workers incur charges continuously, including when idle." With the active-worker rate now a sales conversation, you cannot price an always-on endpoint from the public page.
Cheaper for the same GPU: marketplaces
Vast.ai is a marketplace, not a cloud. Its docs say "hosts set their own prices", so both price and quality vary by listing. Its Secure Cloud filter limits results to vetted datacenter partners. Our Vast.ai alternatives article covers the host types, what "verified" means and how interruptible bidding behaves, so we do not repeat it here.
You give up two things RunPod includes. Bandwidth is not free: "Data transfer costs vary by host and include both upload and download traffic." On-demand rentals have a maximum duration set by the host. In return you get a bidding tier that RunPod's current pricing docs do not describe: interruptible instances, where "you set a bid price" and the instance is stopped when someone bids higher. See RunPod vs Vast.ai for the side-by-side, and our egress reference for transfer fees.
Salad is the other peer-to-peer marketplace we track. In our data every Salad machine is a P2P host, so the trade is the same as Community Cloud or a Vast.ai consumer host: a lower price for work that can restart and data that is not sensitive.
From Community Cloud, a marketplace is a sideways move. From Secure Cloud, it is a step down in assurance unless you filter Vast.ai to its datacenter listings.
More dependable capacity or bare metal: neoclouds
These are clouds in the ordinary sense: you rent from the company, not from an individual host. Here is what our research could verify about each, as of 21 September 2026.
| Provider | What we can verify |
|---|---|
| Lambda | Founded 2012. Instances are "self-serve, first-come access". No charge for ingress or egress. Billed per minute, and an instance can only be launched, restarted or terminated, so there is no stop button. No spot tier on its price list. Details are in our Lambda alternatives article. |
| Hyperstack | Run by NexGen Cloud, which positions itself on European sovereign AI (its Series A announcement, 8 April 2025). |
| Vultr | A general cloud host that also rents whole bare metal GPU servers with a transfer allowance included, per its public plan API. |
| Verda, Massed Compute, Ori | No verified company fact in our research. Their live prices are in the table above. |
| CoreWeave | Its 2025 annual report describes committed contracts as typically "take-or-pay", with initial periods from one to six years. Built for large buyers. |
Two cautions. "More dependable" does not mean "always available": Lambda's first-come model means an instance type can be sold out, and no Lambda document says how often. And for one GPU, the live price and stock matter more than any rating. SemiAnalysis rates GPU providers for cluster workloads in its ClusterMAX system, which says little about renting a single GPU. Our neocloud explainer covers it.
Start with Lambda vs RunPod or Hyperstack vs RunPod. For multi-node jobs, see GPU cluster pricing.
Stay serverless
If you are on RunPod Serverless, compare billing rules before rates. Providers differ on whether cold-start time is charged: RunPod bills it, Beam says it does not charge for server spin-up or image load, and Replicate bills setup and idle time only on private models and deployments, all per their docs on 21 September 2026. Vast.ai also has a serverless product, priced "at the same price as Vast.ai's non-Serverless GPU instances". Our serverless GPU pricing comparison puts the published per-second rates side by side.
Some names you will see elsewhere are not in our live data, so they get no prices here: the serverless platforms Modal, Replicate, Baseten, Cerebrium, Beam and fal, and the neoclouds Nebius and Together AI.
What you lose by leaving
Templates. If you start pods from a RunPod template, write down what it gives you before you go: the image, the ports, the start command and the volume paths. Elsewhere you will rebuild that as a plain container image, or as a setup script on a virtual machine.
Serverless next to pods. RunPod sells pods, endpoints and clusters of up to 64 GPUs on one account. Lambda, for example, lists instances and clusters and no serverless product. Leaving can mean two vendors where you had one.
Community. RunPod said on 20 January 2026 that it had "1M+ developers". A large user base usually means more public write-ups of common errors. Expect fewer on a smaller provider.
Billing habits. Per-second billing with free transfer is not universal. Lambda bills by the minute until you terminate. On Vast.ai, each host sets a per-byte transfer rate.
The decision rule
- Price is the only problem and the work can restart: move to Vast.ai, filter by reliability score, and checkpoint often.
- Price is the problem but the data is sensitive: stay on Secure Cloud or use Vast.ai's Secure Cloud filter. Compare the two in the live table first.
- Your GPU keeps disappearing on restart: before leaving, stop treating stop and resume as a guarantee. Keep data on a network volume and deploy fresh pods. If that still fails, move to a provider in the live table that has your GPU in stock today.
- You need whole servers or no neighbours: a bare metal host such as Vultr, or a large neocloud under contract.
- You run production on RunPod Serverless: stay unless the cost per request, with cold starts counted, is clearly lower elsewhere. Rates alone do not settle it.
- You rent one GPU for a few hours at a time and nothing above applies: there is no reason to leave. Check the live table now and then.
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 unless a date is given.
- RunPod pricing
- RunPod docs: Pods overview
- RunPod docs: Choose a Pod
- RunPod docs: Pod pricing
- RunPod docs: Serverless pricing
- RunPod docs: Endpoint configurations
- RunPod docs: Zero GPUs troubleshooting
- RunPod docs: Network volumes
- RunPod status page
- RunPod press release, 20 January 2026
- Vast.ai docs: General FAQ
- Vast.ai docs: Security FAQ
- Vast.ai docs: Pricing
- Vast.ai docs: Rental types
- Vast.ai docs: Instance types
- Vast.ai docs: Serverless pricing
- Lambda pricing
- Lambda docs: Billing
- Lambda docs: Creating and managing instances
- Lambda blog: agreement with Microsoft, 3 November 2025 (founding year)
- NexGen Cloud: Series A announcement, 8 April 2025
- Vultr bare metal plan API
- CoreWeave annual report for 2025 (10-K), filed 2 March 2026
- SemiAnalysis ClusterMAX provider listing
- Beam pricing
- Replicate docs: Billing