providersguides

GPU Cloud Providers: A Neutral Map With Live Prices

Who rents cloud GPUs, grouped by kind and listed alphabetically, with a live price table, verified facts per provider and five questions to shortlist three.

By Faiz Ahmed•
•10 min read

There is no single best GPU cloud provider, and this page does not pretend to pick one. What you can do is see every tracked provider's current price side by side, understand which of three kinds each one is, and use five questions to cut the list to three worth a trial. The live table comes first, because it answers the question most people arrive with.

Every tracked provider with a current offer

This table shows every provider we track that has a current in-stock, on-demand offer for at least one of these GPUs. It is sorted by the price of the first column. Every price is per GPU per hour.

ProviderRTX 4090 $/GPU-hrA100 $/GPU-hrH100 $/GPU-hrH200 $/GPU-hrB200 $/GPU-hr
Vast.ai$0.45none in stock$2.95$4.68none in stock
RunPod$0.74$1.59$3.19$4.59$6.79
LeaderGPU$1.04$0.68none in stock$6.22none in stock
Massed Computenone in stock$1.35$2.73none in stocknone in stock
QuantaCloudnone in stock$1.48$2.59none in stocknone in stock
Lyceumnone in stock$1.59$2.79$4.29none in stock
VERDAnone in stock$1.74$3.37$4.41none in stock
Lambda Labsnone in stock$1.99$4.19none in stocknone in stock
Vultrnone in stock$2.80none in stocknone in stocknone in stock
Paperspacenone in stock$3.09$5.95none in stocknone in stock
Hyperstacknone in stocknone in stock$2.50none in stocknone in stock
Orinone in stocknone in stock$2.90$3.50none in stock
Scalewaynone in stocknone in stock$3.29none in stocknone in stock
DigitalOceannone in stocknone in stock$4.41none in stocknone in stock
Packet.ainone in stocknone in stocknone in stocknone in stock$3.75
Cheapest in-stock on-demand price per GPU-hour at each provider, from providers with live stock tracking. Latest stock observation: .

"Current" means an in-stock, on-demand offer seen in the last 15 minutes, from a provider whose stock we track live. Stock is rechecked about every minute.

If a provider is missing from the table, it means one of three things:

  1. We saw no in-stock, on-demand offer from it for these GPUs in the last 15 minutes. It may be sold out, or it may not sell these GPUs.
  2. We do not have live stock tracking for that provider. We may still list its prices on its provider page.
  3. Its offers are peer-to-peer. The live tables exclude peer-to-peer hosts, so Salad does not appear, and Vast.ai's column shows only its verified datacenter tier.

An empty cell means the same for that one GPU. Absence is not a verdict on quality, and presence is not an endorsement. One disclosure: QuantaCloud is run by the owner of this site. It is treated like any other row.

The sort order is price and nothing else. We do not rank providers, and the lists further down are alphabetical.

The three kinds of provider

Hyperscalers. These are the general-purpose clouds. The SemiAnalysis playbook of October 2024 lists Google Cloud, Azure, AWS and Oracle among them. GPUs are one product among many, and the discounts come from one-year or three-year commitments.

Neoclouds, or GPU clouds. Uptime Institute described them in February 2025 as providers "specializing in AI infrastructure as a service", which "provide only a handful of product lines". In Uptime's words they focus on "GPU-backed servers and virtual machines".

Marketplaces. A marketplace matches renters with hardware it does not run itself. SemiAnalysis treats brokers, platforms and aggregators as a separate category that does not own GPUs, and Vast.ai's documentation says "hosts set their own prices".

Our neocloud explainer goes deeper on how the three differ in price, contracts and risk. The rest of this page is about the providers we track.

Provider notes

Each note is limited to facts we could verify from the provider's own pages or filings on 21 September 2026, unless another date is given. Providers we have not researched yet are listed at the end of the section, with no claims attached. That is not a judgement on them.

Marketplaces

  • RunPod Community Cloud. RunPod's docs describe Community Cloud as connecting "individual compute providers to users through a vetted, secure peer-to-peer system", and rate its reliability as "Variable". The same docs say RunPod "is no longer accepting new hosts for Community Cloud", though existing resources remain available. RunPod's first-party tier, Secure Cloud, is listed under GPU clouds below.
  • Salad. A peer-to-peer network. Because the live tables exclude peer-to-peer offers, Salad never appears in the table above. See its provider page for what we list.
  • Vast.ai. Its docs call it "a cloud computing, matchmaking, and aggregation service" where "hosts set their own prices". Billing is by the second. Hosts range from individuals to datacenters: the docs say "Individual hosts may have less formal security measures", and a "Secure Cloud" filter limits results to vetted datacenter partners. Data transfer costs "vary by host". Storage keeps billing while an instance is stopped.

GPU clouds

  • CoreWeave. Listed on Nasdaq as CRWV since 28 March 2025. Its annual report for 2025 describes committed contracts that are typically take-or-pay, with initial terms that generally range from one to six years.
  • Crusoe. Named by the SemiAnalysis playbook of October 2024 as one of four "AI Neocloud Giants", with CoreWeave, Lambda and Nebius.
  • DigitalOcean. A general cloud host that Uptime Institute named in February 2025 among established hosts now selling GPUs. Paperspace's pricing is published inside DigitalOcean's product documentation.
  • Hyperstack. The cloud brand of NexGen Cloud, which in April 2025 framed its funding round around "sovereign AI capabilities in Europe".
  • Lambda. Founded in 2012. It sells three lines: instances, 1-Click Clusters and Superclusters. Instances are "billed in one-minute increments", cannot be stopped (only launched, restarted or terminated), and carry no ingress or egress charge. 1-Click Clusters start at 16 GPUs for two weeks.
  • Latitude.sh. A bare metal provider. It bills hourly, monthly or yearly, and every plan on its price page included "20 TB free out / ∞ in".
  • Paperspace. Documented under DigitalOcean's product docs. Its Gradient notebooks still documented a free plan, with a free M4000 GPU machine, on a page DigitalOcean last verified on 13 July 2026. We have not confirmed that free capacity is obtainable in practice.
  • RunPod Secure Cloud. RunPod sells "Pods for dedicated GPU instances, Serverless for API inference, and Clusters for multi-node jobs". Secure Cloud operates "in T3/T4 data centers". Pods are "billed by the second for compute and storage, with no fees for data ingress or egress". Savings plans take a 3-month or 6-month term paid upfront.
  • Scaleway. A French cloud provider. We hold no verified billing facts for it.
  • Voltage Park. Merged with Lightning AI in a deal completed in January 2026. The combined company trades as Lightning AI.
  • Vultr. A general cloud host that also sells GPUs. Its public plan API listed fractional GPU virtual machines, whole-GPU virtual machines and bare metal GPU servers, each bare metal GPU server with 15,360 GB of transfer included.

Tracked, not yet researched

We track prices for these providers, but our research does not yet cover them, so we state no facts and assign no category. In alphabetical order: Atlantic.Net, Cirrascale, Denvr Dataworks, Hot Aisle, LeaderGPU, Lyceum, Massed Compute, Ori, PacketAI, QuantaCloud and Verda. QuantaCloud is run by the owner of this site and is treated like any other row. For all of them, the live table and the provider page show what our own data holds.

Hyperscalers are covered in the neocloud explainer, not here. For what each provider charges to move data out, use the data egress reference.

How to shortlist

Answer five questions in order. Each one removes rows from the table.

1. Which GPU, and how many

The GPU sets which columns you read, and the count sets which providers can serve you at all. A per-GPU price does not tell you whether a provider rents that GPU singly, in fours or only in eights. Check the provider's page for the instance shapes. If you need more than one node, you are buying a different product: see GPU cluster pricing. As one verified example of the jump, Lambda's cluster product starts at 16 GPUs for two weeks.

2. Which region

Latency to your users and data residency rules both narrow the list. Our regions page lists providers by region. Do this before you compare prices, because the lowest row is no use if it sits on the wrong continent.

3. For how long

A few hours is an on-demand purchase. A few months is a commitment question, and the terms differ a lot: RunPod's savings plans run 3 or 6 months, hyperscaler commitments run one or three years, and CoreWeave's contracts run one to six years. Our guide to on-demand, spot and reserved pricing sets out what each model costs you in flexibility.

4. How sensitive the data is

On a marketplace, the machine may belong to an individual. Vast.ai's docs are plain about it: "Individual hosts may have less formal security measures", and "Verified" is an automated check of reliability and configuration, not a security audit. For customer data or proprietary weights, stay with providers that run their own datacenter capacity, or use a marketplace's datacenter-only filter. Our Vast.ai alternatives guide covers this trade in detail.

5. What happens when it is out of stock

Plan for it. SemiAnalysis reported in April 2026 that "On-Demand GPU rental capacity is sold out across all GPU types". Lambda describes its instances as "self-serve, first-come access". RunPod's docs warn that if another user rents your GPU while your pod is stopped, the pod "cannot start with a GPU" until it frees up. The defence is a second provider you have already signed up with and tested your image on.

The rule

Shortlist three, not one.

  1. The lowest current price in the table that survives all five questions.
  2. A fallback from a different provider for the same GPU, for the day the first is out of stock.
  3. One provider that sells the commitment or cluster product you expect to need in six months, so that growing does not mean starting the search again.

Run the same small job on each. Compare what you were billed, not what the price page said, then move the real work to the winner and keep the fallback account open. The full provider list has every provider we track, including those without a current offer in the table above.

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.

Frequently asked questions

Who are the main GPU cloud providers?▾

They fall into three kinds: hyperscalers such as AWS, Azure and Google Cloud, GPU-first clouds often called neoclouds, and marketplaces that match renters with other people's hardware. This page lists the providers we track in each kind, alphabetically.

Which GPU cloud provider is the cheapest?▾

It changes through the day and differs by GPU, which is why we do not name one. The live table on this page is sorted by current price, so the top row is the answer at the moment you load it.

Why is a provider missing from the live price table?▾

There are three possible reasons. We saw no in-stock, on-demand offer from it in the last 15 minutes, or we do not have live stock tracking for it, or its offers are peer-to-peer, which the live tables exclude.

What is the difference between a GPU cloud and a GPU marketplace?▾

A GPU cloud rents capacity it operates itself. A marketplace matches you with hardware run by independent hosts, who on Vast.ai set their own prices.

Does this site rank GPU cloud providers?▾

No. Providers are grouped by kind and listed alphabetically, and the live table is sorted by price alone. QuantaCloud is run by the owner of this site and is treated like any other row.

How many GPU cloud providers should I try?▾

Shortlist three: the lowest current price that passes your checks, a second provider as a fallback for when the first is out of stock, and one that sells the commitment or cluster product you may need later.

TAGGED WITH

#cloud-gpu#neocloud#cost-comparison#runpod#vast-ai#lambda

Related Posts

What a Neocloud Is and When to Rent GPUs From One

A neocloud is a GPU-first cloud. See how neoclouds differ from hyperscalers and GPU marketplaces, who they are, the risks, and which one to rent from.

Read more

RunPod Alternatives, Sorted by Why You Are Leaving

Live prices for the same GPU on RunPod and comparable providers, the documented reasons people leave RunPod, and which alternative fits each reason.

Read more

Vast.ai Alternatives: What the Step Up to Managed Costs

When Vast.ai is safe enough, when it is not, and the live price gap between the marketplace and a managed cloud for the same GPU. Ends with a host checklist.

Read more