Software advice
Best AI Cloud GPU Providers in 2026
An AI cloud GPU provider is a company that rents graphics processors by the hour, minute or second for training, fine-tuning and serving machine-learning models, through virtual machines, bare-metal nodes, clusters or serverless endpoints. It is not a model API that charges per token, and it is not a general-purpose cloud bought mainly for storage, databases or web hosting.
Quick answer
Hyperstack ranks first for public per-minute rates on H100 through B300. RunPod is the pick for pods plus serverless endpoints on one account. Modal suits Python teams that want per-second serverless GPUs with no idle cost.
Renting GPUs is a different purchase from buying a model API. You choose the chip, the interconnect, the region and the commitment, and you carry the engineering work of keeping the job running. Prices swing by a factor of two or more for the same H100, depending on whether the rate is on-demand, spot, reserved or quoted by a sales team. This ranking compares twelve providers on what the price list actually shows, how capacity and compliance are documented, how hard it is to leave, and which kind of workload each one handles best. Prices are as of October 2026, in dollars unless the vendor lists euros.
The 12 best AI cloud GPU providers in 2026, ranked
There is no score. 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. How we rank.
| # | Tool | Best for | Free plan | Entry price |
|---|---|---|---|---|
| 1 | Hyperstack | Teams that want published on-demand H100 and A100 rates, spot capacity and an API without a sales call | No | $2.50 per hour for an H100 (on-demand) |
| 2 | RunPod | Developers who want a pod, a serverless endpoint or a multi-node cluster on the same account | No | $0.59 per hour (A40 or RTX A6000 pod, 48 GB) |
| 3 | Vast.ai | Cost-sensitive trainers and experimenters who can checkpoint work and accept variable host quality | No | Market-set hourly rates; no fixed list price |
| 4 | Modal | Python developers deploying inference, batch jobs, training or sandboxes without managing servers | Trial | $0 plus compute (Starter, with $30 monthly credit) |
| 5 | Nebius | Teams training large models that want an EU-headquartered GPU cloud with managed Kubernetes and Slurm | No | $1.80 per GPU hour (RTX PRO 6000, on-demand) |
| 6 | OVHcloud | European buyers who want GPU instances from a French public cloud with a 99.99% availability SLA | No | $1.80 per hour excl. VAT (1x L40S, l40s-90) |
| 7 | Lambda | ML engineers wanting self-serve H100 and B200 instances with a route to large 1-Click Clusters | No | $0.79 per GPU hour (Tesla V100 16 GB) |
| 8 | DigitalOcean Gradient AI GPU Droplets | Developers already on DigitalOcean who want GPU Droplets next to their existing Droplets and databases | No | $0.76 per GPU hour (RTX 4000 Ada, on-demand) |
| 9 | Crusoe Cloud | Buyers who want GPU instances plus managed inference and fine-tuning from one provider | No | Quoted |
| 10 | CoreWeave | AI labs and enterprises renting whole 8-GPU nodes and clusters for training and large-scale inference | No | $10.00 per hour (8x L40 node) |
| 11 | Verda | European teams that want self-service H100 to B300 instances with EU ownership and published certifications | No | $0.236 per hour (V100 16GB) |
| 12 | Lightning AI | Individual researchers and small teams who want GPUs inside a browser IDE with a free starting allowance | Trial | $0.55 per hour (T4 GPU) |
- 1
Hyperstack
Hyperstack ranks first because every on-demand rate, from an RTX A4000 at $0.15 to a B300 at $7.40 per GPU hour, is printed on one page and billed by the minute. Reservation and spot prices sit beside them, so a buyer can model all three before signing up. The trade-off is a smaller operator with less published enterprise assurance than the largest GPU clouds.
Strong on
- Rates for twelve GPU models are public, from an A4000 at $0.15 to a B300 at $7.40 per hour
- Reserved H100 starts at $1.75 per hour, against $2.50 on-demand
- Go and Python SDKs plus a Terraform provider are listed in the documentation menu
Weak on
- Only SOC 2 Type 2 is named on the pages opened; no ISO certificate was seen
- Spot VMs are limited to five GPU models on the pricing page
Skip it if: Buyers who need an ISO 27001 certificate on file today, or a single vendor covering storage, databases and managed model APIs. Billed per minute. Reservation, spot VMs and larger contracts are priced separately on the same page.
- 2
RunPod
RunPod takes second place because one account covers pods, serverless GPU endpoints and multi-node clusters, with a price list that starts at $0.59 per hour for a 48 GB card. It was updated on October 8, 2026 and shows B300, B200 and H200 next to older cards. Hardware comes from several partner data centers, so security coverage differs by region and product.
Strong on
- One catalog spans 48 GB cards at $0.59 and a 288 GB B300 at $8.99 per hour
- ISO/IEC 27001:2022 certificate valid from August 27, 2026, plus SOC 2 Type II and SOC 3
- Serverless, Pods, Clusters and a Hub of open-source templates share one control plane
Weak on
- Compliance coverage varies by workload, region and data center partner, per the vendor's own page
- Community Cloud and Secure Cloud tiers carry different guarantees, so check which one a price refers to
Skip it if: Regulated buyers who need one uniform compliance posture across every region without a security review of the chosen location. Pods bill per hour or per second; Serverless endpoints and Clusters carry separate meters.
- 3
Vast.ai
Vast.ai is third because it is a marketplace: prices come from supply and demand across 40+ data centers and independent hosts, with per-second billing and three tiers (on-demand, interruptible, reserved). That is where the lowest rates tend to appear, and also where quality varies by host. A Secure Cloud filter limits offers to vetted datacenters for regulated work.
Strong on
- Billing is per second with no minimum hours or rounding
- 68+ GPU types, from RTX 3060 to B200, are listed on the pricing page
- Secure Cloud filter restricts offers to trusted datacenters; SOC 2 Type 2 completed
Weak on
- Standard offers use Docker-level tenant isolation, which the vendor calls cost-effective for non-regulated workloads
- Interruptible instances can be reclaimed mid-job, so work must checkpoint and resume
Skip it if: Production inference that needs a fixed price and fixed uptime, or buyers who cannot vet individual hosts. Prices are set by hosts through supply and demand; interruptible instances run at least 50% cheaper than on-demand.
- 4
Modal
Modal ranks fourth as the serverless option: code is written in Python, containers start on demand and GPUs bill by the second, from a T4 at $0.000164 to a B300 at $0.001972 per second. Idle time costs nothing, which suits bursty inference. Steady 24-hour loads can cost more per GPU hour than a rented node, and pinning a region adds a price multiplier.
Strong on
- Per-second GPU billing with no charge for idle containers
- Starter plan includes $30 monthly credit and 3 workspace seats
- SOC 2 Type II and a GDPR data processing addendum are documented
Weak on
- Pinning a container to one region multiplies usage prices by 1.15x or 1.75x
- Sandboxes bill at 3x list price for non-preemptible rates
Skip it if: Teams that want a plain virtual machine with SSH access and a flat hourly rate, or that do not write Python. Compute is billed per second; Team is $250 per month plus compute with $100 monthly credit.
- 5
Nebius
Nebius is fifth: an Amsterdam-headquartered, Nasdaq-listed GPU cloud with on-demand HGX H100 at $4.50 and H200 at $5.40 per GPU hour, plus live preemptible pricing. Managed Kubernetes and Soperator (Slurm on Kubernetes) are free, and ingress and egress traffic carry no charge. It is aimed at cluster-scale training; the largest systems are quoted by contact.
Strong on
- Managed Kubernetes, Managed Soperator and network traffic are listed at no charge
- SOC 2 Type II with HIPAA, ISO 27001, ISO 27018 and ISO 22301 are listed
- Preemptible H100 and H200 instances start at $0.79 per GPU hour
Weak on
- GB200 and GB300 NVL72 systems are priced by contact only
- Preemptible prices change as often as every 15 minutes, which complicates budgeting
Skip it if: Solo developers who want a $0.50 card for a weekend of experiments, or anyone avoiding a console-and-quota setup. H100 is $4.50 and H200 $5.40 on-demand; preemptible rates start at $0.79 for those cards.
- 6
OVHcloud
OVHcloud is sixth: a French public cloud whose GPU line covers L4, L40S, H100 and H200 instances, with 1x H100 at $2.99 and 1x L40S at $1.80 per hour excluding VAT. The vendor states a 99.99% monthly availability SLA for GPU instances and also sells SecNumCloud-qualified infrastructure elsewhere in its range. GPU choice is narrower than at specialist providers.
Strong on
- Hourly price list includes 1x, 2x and 4x H100 flavors from $2.99 to $11.97
- 99.99% monthly availability SLA stated for every GPU instance
- Anti-DDoS protection is included with every GPU instance at no extra cost
Weak on
- Monthly-billed instances cannot move to hourly billing afterward
- Newest Blackwell parts were not shown on the GPU instance pricing table opened
Skip it if: Teams that need Blackwell-class clusters or a managed Slurm service, which OVHcloud's public cloud GPU pages did not show. A 1x H100 instance is $2.99 per hour; monthly billing cannot be switched back to hourly.
- 7
Lambda
Lambda sits seventh as an AI-only cloud founded in 2012: self-serve instances run from a V100 at $0.79 to a B200 at $6.69 per GPU hour, and 1-Click Clusters scale from 16 to 2,000+ B200 or H100 GPUs. Reserved capacity is priced by contact. Self-serve availability is first-come, so a specific configuration may be sold out.
Strong on
- Instance and cluster rates are public, with H100 SXM clusters from $5.54 to $6.16 per GPU hour
- SOC 2 Type II plus ISO 27001, 27017, 27701 and 22301 are stated
- Lambda Stack and 1-Click Clusters arrive preconfigured with ML software
Weak on
- Self-serve capacity is first-come, so a given instance type may not be available
- Lowest prices require contact for reserved terms, and prices exclude tax
Skip it if: Buyers who want serverless scaling to zero, or teams whose procurement needs one vendor for storage, databases and general cloud services. Instances list B200 at $6.69 and H100 SXM at $3.99 per GPU hour in the 8x configuration.
- 8
DigitalOcean Gradient AI GPU Droplets
DigitalOcean GPU Droplets rank eighth: a familiar Droplet console with Nvidia and AMD cards, from an RTX 4000 Ada at $0.76 to an HGX H200 at $4.47 per GPU hour. New pricing took effect August 1, 2026. AMD MI300X and MI325X are priced below the Nvidia equivalents. The catalog is smaller than at specialist GPU clouds, and powered-off Droplets keep billing.
Strong on
- AMD MI300X is $2.59 and MI325X $3.80 per GPU hour on-demand
- 12-month reserved H100 rate of $3.26 per GPU hour
- SOC 2 Type II report available through the Security Reports and Certifications Center
Weak on
- Powered-off GPU Droplets are billed until you destroy them
- Spot Droplets can be reclaimed, with at least 2 hours' notice aimed for but not guaranteed
Skip it if: Teams that need multi-thousand-GPU training clusters or the newest Blackwell Ultra capacity in self-serve form. Billed per second with a 5-minute minimum; powered-off Droplets are still billed until destroyed.
- 9
Crusoe Cloud
Crusoe Cloud is ninth: it sells GPU instances (Nvidia B200, H200, H100, A100, L40S and AMD MI300X and MI355X) alongside Managed Inference and Serverless Fine-Tuning. Its pricing page publishes only part of the GPU rate card and sends the rest to sales, so a buyer cannot model a full budget from the page alone. It names ISO 27001, ISO 42001 and SOC 2 Type I and II.
Strong on
- Managed Inference and Serverless Fine-Tuning sit on the same pricing page
- Lists ISO 27001, ISO 42001, SOC 2 Type I and II, and GDPR-aligned standards
- Offers both Nvidia and AMD accelerators, including MI355X
Weak on
- Most GPU models show 'Contact sales' instead of a rate
- Headquarters and legal entity were not confirmed on the pages opened
Skip it if: Developers who want to compare list prices across providers before talking to anyone, or hobbyists renting a single card. Pricing page shows a few on-demand and spot GPU rates; most models are contact sales.
- 10
CoreWeave
CoreWeave closes the list because it prices by 8-GPU node rather than per card: HGX H100 at $49.24 and HGX B200 at $68.80 per hour on-demand, with spot rates listed beside them. It is built for cluster-scale AI and ships Blackwell and Vera Rubin generations. A single developer renting one card will find it oversized.
Strong on
- Public node prices include spot rates, such as HGX H100 at $19.71 per hour
- SOC 2, ISO 27001, 27017 and 27018 badges on the Trust Center
- Zero Egress Migration program waives egress fees when moving data in
Weak on
- Compliance reports are shared under NDA, not published
- Pricing is per node, so entry cost is far above single-GPU rentals
Skip it if: Individual developers, students and small teams that need one GPU for hours at a time. Prices are per 8-GPU node: HGX H100 is $49.24 per hour, B200 is $68.80, with spot rates listed.
- 11
Verda
Verda ranks eleventh because its public price list is clear but its catalog is narrower than the hyperscale clouds. The pricing page shows an H100 80GB at $4.793 per hour on-demand and $2.397 on spot, plus H200, B200, B300 and GB300 cards, and confidential-computing variants on Blackwell. It markets EU ownership and EU data centers, which suits buyers with residency rules.
Strong on
- Spot prices are published next to every on-demand GPU, at roughly half the rate
- Trust page lists SOC 2 Type II and ISO 27001, 27017, 27018 and 27701
- Confidential computing instances keep data encrypted in memory on Blackwell cards
Weak on
- Reserved discounts start at only 2% for one month and reach 25% at two years
- Operates from a smaller footprint of data centers than the largest GPU clouds
Skip it if: Teams that need many regions across several continents or a managed Kubernetes and database stack beside the GPUs. On-demand hourly rates with a half-price spot tier; reserved terms cut on-demand rates by 2% to 25%.
- 12
Lightning AI ranks twelfth because it is a development platform first and a GPU rental second. Its pricing page lists T4 at $0.55 per hour up to B200 8-GPU nodes at $9.86 per GPU hour, billed by the second, with interruptible rates for some cards. The free tier gives limited GPU hours, and Pro costs $50 per month, or $20 monthly when billed annually. Bring-your-own-cloud with SOC2 and HIPAA sits in paid tiers.
Strong on
- Per-second billing across single GPUs, multi-GPU nodes and TPUs
- Free tier offers up to 80 GPU hours with no credit card
- Can run on partner clouds or inside your own AWS or GCP account
Weak on
- The free Studio restarts every four hours and caps at two concurrent GPUs
- Plan features and credit allowances vary, so total cost takes some reading
Skip it if: Large training teams that need raw bare-metal clusters and negotiated contracts rather than a Studio-based workflow. Billed by the second; the free tier includes up to 80 GPU hours, and Pro adds a monthly fee.
How these AI cloud GPU providers were judged
Which GPU models and cluster sizes can you actually rent?
Catalog depth decides whether a project fits. RunPod lists cards from a 48 GB A40 to a 288 GB B300, Hyperstack lists twelve models with prices, and Vast.ai counts more than 68 GPU types across its marketplace. Lambda and Nebius sell clusters of 16 to thousands of GPUs, while CoreWeave rents by the 8-GPU node and ships the newest generations first. OVHcloud and DigitalOcean offer a shorter menu, mainly L4, L40S, H100 and H200, plus AMD cards at DigitalOcean. Availability is a separate question: Lambda describes self-serve access as first-come, so check stock for your exact configuration before you plan a launch date.
What does an H100 hour cost, and what changes the price?
The same H100 ranges widely by provider and by buying mode. As of October 2026, Hyperstack lists $2.50 on-demand, OVHcloud $2.99 for one card excluding VAT, DigitalOcean $4.41, Nebius $4.50 per GPU hour for HGX H100 and Lambda $3.99 for an 8x SXM instance. Reserved terms lower the rate: Hyperstack starts at $1.75 and DigitalOcean lists $3.26 for twelve months. Spot and interruptible capacity is cheaper again, from $0.79 at Nebius, but can be reclaimed. Also compare billing granularity (second, minute or hour), minimum periods, storage charges and egress, since those add up on long jobs.
Where does your data sit and which audits cover the provider?
Read compliance claims at the level of region and tier, not at the logo. RunPod holds an ISO/IEC 27001:2022 certificate valid from August 27, 2026 and states that coverage varies by data center partner. Vast.ai separates standard hosts from Secure Cloud datacenters. Nebius lists SOC 2 Type II, ISO 27001 and GDPR processing as an EU-headquartered operator, and Modal documents SOC 2 Type II with a data processing addendum. CoreWeave shares full reports under NDA. Hyperstack names SOC 2 Type 2 for its parent, NexGen Cloud. Ask for the report and the processing region before sending customer data.
How hard is it to move a GPU workload to another provider?
Rented VMs and bare-metal nodes move easily if your code lives in containers and your data in object storage you can copy out. Lock-in grows with proprietary layers. Modal's decorators and image definitions tie code to its runtime, while RunPod Serverless endpoints and DigitalOcean's managed Droplets sit closer to standard containers and VMs. Reserved contracts at Lambda, Nebius, CoreWeave and Hyperstack trade a lower rate for a fixed term. CoreWeave runs a Zero Egress Migration program for moving data in, and Nebius lists network traffic at no charge, which lowers the cost of leaving.
Which provider fits a startup, a research lab or an enterprise?
Match the provider to how long the GPU stays busy. For experiments and short fine-tunes, Vast.ai, RunPod and Hyperstack give low entry rates and per-second or per-minute billing. For bursty inference without idle cost, Modal fits. For multi-week training on hundreds of GPUs, Lambda, Nebius and CoreWeave sell clusters with interconnects and reserved pricing. For a team already on DigitalOcean or a French public cloud, adding GPU instances beside existing services can beat a separate vendor. Regulated workloads should start from the compliance pages, not the price table.
How to choose an AI cloud GPU provider in three steps
- Size the job in GPU hours and memory. Write down the model size, the VRAM it needs, the number of GPUs and the hours per month. A job needing 80 GB per card rules out many 48 GB offers, and a job that runs for days belongs on reserved or spot capacity rather than on-demand. Multiply the hourly rate by realistic hours, then add storage and egress. A cheap rate on a card that is too small, or a card that sits idle between runs, costs more than a pricier match.
- Pick the buying mode before the vendor. Decide between on-demand, spot or interruptible, reserved, and serverless. On-demand suits unpredictable work, spot suits jobs that checkpoint, reserved suits steady training, and serverless suits traffic with long quiet periods. Each vendor in this ranking sells a different mix, so the mode narrows the list quickly. Then compare the same mode across providers instead of setting a spot price against an on-demand price, which tells you nothing about the real gap between two vendors, and write down the minimum billing period each one applies.
- Test one real job on your shortlist. Run the same training step or inference benchmark on two or three providers with your own container. Measure start-up time, throughput, disk speed and how long it takes to get a node at the hour you need one. Ask for the compliance report and the processing region while the trial runs. A one-day test shows capacity and support behavior that no price page can, and it costs far less than discovering the gap in the middle of a project, when moving data and rewriting launch scripts is expensive.
Frequently asked questions
What is the best AI cloud GPU provider in 2026?
Hyperstack ranks first in this list because it publishes on-demand, reserved and spot rates for many GPU models and bills by the minute. RunPod is the pick for pods plus serverless, and Modal for per-second serverless GPUs. The best choice depends on whether your workload is bursty, steady or cluster-scale.
How much does it cost to rent an H100 GPU?
As of October 2026, published on-demand H100 rates in this ranking run from $2.50 per hour at Hyperstack to $4.50 per GPU hour at Nebius, with OVHcloud at $2.99 excluding VAT. Reserved and spot capacity can be lower, and CoreWeave prices an 8-GPU H100 node at $49.24 per hour.
Is a marketplace like Vast.ai safe for company data?
Vast.ai separates standard hosts, which use Docker-level tenant isolation and suit non-regulated work, from a Secure Cloud tier of trusted datacenters for regulated industries. It reports SOC 2 Type 2 for the platform. For company data, filter to Secure Cloud and read its data processing terms first.
When does serverless GPU beat renting a dedicated instance?
Serverless wins when traffic is spiky and a dedicated GPU would sit idle. Modal bills per second and charges nothing for idle containers. A GPU that is busy around the clock usually costs less on a reserved or on-demand instance, because serverless per-second rates are set above the equivalent hourly rate.
Are there GPU clouds headquartered in Europe?
Yes. Nebius is headquartered in Amsterdam and OVHcloud is a French provider, and Hyperstack's operator NexGen Cloud has a head office in London. Their pages describe GDPR processing and, for OVHcloud, SecNumCloud-qualified infrastructure elsewhere in its range. Confirm the exact data center region for your instance.
What is the difference between on-demand, spot and reserved GPUs?
On-demand is billed while you use it with no commitment. Spot or interruptible capacity is cheaper but the provider can reclaim it, so jobs must checkpoint. Reserved capacity is a fixed term, such as 12 months at DigitalOcean, in exchange for a lower hourly rate and guaranteed supply.
Do I need to pay for egress and storage on top of GPU hours?
Often yes, and the amounts differ by vendor. Modal charges $0.04 per GiB of egress beyond plan allowances and $0.09 per GiB per month for volumes, while Nebius lists network traffic at no charge and bills storage separately per GiB per month. Add data transfer and storage to the GPU rate before comparing totals.
Which providers publish a full price list without contacting sales?
Hyperstack, RunPod, Modal, Lambda, Nebius and DigitalOcean show per-hour or per-second GPU rates for most models. CoreWeave prints per-node prices. Crusoe shows only some GPU rates and uses contact sales for the rest, and large reserved terms at most vendors are also quoted privately by their sales teams.