Cloud GPU Pricing Guide 2026: Compare Azure, AWS, RunPod, Vast.ai and More

An October 2026 market comparison published by Hostinger evaluates 10 prominent GPU cloud providers across hardware selection, billing mechanics, and deployment models: Hostinger, RunPod, Vast.ai, Lambda, CoreWeave, Modal, Google Cloud, AWS, Microsoft Azure, and Nebius. Because Hostinger is itself a vendor in the space and places its own service at the top of its guide, seasoned IT pros should read the ranking as a vendor's editorial perspective rather than an objective trophy ceremony. However, when you strip away the marketing polish and cross-check the claims against official provider documentation, a genuinely useful framework emerges—along with several critical pricing caveats that every buyer needs to understand before spinning up an instance.

The Four Cloud GPU Operating Models​

Rather than crowning a single "best" provider, it makes far more technical sense to divide these ten platforms by how they actually operate and bill for compute. What works for a solo developer testing a LoRA adapter in ComfyUI from a Windows 11 terminal will bankrupt or frustrate an enterprise team orchestrating distributed training across Microsoft Azure or AWS.

Summary: Cloud GPU providers fall into four distinct buckets: simple root-access VMs, serverless/container platforms, open hardware marketplaces, and heavy-duty hyperscaler or InfiniBand cluster operators.

Why Headline GPU Prices Can Mislead You​

Before looking at individual vendors, it helps to examine why a static price table is already outdated the moment you read it. Comparing official provider pricing documentation against Hostinger's October 2026 roundup reveals five structural traps that routinely catch buyers off guard:

1. Per-GPU Rates vs. Mandatory Full-Node Rentals: A provider might advertise an H100 at roughly $6.16 per GPU-hour or an A100 at $2.70 per GPU-hour. On CoreWeave's on-demand training tier, however, those figures represent one-eighth of an 8-GPU HGX node. You cannot rent just one; you are paying $49.24/hour for the full H100 HGX node or $68.80/hour for an 8-GPU B200 HGX node. (CoreWeave only breaks out single-GPU hourly pricing on its separate inference platform.) The same applies to high-end AWS P5/P6 instances, Google Cloud A3 Mega/Ultra VMs, and Azure ND-series virtual machines.
2. Configuration Size Changes the Unit Price: On Lambda's official instance sheet, per-GPU hourly pricing changes depending on how many GPUs you bundle together. While Lambda's 8x NVIDIA B200 SXM6 plan lists at $6.69 per GPU/hour, dropping down to a 4x configuration raises the per-GPU rate to $6.79/hour, a 2x plan to $6.89/hour, and a 1x B200 instance to $6.99/hour. Similarly, a single H100 SXM on Lambda is $4.29/hour, compared to $3.99/hour per GPU in an 8x node.
3. Different Purchasing Mechanisms Mixed Together: Hostinger's overview lists AWS p5.4xlarge (1x H100) at $5.19/hour and p6-b200.48xlarge (8x B200) at $98.84/hour, noting these are EC2 Capacity Block rates. Capacity Blocks require scheduling and paying upfront to reserve hardware for a defined window; they are not standard click-to-launch On-Demand rates. Meanwhile, Google Cloud rates depend heavily on the specific zone (such as us-central1 in Iowa) and whether you attach sustained-use or committed-use discounts, while Nebius publishes base rates alongside dated price adjustments and dynamic Spot rates that can fluctuate every 15 minutes.
4. Live Price Drift and Metering Fine Print: Even within a single vendor's ecosystem, numbers shift quickly. Hostinger's tutorial cites an entry price of $0.38/hour for an RTX 4090 (and $1.43/hour for an A100) with per-minute metering, whereas Hostinger's live GPU hosting product page also displays RTX 4090 instances starting from $0.42/hour ($0.64/hour for an RTX 5090 and $1.55/hour for an A100) and notes in its FAQ that hourly credit billing rounds up to the next full hour when an instance runs past the hour mark. Always check the live checkout screen before deploying.
5. The "Stopped Instance" Storage Tax: Across Hostinger, Vast.ai, and RunPod, halting your training script does not necessarily halt your bill. On Hostinger, compute charges continue even when the GPU sits completely idle until you permanently destroy the instance. On RunPod, persistent Pod volumes double in price from $0.10/GB/month while running to $0.20/GB/month while your Pod is idle (likely to discourage users from squatting on local NVMe storage on physical host machines). On Vast.ai, storage charges continue accruing for as long as an instance exists, even when inactive.

Summary: Never compare cloud GPUs by dividing a node price by eight or trusting a static table; verify whether you are renting a single card or an entire server, how idle storage is penalized, and whether billing is rounded by the second, minute, or hour.

Deep Dive: Direct VMs, Marketplaces, and Serverless Platforms​

For independent developers, startups, and enterprise innovation labs looking for rapid provisioning without hyperscaler bureaucracy, six specialized providers dominate the conversation.

1. Hostinger: Preconfigured Simplicity vs. Bare-Metal Control​

Hostinger positions its GPU Hosting service as a bridge between bare-metal Linux servers and turnkey AI apps. You get full root, browser terminal, and SSH access, alongside one-click preconfigured templates for tools like Ollama, Open WebUI, ComfyUI, and Stable Diffusion that come with NVIDIA drivers and containers preinstalled.

• Hardware & Pricing Snapshot: Options span from a 24GB RTX 4090 ($0.38/hour in the tutorial; $0.42/hour on the current storefront) and a 32GB RTX 5090 ($0.64/hour on the storefront), up through a 48GB L40S ($0.92–$0.96/hour), a 96GB RTX PRO 6000 Server ($0.60–$0.65/hour), an 80GB A100 ($1.43–$1.55/hour), and a 192GB Blackwell B200 ($4.50/hour shared, or $7.08/hour dedicated).
• The Catch: You must actively manage persistent storage because destroying an instance—the only way to stop compute billing—wipes everything not saved to external persistent volumes. Egress traffic, however, is free.

2. RunPod: Combining Persistent Pods with Serverless Endpoints​

RunPod is built for teams that want to interactively build or fine-tune a model inside a containerized Pod and then serve that exact model via an autoscaling Serverless API endpoint without switching clouds. It spans 31 regions and divides inventory between Community Cloud and Secure Cloud tiers.

• Hardware & Pricing Snapshot: Catalog highlights in Hostinger's comparison include the 24GB RTX A5000 ($0.27/hour), 48GB A40 ($0.49/hour), 48GB L40S ($1.09/hour), 80GB A100 PCIe ($1.59/hour), 80GB H100 PCIe ($2.89/hour), 141GB H200 ($4.59/hour), 180GB B200 ($6.79/hour), and 288GB B300 ($7.89/hour), though live portal rates fluctuate by tier and region.
• The Catch: Compute is billed by the second with zero ingress or egress fees, and multi-node Clusters are available. However, teams must navigate separate network volume fees ($0.07/GB/month under 1TB; $0.05/GB/month above 1TB) and idle Pod volume penalties.

3. Vast.ai: The Open Marketplace Approach​

Think of Vast.ai as a live commodity exchange for GPU compute. It aggregates more than 20,000 GPUs across 40+ data centers and 68+ GPU models from independent hosts, alongside newer Serverless and InfiniBand Cluster options.

• Hardware & Pricing Snapshot: Because pricing is driven by supply and demand, the spread between the lowest starting offer and the median market price is crucial. In the comparison snapshot, a 24GB RTX 3090 starts at $0.07/hour (median $0.15/hour), an RTX 4090 starts at $0.12/hour (median $0.36/hour), a 32GB RTX 5090 starts at $0.29/hour (median $0.44/hour), an 80GB H100 SXM starts at $1.33/hour (median $2.00/hour), a 141GB H200 starts at $2.63/hour (median $4.61/hour), and a 192GB B200 starts at $5.31/hour (median $6.00/hour).
• The Catch: Two listings for an RTX 4090 can have wildly different system RAM, PCIe lane speeds, disk I/O, bandwidth pricing, and host uptime. While per-second billing and 50%+ discounts on interruptible instances are great for checkpointed batch training, Vast.ai is ill-suited for compliance-heavy enterprise workloads that require uniform hardware guarantees.

4. Lambda: Built by ML Engineers, for ML Engineers​

Lambda skips the consumer gaming cards and focuses strictly on deep learning infrastructure. Every instance boots with the Lambda Stack—meaning PyTorch, CUDA, cuDNN, and essential ML libraries are ready out of the box.

• Hardware & Pricing Snapshot: Lambda bundles generous vCPUs, system RAM, and local NVMe SSD storage directly into its per-minute GPU rates with zero egress fees. For example, an 8x H100 SXM node ($3.99/GPU/hour) comes packed with 208 vCPUs, 1,800 GiB of RAM, and 22 TiB of SSD storage. Other 8x per-GPU rates include the 16GB V100 ($0.79/hour), 40GB A100 SXM ($1.99/hour), 80GB A100 SXM ($2.79/hour), and 180GB B200 SXM6 ($6.69/hour), plus 1x options like the 96GB GH200 ($2.29/hour) and A6000 ($1.09/hour).
• The Catch: Self-service instances are allocated first-come, first-served, so popular single-GPU configurations can sell out during peak hours. For massive workloads, Lambda offers 1-Click Clusters scaling from 16 to over 2,000 H100 or B200 GPUs.

5. Modal: Pure Python Serverless Execution​

If you hate managing Linux VMs, Dockerfiles, and Kubernetes manifests, Modal abstracts the server away entirely. You write standard Python decorators—such as @app.function(gpu="A100") or @app.function(gpu="H100:8")—and Modal spins up containers on demand, bills by the second, and scales down to zero when requests stop.

• Hardware & Pricing Snapshot: Per-second rates equate to roughly $0.59/hour for a T4, $0.80/hour for an L4, $1.10/hour for an A10, $1.95/hour for an L40S, $2.50/hour for an A100 80GB, $3.03/hour for an RTX PRO 6000, $3.95/hour for an H100 SXM5, $4.54/hour for an H200 SXM, $6.25/hour for a B200, and $7.10/hour for a B300. Interestingly, Modal's documentation notes that it may automatically upgrade A100 (40GB) requests to 80GB A100s, or H100 requests to 141GB H200s, at no extra charge (unless you pass gpu="H100!" for strict benchmarking). It also supports fallback arrays like gpu=["H100", "A100-40GB:2"].
• The Catch: Requesting NVIDIA Blackwell Ultra B300 GPUs on Modal requires your container image and libraries to support CUDA 13.1+. You also pay separately for CPU and memory, and while the Starter workspace is free (with $30/month in compute credits and up to 10 concurrent GPUs), the Team plan adds a $250/month platform fee (including $100/month in credits and up to 50 concurrent GPUs). Cold starts are mitigated via memory snapshots, and Modal even supports Microsoft Entra SSO for enterprise teams.

6. Nebius: Granular VMs That Scale to InfiniBand Clusters​

Nebius is an AI-native cloud that lets engineers start on a single GPU virtual machine billed by the second and scale up to multi-node clusters wired with NVIDIA Quantum-2 InfiniBand and Managed Kubernetes.

• Hardware & Pricing Snapshot: Published base GPU-hour charges in the comparison include the L40S ($1.55/hour with AMD CPU or $1.82/hour with Intel CPU; preemptible from $0.74–$0.90/hour), RTX PRO 6000 ($1.80/hour; $0.95/hour preemptible), H100 ($3.85/hour; $2.15/hour preemptible), H200 ($4.50/hour; $2.45/hour preemptible), B200 ($7.15/hour; $3.95/hour preemptible), and B300 ($7.85/hour; $4.30/hour preemptible), alongside massive GB300 NVL72 racks.
• The Catch: CPU, RAM, and erasure-coded block storage ($0.071/GiB/month) can add to the base GPU price. Furthermore, regional footprints are strict: while H200 spans multiple US and European regions (eu-north1, eu-north2, eu-west1, us-central1), H100 is restricted to eu-north1, RTX PRO 6000 to us-central1, and B200/B300 to specific US, UK, European, and Middle Eastern zones.

Summary: Specialized AI clouds offer faster onboarding, per-second or per-minute billing, and zero-egress perks, provided you match their architecture—serverless Python on Modal, marketplace hunting on Vast.ai, or prebuilt ML stacks on Lambda—to your team's workflow.

Enterprise Heavyweights: Azure, AWS, Google Cloud, and CoreWeave​

For enterprise IT administrators and cloud architects—especially those managing Windows Server estates, hybrid Entra ID environments, and strict data-governance pipelines—moving gigabytes of proprietary training data to a third-party startup cloud is often a non-starter. That is where the hyperscalers and CoreWeave step in.

Microsoft Azure: Tight Integration for Microsoft Shops​

If your organization already lives inside Azure Kubernetes Service (AKS), Azure Managed Disks, and Azure Monitor, spinning up Azure GPU Virtual Machines keeps your security boundary, billing, and Virtual Network peering under one roof.

• NC vs. ND Series: Azure splits its heavy AI compute between smaller 1-to-2 GPU VMs (the NC series) and massive 8-GPU distributed training nodes (the ND series) equipped with InfiniBand, NVLink, and GPUDirect RDMA.
• Pricing Structure: In Hostinger's snapshot, Azure's rates are quoted as full-month VM prices covering the GPU, vCPUs, and system RAM:
  • NC24ads A100 v4 (1x A100 80GB, 24 vCPUs, 220 GiB RAM): $2,681.29/month pay-as-you-go ($495.50/month Spot).
  • NC40ads H100 v5 (1x H100 NVL 94GB, 40 vCPUs, 320 GiB RAM): $5,095.40/month pay-as-you-go ($941.63/month Spot).
  • ND96asr A100 v4 (8x A100, 96 vCPUs, 900 GiB RAM): $19,853.81/month pay-as-you-go ($4,367.84/month Spot).
  • ND96amsr A100 v4 (8x A100 80GB, 96 vCPUs, 1,900 GiB RAM): $23,922.10/month pay-as-you-go ($6,162.33/month Spot).
  • ND96isr H100 v5 (8x H100, 96 vCPUs, 1,900 GiB RAM): $71,773.60/month pay-as-you-go ($13,263.76/month Spot).
• Admin Takeaway: Azure Spot VMs offer massive discounts (over 80% off pay-as-you-go in those configurations), but Azure can evict your workload whenever capacity tightens. Use Spot strictly for stateless batch inference or training runs that write frequent checkpoints to persistent storage. Also remember to request regional vCPU/GPU quota approvals early, as ND-series H100, H200 (141GB), and RTX PRO 6000 Blackwell availability varies sharply by region.

AWS: Elastic Scale and Software Stack Precision​

AWS delivers GPU muscle through Amazon EC2 accelerated computing families—ranging from L4 (G6) and L40S (G6e) to A100 (P4), H100/H200 (P5, P5e, P5en), and Blackwell B200/B300 (P6, plus 36- or 72-GPU P6e UltraServers)—as well as managed orchestration via Amazon SageMaker.

• Capacity Block Snapshot (US East): p4d.24xlarge (8x A100) at $11.80/hour; p5.4xlarge (1x H100) at $5.19/hour; p5.48xlarge (8x H100) at $41.53/hour; p5e.48xlarge (8x H200) at $47.76/hour; p5en.48xlarge (8x H200 with enhanced networking) at $54.92/hour; p6-b200.48xlarge (8x B200) at $98.84/hour; and p6-b300.48xlarge (8x B300) at $112.32/hour.
• Critical Technical Caveat from AWS Documentation: If you build custom AMIs rather than using preconfigured AWS Deep Learning AMIs, pay close attention to driver matrices and CPU C-states (which should be disabled for low-latency HPC). AWS's official EC2 documentation flags a specific compatibility issue on p5.4xlarge instances when GPU-to-GPU communication uses Elastic Fabric Adapter (EFA) and the NVIDIA Collective Communications Library (NCCL): admins must set the environment variable FI_HMEM_DISABLE_P2P=1 and install EFA version 1.43.1 or newer. Additionally, P6-B200 instances require minimum NVIDIA driver 570 / CUDA 12.8, while P6-B300 requires driver 580 / CUDA 13.0.

Google Cloud & CoreWeave: Cluster Specialists​

• Google Cloud Compute Engine: Bundles GPUs into accelerator-optimized VMs. Using Iowa (us-central1) as a reference benchmark, a g2-standard-4 (1x L4, 4 vCPUs, 16 GiB RAM) runs $0.71/hour ($0.42/hour Spot); an a2-highgpu-1g (1x A100, 12 vCPUs, 85 GiB RAM) is $3.67/hour ($2.12/hour Spot); a g4-standard-48 (1x RTX PRO 6000, 48 vCPUs, 180 GiB RAM) is $4.50/hour ($1.61/hour Spot); an a2-ultragpu-1g (1x A100 80GB, 12 vCPUs, 170 GB RAM) is $5.07/hour ($2.93/hour Spot); an a3-ultragpu-8g (8x H200, 224 vCPUs, 2,952 GB RAM) is $84.81/hour ($49.01/hour Spot); and an a3-megagpu-8g (8x H100, 208 vCPUs, 1,872 GB RAM) hits $93.40/hour ($56.03/hour Spot). Google Cloud's official GPU pricing documentation notes that Spot discounts generally range from 60% to 91% off on-demand rates, though discounts on local SSDs and A3 machines are smaller.
• CoreWeave: Built from the ground up for multi-node AI training with NVIDIA Quantum-2 InfiniBand and CoreWeave Kubernetes Service (CKS). Full 8-GPU node on-demand pricing in North America and Europe includes 8x L40 ($10.00/hour), 8x L40S ($18.00/hour), 8x RTX PRO 6000 Blackwell High Memory ($20.00/hour), 8x A100 ($21.60/hour), 8x H100 HGX ($49.24/hour; Spot ~$19.51–$19.71/hour), 8x H200 HGX ($50.44/hour; Spot ~$20.64–$20.93/hour), and 8x B200 HGX ($68.80/hour; Spot ~$34.11–$34.87/hour), plus 2-Superchip GB200 NVL72 instances ($42.00/hour). Unlike the traditional Big Three hyperscalers, CoreWeave waives internet egress and internal data transfer fees, charging $0.06/GB/month for Hot AI Object Storage down to $0.0125/GB/month for Archive.

Summary: Hyperscalers and CoreWeave dominate multi-node distributed training and enterprise security compliance, but they require careful quota planning, egress cost modeling, and attention to full-VM minimum commitments.

Step-by-Step Guide: How to Right-Size and Audit Your Cloud GPU Deployment​

Before you hand over a corporate credit card or consume your Azure/AWS enterprise commit, walk through this five-step checklist to avoid paying for idle silicon:

1. Calculate Your Exact VRAM Floor First:
   Do not shop by GPU brand name; shop by video memory (VRAM) and memory bandwidth. A 7-billion to 13-billion parameter model in 8-bit or 4-bit quantization or a Stable Diffusion / ComfyUI pipeline fits comfortably on a 24GB card (RTX 4090, RTX A5000, L4) or 48GB card (L40S, A40). Serving a 70B+ parameter LLM at high concurrency or fine-tuning large weights requires 80GB to 192GB+ of HBM memory (A100 80GB, H100, H200 141GB, or B200 180GB–192GB). Remember: for single-prompt, small-batch LLM inference, your bottleneck is memory bandwidth rather than raw arithmetic FLOPS—meaning a pricey Blackwell B200 may not deliver a proportional speedup over an H200 if the chip is just waiting on memory reads.
2. Audit the "All-In" Hourly Billable Unit:
   Write down the exact formula for 100 hours of runtime PLUS 1TB of persistent dataset storage PLUS 500GB of outbound model checkpoints:
   • Are CPU and system RAM bundled (Lambda, Azure, AWS, Google Cloud) or billed as separate line items (Modal, Nebius)?
   • Will egress fees add 20% to 40% to your bill when pulling checkpoints back to on-premises Windows storage, or does the provider offer zero egress (Hostinger, RunPod, Lambda, CoreWeave)?
3. Verify CUDA and Driver Compatibility for Next-Gen Silicon:
   Jumping onto NVIDIA Blackwell (B200) or Blackwell Ultra (B300) hardware requires modern software stacks. As documented by Modal and AWS, B200 requires at least CUDA 12.8+, and B300 requires CUDA 13.0 / 13.1+. If your organization's legacy Python libraries or custom C++ kernels are pinned to CUDA 11.8 or 12.1, stick with Hopper (H100/H200) or Ampere (A100) where precompiled kernels are mature.
4. Automate Checkpointing and Instance Teardown:
   If you use Spot, preemptible, or interruptible instances (which can slash compute costs by 45% to 80% on Azure, Google Cloud, CoreWeave, Nebius, and Vast.ai), configure your training loop to save weights to persistent network storage every 15 to 30 minutes. Furthermore, if you use a provider like Hostinger where stopping a job leaves the hourly meter running, add a teardown script or webhook alert so an idle VM doesn't quietly drain your credits over a long weekend.
5. Test Regional Latency and Quota Before Committing:
   Verify that your target GPU is actually in stock in the region closest to your data lake. An advertised rate means nothing if the SKU is restricted to a single European data center or locked behind a two-week enterprise quota review.

Whether you are remoting into a $0.40/hour RTX 4090 from Windows Terminal to test an open-weight vision model or orchestrating an 8-node H200 cluster inside Azure, the golden rule of 2026 cloud computing remains the same: match the billing mechanics to your workload's pulse, and never leave a meter running on an empty room.