GPU cloud providers rent graphics processors by the hour, month or token.
This guide ranks thirteen of them by what a first month costs once storage and idle time are added, by how much of the price is printed, and by whether you can start without a salesperson, with independent vendors placed ahead of larger groups.
Vendors can pay for more visibility on this website, for example a higher place in a ranking or more room on a page. Commercial relationships can also affect which products we feature and how prominently. Every vendor link is marked the same, whatever the arrangement behind it. How that works.
In short
What GPU cloud providers software does
A GPU cloud rents graphics processors on demand, as hourly instances, dedicated servers or per-token endpoints, so teams can train and serve models without buying hardware.
To rank GPU cloud providers tools we asked, in turn, how much work setup takes, what it truly costs, whether data can be taken back out, whether leaving is possible, and who the product suits. Vendors outside the best-known names fill places 1 to 3 when they exist, and the big brands rank from place 4. We explain on how we work why the tests are these five and why no product receives a score.
Published per-second rates by the GPU-hour, with separate Secure Cloud, Community Cloud and serverless price lists; clusters up to 64 GPUs on pay as you go
—
Engineers who want short per-second GPU sessions at several price levels
Teams needing a stated jurisdiction or a fixed monthly bill
Published USD rates per GPU-hour metered by the minute, spot rates on some cards, per-million-token billing for open models, sales contact from 25 GPUs
—
Solo developers and small teams renting one or two GPUs by the minute
Companies that need a named legal entity before signing
On-demand per-second billing, interruptible rentals and reserved terms of one to six months through sales; a calculator estimates cost by GPU and hours
—
Cost-driven users whose jobs can stop and resume from checkpoints
Anyone needing a price list to paste into a budget
Per-second GPU pods with a cheaper community tier and serverless endpoints
Ranked #1 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
The H100 SXM card is listed at 3.99 dollars per hour on Secure Cloud and 2.69 on Community Cloud, so the same hardware class has two price points. Serverless is priced by VRAM tier, and Public Endpoints charge per use for about 45 models.
Reserved clusters and enterprise contracts go through sales. Container disk is 0.10 per GB-month, so long-lived environments need a storage line in the budget.
What stands out
Per-second billing
Community Cloud tier
Serverless by VRAM
Where it costs you
Volume disks cost twice as much while idle as while running.
No country or jurisdiction is stated on the pages read.
Right for
Engineers who want short per-second GPU sessions at several price levels
Wrong for
Teams needing a stated jurisdiction or a fixed monthly bill
Not stated by the vendorPublished per-second rates by the GPU-hour, with separate Secure Cloud, Community Cloud and serverless price lists; clusters up to 64 GPUs on pay as you go
Per-minute GPU rentals plus a token API for open models
Ranked #2 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
An H200 SXM costs 4.59 dollars per hour on demand and 2.29 on spot, and an H100 SXM 3.49 against 1.54. Cheaper cards run from the A100 80 GB at 1.49 down to the L4 at 0.44.
CPU virtual machines are on demand only, storage is 0.10 per GB-month and in-region egress is not charged. The pricing page says no commitments apply, which keeps a trial run cheap and lets you stop paying whenever the job ends.
What stands out
Per-minute billing
Spot GPU rates
Token Factory API
Where it costs you
Savings plans of a month or longer need a sales conversation.
Legal entity and country are not stated on the pages read.
Right for
Solo developers and small teams renting one or two GPUs by the minute
Wrong for
Companies that need a named legal entity before signing
Not stated by the vendorPublished USD rates per GPU-hour metered by the minute, spot rates on some cards, per-million-token billing for open models, sales contact from 25 GPUs
Marketplace GPU servers from independent hosts, starting with a small deposit
Ranked #3 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
Listed rates start at 2.25 dollars per hour for an H100 SXM5, 1.80 for an A100 SXM4 and 0.35 for an RTX 4090, with consumer GPUs from 0.12 and CPU cloud from 0.012. The home page mentions KVM-based root access and Windows machines.
Servers sit across more than 100 locations, so latency to your users depends on the listing chosen. A 5 dollar deposit opens the account, and enterprise needs go through sales.
What stands out
Host marketplace
KVM root access
Consumer GPUs
Where it costs you
Instant VMs carry a beta label on the home page.
No company country or legal entity is stated.
Right for
Hobbyists and researchers who want the lowest listed hourly rate on consumer cards
Wrong for
Buyers who need one named operator running every server
Not stated by the vendorPublished hourly rates by GPU model, pay as you go from a minimum deposit, sales contact for reserved or long-term rentals
UK-headquartered GPU cloud with per-minute billing and reservation discounts
Ranked #4 of 13 in Best GPU Cloud Providers in 2026.
Published pricingEurope
On demand, an H200 SXM is 3.99 dollars per hour, an H100 SXM 3.20 and an A100 80 GB 1.35, all billed per minute. Reservations start at 2.79 for the H200 and 2.72 for the H100, and spot H100 PCIe is 2.00.
AI Studio sells Llama and gpt-oss models by the million tokens. The Kubernetes master node is not charged, though worker GPUs are, and reservations arrive on a monthly invoice.
What stands out
Per-minute billing
Spot and reserved VMs
AI Studio tokens
Where it costs you
Pay as you go needs a prepaid balance paid by card.
The vendor is a flagship product of NexGen Cloud, not an independent company.
Right for
Teams in the United Kingdom wanting listed rates and reservation discounts
Wrong for
Buyers who insist on a single-product, standalone vendor
United KingdomPublished minute-metered GPU-hour rates, spot VMs, discounted reservations invoiced monthly, prepaid card balance for pay as you go
Three billing modes, from interruptible rentals to reserved terms
Ranked #5 of 13 in Best GPU Cloud Providers in 2026.
Pricing on requestElsewhere
The discount structure is the draw: interruptible rentals are called 50% or more cheaper than standard, and reserved terms go up to 50% off. The trade is predictability, since reclaimed instances end a job and reserved terms are arranged through sales.
The page gives no per-card rate in its text, so the calculator, which takes a GPU model, a count and hours, is the only way to see a figure before renting.
What stands out
Interruptible rentals
Per-second billing
Cost calculator
Where it costs you
Interruptible instances can be reclaimed by the platform.
The fetched pricing page shows no numeric prices, only a calculator.
Right for
Cost-driven users whose jobs can stop and resume from checkpoints
Wrong for
Anyone needing a price list to paste into a budget
Not stated by the vendorOn-demand per-second billing, interruptible rentals and reserved terms of one to six months through sales; a calculator estimates cost by GPU and hours
Bare metal and virtual machines with GPU plans billed hourly or monthly
Ranked #6 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
A single H100 80 GB is 1.68 dollars per hour or 1,230 per month, and eight RTX PRO 6000 cards are 24 per hour. Bare metal runs from 0.41 per hour for 6 cores to 5.44 for 64 cores, and virtual machines start at 0.02.
Unlisted GPU plans are requested through a Wishlist form, and bespoke bare metal goes through sales, so the catalogue is narrower than the largest clouds.
What stands out
Bare metal GPU
Hourly or monthly
Bandwidth included
Where it costs you
No legal entity or country is named anywhere on the page.
Storage and networking prices sit on other pages that were not read.
Right for
Teams that want a whole GPU server and one bill for GPUs and CPUs
Wrong for
Buyers who need a vendor entity named on the page
Not stated by the vendorPublished hourly and monthly rates in dollars, reserved discounts on monthly and yearly billing, a wishlist form for unlisted GPUs
Dedicated GPU servers in Germany and Finland at fixed monthly rates
Ranked #7 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
The GEX45 inference server is 214 euro per month, or 0.3429 per hour, plus a 209 setup fee. The GEX63 is 999 per month, the GEX131 starts at 1,499 and the GEX430 is 3,599, all excluding VAT.
Each plan links straight to configuration and no sales route is mentioned, which keeps costs readable but offers no cluster tier on the page read.
What stands out
Dedicated servers
Monthly pricing
German and Finnish sites
Where it costs you
The GEX45 entry model carries a 209 euro setup charge.
No trial is mentioned, and the catalogue is only four server types.
Right for
Teams running steady inference who want a flat monthly server bill
Wrong for
Researchers who need to scale from one GPU to many
Not stated by the vendorPublished monthly prices in euro excluding VAT, an hourly rate on the entry model, a one-time setup fee
Per-GPU-hour instances and reserved clusters from a San Jose company
Ranked #8 of 13 in Best GPU Cloud Providers in 2026.
Published pricingNorth America
One-click clusters last from two weeks to a year, and the rate falls as the cluster grows: HGX B200 goes from 9.86 dollars at 16 GPUs to 8.87 at 256 or more.
Self-serve instances list B200 at 7.69, H100 at 4.59 and A10 at 1.29. Prices exclude sales tax, VAT and GST, so a European or Indian invoice will total more than the card shows.
What stands out
1-Click Clusters
HGX B200 and H100
Reserved Superclusters
Where it costs you
Superclusters show no price and need a sales conversation.
Self-serve instance prices change in the middle of November.
Right for
Machine-learning teams wanting a short cluster rental without a long contract
Wrong for
Anyone who needs a trial credit before spending
United StatesPublished rates by the GPU-hour for instances and 1-Click Clusters, with Superclusters priced through sales
Self-serve GPU deployments and token-priced inference with scheduled price rises
Ranked #9 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
B200 moves from 9.65 to 10.20 dollars per GPU-hour on 1 December, H100 from 5.50 to 6.30, and B300 becomes self-serve at 13.80. Serverless Inference runs per million tokens, for example gpt-oss 120B at 0.05 input and 0.25 output.
Managed Kubernetes is 0.10 per cluster-hour, persistent disks 0.08 per GiB-month. Provisioned Throughput, Tailored Deployments and committed capacity all go through sales rather than a price list.
What stands out
Serverless inference
Managed Kubernetes
Dec 2026 price change
Where it costs you
Every self-serve GPU rate rises from 1 December 2026.
Managed Kubernetes adds a per-cluster-hour charge.
Right for
Teams that want GPUs and token-priced inference from one account
Wrong for
Buyers who need a fixed price beyond next month
Not stated by the vendorPublished per-GPU-hour rates now and from 1 December, per-token serverless inference, committed capacity through sales
On-demand and spot GPU instances billed per GPU-hour, no network charges
Ranked #10 of 13 in Best GPU Cloud Providers in 2026.
Published pricingElsewhere
On demand, an H100 is 4.50 dollars per GPU-hour and a B200 8.50, while preemptible rates begin at 0.79 for H100 and H200. Object Storage Standard is 0.0147 per GiB-month and Managed Kubernetes costs nothing.
GB200 and GB300 NVL72 racks are sales-only. Companies can pay by card or bank transfer, individuals by card, all processed through Stripe, and the pages read give no trial.
What stands out
Preemptible GPUs
H100 to B300
Managed Kubernetes
Where it costs you
Spot prices can change as often as every 15 minutes.
Accounts need a minimum opening payment.
Right for
Buyers who want a wide GPU range from H100 through B300
Wrong for
Anyone who needs to try the platform before paying
Not stated by the vendorPublished on-demand and preemptible rates per GPU-hour plus storage; top NVL72 racks through sales; minimum first payment applies
Notebook platform with a hobby plan, now sold under DigitalOcean
Ranked #11 of 13 in Best GPU Cloud Providers in 2026.
Free tierPublished pricingElsewhere
Gradient plans run from a free tier to a 39 dollar team plan, with contact-sales levels above. Separately, a dedicated H100 is listed at 3.09 per hour, an A100-80G at 1.15 and an A6000 at 1.89, with lower H100 and A100 rates under a three-year commitment.
Flat monthly A100 and H100 plans are 1,994 dollars. The Pro price appears as both 8 and 12 on the page.
What stands out
Gradient notebooks
DigitalOcean trading name
Flat monthly GPUs
Where it costs you
The page says refunds are not offered.
The free plan limits users to public projects and a 12-hour auto-shutdown.
Right for
Students and individuals starting in notebooks before renting a dedicated GPU
Wrong for
Teams wanting a standalone vendor with a stated country
Not stated by the vendorFree Gradient plan plus paid monthly plans, published hourly dedicated GPU rates and flat monthly GPU plans
Large-cluster GPU compute priced per hour with committed-use discounts
Ranked #12 of 13 in Best GPU Cloud Providers in 2026.
Published pricingNorth America
HGX H100 is listed at 49.24 dollars per hour on demand and 19.71 on spot, with B200 at 68.80. Committed usage can reach 60% off. AI Object Storage is 0.06 per GB-month hot, and egress and data transfer are not charged.
Direct Connect 10G is 1,250 per month. Several large GPUs are sales-only, so the shopping basket of an ordinary team may include a quote.
What stands out
HGX node pricing
Spot GPU rates
AI Object Storage
Where it costs you
The newest GPUs and Reserved Compute Capacity are contact-sales only.
Public IP addresses add 4 dollars each per month.
Right for
Large training teams ready to commit usage for a discount
Wrong for
Individuals wanting one cheap GPU for a weekend
United StatesPublished on-demand and spot rates by the hour, committed usage discounts, several newest GPUs through sales
Token-billed AI model endpoints from a French public cloud
Ranked #13 of 13 in Best GPU Cloud Providers in 2026.
Pricing on requestEurope
Base mode bills input and output tokens separately; Batch mode, still beta, is half of that rate. A Playground and the Model Catalogue let you test models through a sandbox or the API without paying.
A trial credit of 200 dollars covers one first project for a month, needs a saved payment method, and cannot be refunded or transferred to another project.
What stands out
Per-token endpoints
Batch mode beta
Model playground
Where it costs you
Token amounts are only reachable through the model catalogue.
Dedicated clusters are offered through a contact route.
Right for
Teams that buy model inference by the token and want a French provider
Wrong for
Training teams that need to rent raw GPU hours
FranceBilled per input and output token, batch mode at half the base rate, dedicated clusters through sales, one-month trial credit
A GPU cloud rents graphics processors on demand, as hourly instances, dedicated servers or per-token endpoints, so teams can train and serve models without buying hardware. These are the questions that actually differ between vendors, in the order we would ask them.
01
Decide whether you need a published price
The split is 11 published against 2 quoted. If you want to compare numbers first, start with RunPod, Jarvislabs, TensorDock, Hyperstack, Latitude.sh, Hetzner GPU, Lambda, Crusoe, Nebius, Paperspace, CoreWeave.
02
Look at where the vendor sits
The 13 companies behind this list are established in 4 countries, by region Elsewhere 9, North America 2, Europe 2. Disclosure obligations attach to the company, so they travel with the vendor even when the servers stay at home.
Match GPU cloud providers to the job before comparing hourly rates
A training run, an inference service and a notebook session need different things from a GPU cloud. Training wants many cards joined by fast networking, which is why Lambda sells 1-Click Clusters and CoreWeave sells HGX nodes. Inference wants steady capacity at a low monthly figure, which is where Hetzner GPU and its GEX45 server fit.
Experiments want per-minute or per-second billing, which Hyperstack, Jarvislabs and RunPod offer. OVHcloud AI sells model access by the token, which is a different purchase from renting a card. Write the job down first, then compare only the providers whose product shape matches it, because an hourly rate for the wrong shape tells you little.
Name the job as training, inference or experiments before opening any pricing page.
Check whether the offer is a card, a whole server or a token endpoint.
Note the largest cluster size you can reach without talking to sales.
Read the GPU-hour price the way each pricing page prints it
GPU-hour figures are not directly comparable. CoreWeave prints rates far above single-card rivals, so confirm whether a row is one card or a node before comparing it with RunPod or Nebius. Lambda says its self-serve rates change on 16 Nov 2026, and Crusoe lists current rates beside higher ones from 1 December, so a budget built today can be wrong next month.
Lambda also excludes sales tax, VAT and GST, and Hetzner GPU excludes VAT, so the invoice total differs from the card. Vast.ai shows no numeric rates in its page text at all. Put each candidate in one table with unit, currency, tax treatment and the date the price takes effect.
Confirm whether the price is per card or per node.
Look for a stated date on which the rate changes.
Add tax to the rate before comparing invoices.
Add storage, addresses and idle time before choosing a cheaper rate
The hourly rate is only part of the bill. RunPod charges idle volume disks at twice the running rate, so a stopped project still costs. CoreWeave charges 4.00 dollars per month for each public IP and 1,250 for a 10G Direct Connect, while it lists egress as free.
Nebius, Hyperstack and Crusoe do not charge for network traffic at present, though Crusoe words that as true only at this time. TensorDock asks for a small deposit, and Nebius requires a first payment of at least 25 dollars. Paperspace states that refunds are not offered. Add a month of storage and one idle week to each estimate before you decide which provider is cheaper.
Price one idle week of attached storage for each candidate.
Check whether network traffic is billed now, and whether that is temporary.
Read the refund and minimum-payment lines before depositing.
Choose between marketplace capacity and dedicated servers deliberately
Cheap GPU capacity comes in two forms, and they fail differently. Vast.ai offers interruptible rentals that can be reclaimed, and TensorDock rents from independent hosts whose hardware and location differ by listing, so both suit work that checkpoints and restarts. Hetzner GPU and Latitude.sh sell whole servers by the month, which suits steady inference but charges you for idle hours.
Hyperstack and Lambda sit between them: Hyperstack adds spot and reserved rates beside on-demand ones, and Lambda adds reserved clusters. Match the failure you can tolerate to the product: a reclaimed instance costs a restart, while an unused dedicated server costs a monthly fee, so pick the form whose worst case your team can afford.
Ask whether your job can stop at a checkpoint and resume.
Compare a month of dedicated server cost with your expected busy hours.
Check who operates the hardware: the vendor or an independent host.
What goes wrong most often when buying GPU cloud providers software
Comparing an on-demand rate for one provider with a reserved or spot rate for another, as if they described the same capacity.
Ignoring dated price changes such as Lambda's 16 Nov 2026 rates or Crusoe's 1 December rise when building a yearly budget.
Leaving volumes attached after a run, so RunPod-style idle storage keeps charging while no GPU is working.
Assuming an AI endpoint like OVHcloud AI can give you a raw GPU to train on, when its dedicated clusters sit behind a contact link.
07
Frequently asked questions
6 answers
What is the best GPU cloud providers in 2026?
RunPod leads our ranking of 13. Runpod Inc. appears in structured data on the pricing page, which names no country. Pods are billed per second with no commitment, and a Community Cloud tier sits below the Secure Cloud rate for the same GPU families.
Idle volume disks cost twice the running rate, and nothing on the page offers a trial or credit, so testing is paid for from the first minute.
Which GPU cloud providers tools publish their pricing?
Published pricing exists for 11 products on this page:
RunPod: Published per-second rates by the GPU-hour, with separate Secure Cloud, Community Cloud and serverless price lists; clusters up to 64 GPUs on pay as you go.
Jarvislabs: Published USD rates per GPU-hour metered by the minute, spot rates on some cards, per-million-token billing for open models, sales contact from 25 GPUs.
TensorDock: Published hourly rates by GPU model, pay as you go from a minimum deposit, sales contact for reserved or long-term rentals.
Hyperstack: Published minute-metered GPU-hour rates, spot VMs, discounted reservations invoiced monthly, prepaid card balance for pay as you go.
Latitude.sh: Published hourly and monthly rates in dollars, reserved discounts on monthly and yearly billing, a wishlist form for unlisted GPUs.
Hetzner GPU: Published monthly prices in euro excluding VAT, an hourly rate on the entry model, a one-time setup fee.
Lambda: Published rates by the GPU-hour for instances and 1-Click Clusters, with Superclusters priced through sales.
Crusoe: Published per-GPU-hour rates now and from 1 December, per-token serverless inference, committed capacity through sales.
Nebius: Published on-demand and preemptible rates per GPU-hour plus storage; top NVL72 racks through sales; minimum first payment applies.
Paperspace: Free Gradient plan plus paid monthly plans, published hourly dedicated GPU rates and flat monthly GPU plans.
CoreWeave: Published on-demand and spot rates by the hour, committed usage discounts, several newest GPUs through sales.
Everything else on the list is priced on request.
Is there a free GPU cloud providers tool?
Paperspace offer a free tier or a free self-hosted edition.
Where are these GPU cloud providers vendors established?
Vendor countries on this page, Elsewhere 9, North America 2, Europe 2:
RunPod: Not stated by the vendor.
Jarvislabs: Not stated by the vendor.
TensorDock: Not stated by the vendor.
Hyperstack: United Kingdom.
Vast.ai: Not stated by the vendor.
Latitude.sh: Not stated by the vendor.
Hetzner GPU: Not stated by the vendor.
Lambda: United States.
Crusoe: Not stated by the vendor.
Nebius: Not stated by the vendor.
Paperspace: Not stated by the vendor.
CoreWeave: United States.
OVHcloud AI: France.
What should you use instead of RunPod?
The next in line are Jarvislabs, aimed at Solo developers and small teams renting one or two GPUs by the minute, and TensorDock, aimed at Hobbyists and researchers who want the lowest listed hourly rate on consumer cards.
Who should not buy RunPod?
Teams needing a stated jurisdiction or a fixed monthly bill. Volume disks cost twice as much while idle as while running..
Makers of GPU cloud providers products can ask us to look at theirs. Say what it does and who uses it. We choose what to add, and nothing can be bought; see software advice for what a listing means.
