Best NVIDIA A100 GPU cloud providers in 2026

Choosing an NVIDIA A100 cloud provider is not just a matter of finding the lowest hourly rate. The right option depends on whether you need 40 GB or 80 GB of GPU memory, one GPU or a connected multi-GPU node, and raw infrastructure or a platform that also manages the surrounding application.

This guide compares five NVIDIA A100 cloud providers by GPU configuration, deployment workflow, access model, pricing scope, and fit for training, fine-tuning, inference, and high-performance computing (HPC).

• Northflank combines managed A100 40 GB and 80 GB capacity with services, jobs, CI/CD, autoscaling, observability, databases, and storage. You can also use self-serve BYOC or import an existing Kubernetes cluster through BYOK, including eligible on-premises and bare-metal clusters with compatible GPU-enabled nodes.
• Amazon EC2 provides fixed eight-GPU P4 instances with NVIDIA A100 40 GB or 80 GB GPUs for workloads running within AWS.
• Google Cloud Compute Engine offers A2 configurations from one A100 to large multi-GPU VM shapes, with 40 GB and 80 GB options.
• Lambda Cloud provides A100 VMs through its UI, API, and CLI, with per-GPU pricing billed by the minute.
• Runpod Serverless provides A100 80 GB workers for autoscaled inference, including workers that scale to zero when idle.

Evaluate each provider across the same practical criteria before comparing prices or choosing a deployment model.

• GPU configuration: Compare A100 40 GB and 80 GB options, PCIe and SXM form factors, GPU count, and interconnect. A single GPU for development is a different product from an eight-GPU node designed for distributed training.
• Operating model: Check whether the provider offers complete VMs, GPU-focused infrastructure, serverless workers, or an application platform. Also account for quotas, regional capacity, deployment tooling, and how much infrastructure your team must operate.
• Total workload cost: Full-machine, bundled GPU, GPU-only, and serverless GPU-second rates include different resources. Check whether CPU and RAM are included, then account for storage, network transfer, idle time, interruption risk, and commitments rather than comparing one headline figure.

These five providers represent distinct A100 operating models: an application platform, two hyperscale clouds, a GPU-focused VM provider, and a serverless GPU platform. Other cloud providers also offer A100 capacity.

The table below compares deployment models, A100 prices, and the resources included with each option. Prices are accurate as at 30 July 2026.

These rates are not like for like. Included GPU count, CPU, RAM, storage, networking, purchasing terms, and platform services differ.

Match each provider's operating model to the A100 workload, infrastructure requirements, and level of control your team needs.

Choose Northflank when your A100 workload is one part of a larger production application. It is particularly relevant when you have limited platform-engineering capacity or need infrastructure choice, governance, and a consistent workflow across environments.

• A100 choice and pricing: The Northflank GPU platform offers A100 40 GB at $1.42/hour and A100 80 GB at $1.76/hour, as at 30 July 2026. Per-second rates include the GPU and attached CPU and RAM; storage, networking, and separate CPU workloads cost extra. Check the pricing calculator for a current estimate. Enterprise pricing may apply to high-scale or custom deployments.
• Complete application platform: A100-backed inference, training, and batch jobs can run beside APIs, workers, scheduled jobs, managed databases, persistent volumes, object storage, networking, secrets, CI/CD, autoscaling, logs, metrics, and release workflows. Engineers use the same UI, CLI, API, and Git-based workflow for CPU and GPU services.
• Infrastructure choice: You can deploy GPUs on Northflank Cloud, deploy GPUs in your own cloud, or import an existing Kubernetes cluster through BYOK. BYOC covers providers including AWS, Google Cloud, Microsoft Azure, Oracle, and CoreWeave. BYOK can extend the workflow to eligible on-premises and bare-metal Kubernetes clusters, provided they meet Northflank's import requirements and have compatible GPU-enabled nodes available for workload scheduling.
• Governance and integration: Enterprise capabilities include SSO with SAML or OIDC, audit logs, global backups and HA/DR options, secure runtime and on-premises deployment, plus integrations for existing registries, Vault, and DNS. BYOC and BYOK can retain control over infrastructure, regional placement, networking, IAM, data residency, and cloud billing.
• GPU operations: Use the GPU documentation to deploy GPU-backed services and jobs. Use the configuration and optimisation guide for compatible images, CPU and memory sizing, persistent storage, scaling, health checks, ports, and startup commands. Northflank also supports NVIDIA MIG and time slicing, although time-sliced workloads do not receive fault or memory isolation from one another.

The distinction between these providers is the scope each one manages. Amazon EC2 and Google Cloud provide native infrastructure that your engineering team configures and operates. Lambda Cloud provides preconfigured GPU VMs, while Runpod Serverless provides autoscaling inference workers. Northflank manages the GPU workload and its surrounding application services on Northflank Cloud, in your own cloud through BYOC, or on eligible existing Kubernetes clusters through BYOK.

Choose Amazon EC2 when you already operate data, networking, security controls, and ML workloads on AWS. Its P4 family packages A100 GPUs into fixed multi-GPU VMs.

• GPU configuration: p4d.24xlarge contains eight A100 40 GB GPUs. p4de.24xlarge contains eight A100 80 GB GPUs.
• Host resources: Both shapes provide 96 vCPUs and 1,152 GiB of system memory.
• Purchasing options: You can use on-demand or Spot instances, commitments, or Capacity Blocks for ML where available.
• AWS fit: P4 instances can remain within an existing AWS architecture, account-governance model, and procurement relationship.

Choose Amazon EC2 when your workload can use a complete eight-GPU node and your team wants direct control within existing AWS identity, VPC, automation, governance, and purchasing workflows. P4 may be oversized for development or smaller inference workloads, and its full-instance or per-accelerator equivalent is not directly comparable with an independently rented single A100.

Choose Google Cloud Compute Engine when you want both single-GPU and large multi-GPU VM shapes inside the Google Cloud ecosystem.

• GPU configuration: A2 Standard uses A100 40 GB GPUs. A2 Ultra uses A100 80 GB GPUs and includes Local SSD.
• Scaling range: A2 Ultra ranges from one to eight GPUs per VM. A2 Standard also includes the 16-GPU a2-megagpu-16g shape.
• Platform integration: A2 VMs integrate with Google Kubernetes Engine and Google Cloud AI services.
• Purchasing options: A2 uses bundled accelerator-optimised VM pricing with on-demand, committed-use, and Spot VM options.

Choose Google Cloud Compute Engine when you want direct control over A2 VMs or GKE nodes and already have the expertise to operate the surrounding Google Cloud infrastructure. Check project quota, regional support, and live capacity because an A2 shape may be supported without being immediately allocatable in every project or region.

Choose Lambda Cloud when you want a preconfigured ML environment and a direct VM deployment workflow.

• GPU configuration: Choose a single A100 40 GB SXM or PCIe instance, a two- or four-GPU A100 40 GB PCIe instance, or an eight-GPU A100 40 GB or 80 GB SXM instance.
• Interface: Instances can be launched and managed through the UI, API, or CLI.
• Pricing: As at 30 July 2026, the A100 40 GB SXM and PCIe configurations cost $1.99 per GPU-hour, while the eight-GPU A100 80 GB SXM configuration costs $2.79 per GPU-hour, plus applicable sales tax. Usage is billed in one-minute increments.
• Included workflow: Lambda Cloud includes its ML software stack and does not charge egress fees for on-demand instances.

Choose Lambda Cloud when direct VM access and a preconfigured ML stack are the main requirements. Capacity is available on a first-come basis, and the available A100 memory and form factor depend on the instance size, so your preferred configuration may not always be immediately available.

Choose Runpod Serverless when your inference workload benefits from autoscaled workers rather than persistent VMs.

• GPU configuration: Runpod Serverless offers A100 80 GB GPUs for autoscaling inference workers.
• Worker models: Flex workers scale to zero when idle. Active workers remain running for consistent traffic and low-latency requirements.
• Pricing: As at 30 July 2026, Runpod lists A100 80 GB Serverless Flex workers at an hourly equivalent of $2.72. Serverless usage is billed per second. Storage is billed separately, and dedicated Pods use different pricing.
• Workload fit: The worker model suits variable inference demand where scaling behavior is more important than maintaining a persistent machine.

Choose Runpod Serverless when your workload benefits from autoscaling workers. These rates apply only to Serverless compute. Runpod Pods use a separate deployment and pricing model for persistent GPU instances.

Choose based on the layer you want the provider to manage:

• Choose Northflank when you need A100 services or jobs alongside application infrastructure, with a choice of managed cloud, BYOC, or BYOK.
• Choose Amazon EC2 when an eight-GPU A100 node fits the workload and the wider system already operates on AWS.
• Choose Google Cloud Compute Engine when you need a broad range of A2 sizes inside Google Cloud.
• Choose Lambda Cloud when direct VM access and a preconfigured ML stack are the priorities.
• Choose Runpod Serverless when the workload maps naturally to autoscaled inference workers.

Before reserving capacity, confirm the exact GPU memory and form factor, GPU count and interconnect, target region, quota or approval, live capacity, CPU and RAM allocation, storage and network charges, interruption terms, and minimum commitment. Test a representative workload before relying on hourly rates alone.

Yes, when its 40 GB or 80 GB memory, mature CUDA support, availability, and price-performance fit the workload. A100 capacity remains available for training, fine-tuning, inference, and HPC. For new deployments, benchmark the A100 against newer accelerators because a lower hourly price does not necessarily mean a lower cost per completed job.

Choose 80 GB when model weights, batches, activations, or working data cannot fit comfortably within 40 GB. A 40 GB option can cost less when the workload fits. Also check PCIe or SXM form factor and multi-GPU interconnect because memory capacity alone does not define the complete node.

Yes. Northflank lets you deploy a single A100 40 GB or 80 GB GPU for a service or job on Northflank Cloud, with BYOC and BYOK options when you need to use your own infrastructure. Google Cloud Compute Engine and Lambda Cloud also offer single-A100 instances, while Runpod Serverless provides A100 workers backed by a single GPU. Amazon EC2 P4 instances start with eight A100 GPUs.

Confirm the region, quota or access approval, live capacity, GPU memory and form factor, multi-GPU topology, CPU and RAM allocation, persistent storage, egress charges, interruption terms, minimum commitment, software image, and monitoring setup.

• How much does an NVIDIA A100 GPU cost? explains the cost of buying and renting A100 capacity.
• 12 best GPU cloud providers for AI and ML compares a broader range of providers and GPU models.
• NVIDIA H100 vs A100 compares the two GPU generations by architecture, memory, workload, and cloud cost.
• Deploy GPUs on Northflank Cloud walks through creating a GPU-enabled project and workload.
• Best options for BYOC in cloud computing compares approaches for keeping workloads in your own cloud account.