7 cheapest cloud GPU providers in 2026

Cheap cloud GPUs can reduce training and inference costs, but an inexpensive GPU that runs out of memory or needs repeated restarts can cost more to use. This comparison covers seven GPU cloud providers, their rates, and the decisions that determine your actual bill.

These seven providers offer different ways to reduce GPU costs. They aren’t ordered from cheapest to most expensive, because the lowest-cost option depends on the GPU, configuration, availability, and billing model. The examples below distinguish GPU charges from configured machines and keep live marketplace and spot prices separate.

Prices are in USD, checked on 8 October 2026, before applicable taxes. They are examples rather than equivalent complete deployments or guarantees of available capacity. PCIe and SXM variants, host resources, regions, and minimum GPU counts can change both performance and cost.

Smaller workloads may fit lower-cost cards; larger models may need more memory or suitable interconnects. Our guide to choosing a GPU for AI covers that hardware decision.

Find the lowest cost of completing your workload at the performance and reliability you need. Estimate the full deployment rather than the accelerator alone.

Compare the exact model and memory size, including whether the offer is a whole GPU or a partition. An A100 40GB and an A100 80GB are different options, as are H100 PCIe and H100 SXM machines. NVIDIA's H100 specifications explain the hardware differences between those variants. For multi-GPU training, check the interconnect and the number of GPUs you must rent together.

GPU cloud pricing can include more than the accelerator: account for CPU, system memory, persistent storage, model downloads, checkpoints, and network transfer. Some providers bundle host resources; others let you configure and pay for them separately.

Benchmark a representative workload with the same model, precision, batch size, and output requirements. Measure inference throughput at an acceptable latency, or training time to the required result. MLPerf Inference benchmarks provide a reference for comparing system performance, but test your own workload before estimating its running cost.

For a hypothetical training run, a $1/hour configuration taking ten hours costs $10 in compute. A $2/hour configuration completing the same work in four hours costs $8. The second has a higher hourly rate but a lower job cost. Add storage, transfer, and any failed runs to both estimates.

Check when billing starts and stops. Model loading, warm workers waiting for requests, and retained disks can all contribute to the bill. Stopping GPU compute does not necessarily delete its storage or end every associated charge.

Confirm that your required GPU count is available in a suitable region and that the account can provision it. For a deadline or sustained production load, evaluate reserved capacity and support terms alongside the price. A public price listing alone does not reserve hardware for you.

These providers cover managed workloads, marketplace offers, serverless execution, virtual machines, and capacity commitments. Choose around your application and the operations your team wants to own.

Northflank runs GPU services and jobs within a platform that also manages application services, databases, and delivery workflows. It suits teams building production AI applications, where the GPU is one part of a system that needs deployment, scaling, and monitoring.

Capabilities include:

• Container-based GPU services and jobs for inference, training, and batch processing.
• Configurable GPU counts and CPU/memory resources, with deployment and autoscaling settings.
• Application services, managed databases, secrets, logs, and metrics alongside GPU workloads.
• Managed cloud hosting and bring your own cloud (BYOC) options. With BYOC, workloads run inside your cloud account and VPC, while Northflank acts as the control plane managing Kubernetes and deployments.

Where it can reduce costs: Northflank combines GPU provisioning with application operations, reducing the infrastructure your team needs to assemble and maintain separately. For eligible workloads on your own cloud, GPU node pools and spot scheduling provide ways to use existing capacity and cloud purchasing arrangements within the same deployment workflow.

The Weights customer story illustrates the value of optimising more than the hourly rate: the team reduced model-loading time from seven minutes to 55 seconds using a shared cache layer. That result reflects its particular workload and architecture.

Best for: Teams deploying production inference and training workloads alongside application services, organisations that need workloads in their own VPC, and developers who want managed GPU deployment and operations.

Pricing and limits: Pay-as-you-go GPU pricing includes L4, A100, and H100 options; the table above separates their GPU rates from other resources. GPU workloads are billed by the second once provisioned. Enterprise plans support custom requirements, volume discounts, and 24/7 support. GPU models and capacity vary by region.

Trade-off: GPU availability varies by region, so your preferred model and GPU count may affect where you can deploy.

Vast.ai is a GPU marketplace where providers list hardware with their own prices and configurations. You can select an offer around the GPU, location, host characteristics, and rental type rather than a single fixed instance catalogue.

Capabilities include:

• GPU offers spanning consumer RTX cards and data-centre accelerators.
• On-demand, interruptible, and reserved rental options.
• Templates, SSH access, and CLI/API tools for deploying workloads.

Where it can reduce costs: Host-level selection gives you more variables to adjust when looking for a lower quote. Interruptible offers can fit experiments or batch jobs that save progress and tolerate pauses; on-demand offers serve different continuity requirements.

Best for: Researchers, developers, and teams prepared to evaluate individual offers and manage their workload's recovery behaviour.

Pricing and limits: Compute prices vary by host and are charged per second while an instance runs. Storage and upload/download traffic have separate host-dependent charges. Select an actual offer before estimating the total; a platform-wide minimum does not describe every GPU or rental type.

Trade-off: Price comparison takes more work because host resources, bandwidth charges, and reliability scores differ. A cheaper listing may not be the cheapest place to complete a data-heavy job.

Runpod provides GPU containers called Pods and a separate Serverless service for inference workers. This gives teams a choice between maintaining a running environment and scaling workers around incoming work.

Capabilities include:

• GPU Pods with templates and configurable storage.
• Serverless Flex workers that can scale to zero and Active workers that remain running.
• On-demand Pods and prepaid savings plans for longer-running workloads.

Where it can reduce costs: Pods fit sustained development or compute sessions, while Serverless can reduce idle GPU time for variable inference traffic. The right choice depends on how often requests arrive and how expensive it is to initialise the model.

Best for: AI teams comparing a persistent GPU environment with managed inference workers.

Pricing and limits: The table shows Pods prices, not Serverless prices. Pod storage is separate from compute, and retained storage can remain billable after stopping a Pod. Serverless bills the worker lifecycle, including startup, request execution, and the configured idle period before shutdown.

Trade-off: Scaling to zero trades lower idle usage for startup work when demand returns. Measure model loading and request latency before assuming Serverless will cost less than a continuously running Pod.

For a closer comparison of deployment models, see Runpod vs Vast.ai vs Northflank.

TensorDock provides GPU virtual machines through a network of hosts. Its configurable resource model suits workloads that need control over the operating system and the CPU, memory, and storage paired with the accelerator.

Capabilities include:

• KVM virtual machines with root access and dedicated GPUs.
• Separately configurable CPU, RAM, and storage allocations.
• An API for provisioning and managing GPU servers.

Where it can reduce costs: You can size the host resources around the workload rather than accepting one fixed CPU-to-GPU ratio. That is useful when a job's system-memory or CPU needs differ from a provider's standard bundle.

Best for: Teams and developers who want a GPU VM with operating-system control and flexible resource sizing.

Pricing and limits: Advertised examples include RTX 4090 GPUs from $0.35/hour and H100 SXM5 GPUs from $2.25/hour. Those are GPU rates: CPU, RAM, and storage add to the total, and the available host determines the final configuration and quote.

Trade-off: VM control also leaves more of the software environment and workload operations with your team. Budget for setup, updates, and recovery procedures alongside the rental price.

Amazon EC2 provides GPU instances within AWS, with several purchasing models and integrations with its storage, networking, and identity services. It is relevant when moving the workload elsewhere would also mean moving substantial data or replacing existing infrastructure integrations.

Capabilities include:

• GPU instance families for training and inference, including single- and multi-GPU H100 configurations.
• Spot capacity for interruptible workloads and Capacity Blocks for scheduled GPU reservations.
• Integration with AWS storage and container orchestration services.

Where it can reduce costs: Spot capacity can lower compute spend for jobs that checkpoint and resume. AWS advertises savings of up to 90% for EC2 Spot generally, but the discount for your GPU instance depends on its actual spot and on-demand rates.

Best for: Teams already operating in AWS that can use interruptible capacity or plan GPU requirements in advance.

Pricing and limits: EC2 bills the whole instance, not an independently rentable fraction of a multi-GPU machine. Spot rates depend on the instance and location. Capacity Blocks require a prepaid reservation; their effective hourly figures should not be treated as on-demand prices.

Trade-off: Spot capacity can be reclaimed. Your job must handle interruption and restart, and a replacement instance may not be immediately available. The AWS interruption documentation explains the lifecycle to design around.

Lambda offers GPU instances with predefined CPU, memory, and local SSD allocations. Its instance options and machine-learning software environment suit teams that want to run training or inference without assembling a hardware configuration from separate components.

Capabilities include:

• Single- and multi-GPU instance sizes.
• A machine-learning software stack and direct access to the instance environment.
• Persistent filesystems and larger cluster options for workloads with different scale requirements.

Where it can reduce costs: Bundled host resources make it easier to estimate the compute configuration, particularly when your job uses the included memory and local storage. Match the bundle to the workload rather than comparing GPU rates alone.

Best for: Researchers and engineering teams running training or inference on defined GPU instance configurations.

Pricing and limits: Single-GPU examples include A100 40GB at $1.99/hour and H100 PCIe at $3.29/hour. On-demand usage is billed in one-minute increments, including idle running time. A100 80GB at $2.79 per GPU-hour belongs to an eight-GPU configuration, costing $22.32 per instance-hour.

Trade-off: A lower per-GPU rate on a larger machine requires renting the larger allocation. If your workload cannot use those GPUs effectively, a smaller instance can have a lower total cost.

Paperspace, part of DigitalOcean, provides GPU Machines and browser-based Notebooks. It suits teams that value an interactive development environment for exploring models and datasets alongside access to GPU compute.

Capabilities include:

• GPU virtual machines for training, inference, and development.
• Jupyter-based Notebooks with shared persistent storage.
• Notebook templates and tools for working with datasets and custom containers.

Where it can reduce costs: A notebook workflow can reduce the setup needed for exploratory work. For occasional experiments, choosing suitable hardware and shutting down unused machines matters more than keeping a large GPU running between sessions.

Best for: Teams, researchers, and developers working interactively with models and datasets.

Pricing and limits: Machine examples include A100 40GB at $3.09/hour, A100 80GB at $3.18/hour, and H100 at $5.95/hour. Notebook subscription plans and machine usage are separate considerations. These are Paperspace rates, not DigitalOcean GPU Droplet prices.

Trade-off: Powering off a Machine stops its compute billing, but storage, public IP addresses, and other retained resources can continue to accrue charges. Include those costs in the budget for an occasional-use environment.

Match the purchasing model to utilisation, and make the workload efficient to start, run, and recover.

Spot or interruptible capacity can suit batch inference, experiments, and training jobs that save checkpoints. Store progress somewhere that survives loss of the compute instance, and test that a replacement can resume correctly. Retried work, model loading, and waiting for capacity belong in the cost estimate.

For more detail, see how spot GPUs work. Orchestration helps schedule and replace workloads; it does not make an interruption-sensitive application immune to reclaimed hardware.

Serverless can be cheaper when GPU demand is intermittent and workers shut down between requests. It is not automatically cheaper for steady traffic, particularly if repeated model loading or keeping warm workers consumes much of the day. Compare the actual billable worker time with the cost of an always-running configuration at the same latency target.

A commitment can make sense when you have sustained utilisation and a clear requirement for a particular GPU configuration. Check its duration, cancellation rules, capacity guarantees, and minimum allocation. A discounted resource you leave idle can cost more than a flexible resource you use only when needed.

Bring your own cloud (BYOC) can help when you already have eligible credits, negotiated rates, or infrastructure that needs to remain in your account. Northflank acts as the control plane for GPU workloads running inside your cloud account and VPC, managing Kubernetes and deployment workflows. Estimate the cloud resources and platform costs together; eligibility for credits or discounts depends on your cloud agreement.

Choose Northflank when you want GPU workloads and the surrounding application managed together; Vast.ai or TensorDock when selecting individual offers or controlling a VM matters most; and Runpod when comparing persistent Pods with serverless inference. Lambda suits defined training configurations, Paperspace suits notebook-led work, and EC2 fits AWS-dependent workloads. Test a representative job and compare its complete cost before making a longer commitment.

These questions cover the hardware and billing choices that most affect a GPU rental decision.

The cheapest offer changes with the H100 variant, region, and rental type. TensorDock advertises H100 SXM5 GPU rates from $2.25/hour, while Northflank lists H100 80GB at $2.74/hour; both require additional host resources. Runpod lists an H100 PCIe Pod at $2.89/hour with CPU/RAM allocated. Match the variant, complete configuration, and interruption terms before comparing totals.

The cheapest option is the configuration that fits your model and delivers the required throughput and latency at the lowest total cost. Smaller workloads may fit lower-cost L4 or RTX-class GPUs; larger models or higher concurrency may justify more memory or faster hardware. Benchmark the actual model and include idle time rather than choosing solely by the GPU's hourly price.

They can be cheaper for intermittent traffic because workers can scale down between requests. Dedicated running instances can be more economical for sustained demand. Include startup, model loading, any billed idle time, storage, and latency requirements in the comparison; serverless billing does not always cover request execution alone.

Renting can cost less for short projects, variable demand, or workloads that need different hardware over time. Buying can make sense with sustained utilisation, but compare the complete system cost, power, cooling, maintenance, and replacement needs over your expected ownership period. There is no universal break-even point that applies to every GPU and workload.

These guides cover hardware selection, rental planning, and broader provider comparisons:

• Best GPU cloud providers for AI and machine learning
• Best GPUs for AI workloads
• How much does an NVIDIA A100 GPU cost?
• Rent GPUs for AI