The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
There is no defensible universal cheapest winner among AWS, Azure, and Google Cloud from the currently published figures. The examples use different regions, GPU and VM configurations, billing models, and service scopes: Azure publishes monthly machine-learning and Databricks totals, while Google publishes a whole-VM Spot example and AWS’s cited material provides configurations and purchase options but no matching price. A fair decision requires matching the GPU model and count, vCPU and RAM, region, currency, utilization, purchase term, interruption tolerance, and whether storage, networking, software, or a managed-service fee is included.
Which cloud GPU is cheapest?
That question cannot be answered fairly by comparing one headline number from each provider. A GPU-only rate can omit the machine, and a monthly platform total can include services that a bare virtual machine does not. Spot prices also buy a different reliability profile from on-demand capacity.
As an Amazon Associate I earn from qualifying purchases.
Recommended Free Tools
The figures available for this comparison illustrate the problem:
| Provider and example | Configuration | Billing and cost scope | Displayed price | What the figure does not prove |
|---|---|---|---|---|
| Azure Machine Learning | ND96asr A100 v4; eight A100 GPUs | Pay-As-You-Go monthly total shown on the Azure Machine Learning pricing page | $19,853.810 per month | It is not a region-matched quote against AWS or Google Cloud, and the page also displays plan figures rather than a common hourly comparison. |
| Azure Machine Learning | ND96asr A100 v4; eight A100 GPUs | Displayed savings-plan totals | $16,512.42 per month for one year; $9,946.77 per month for three years | These are the page’s displayed plan scenarios, not proof of a universal Azure discount or a cross-provider break-even point. |
| Azure Databricks, US example | NC24ads A100 v4 | Pay-As-You-Go cluster total including VM compute and Databricks unit (DBU) charges | $7,500.75 per month | It is a managed-service total, not a standalone Azure VM price. |
| Google Cloud Spot VM pricing | A3 High; one-GPU H100 Spot configuration shown on the page | Spot whole-VM hourly total | $6.307365068 per hour | It is a Spot example whose machine configuration and region matter; it is not an on-demand H100 rate or a general H100 price. |
| AWS EC2 accelerated-computing documentation | P5.4xlarge with one H100; P5.48xlarge with eight H100 GPUs; P5e configurations with eight H200 GPUs are also listed | Configuration documentation, not a price quote | Price not stated on the cited configuration page | AWS cannot be ranked against the displayed Azure and Google figures without a matching quote. |
The Azure figures were displayed on official pricing pages accessed on October 25, 2026. Dynamic pricing pages do not provide an authoritative publication date for each number, so treat every amount as a snapshot to recheck immediately before committing money.
#1 Best Overall
What must be matched before comparing GPU prices?
A useful comparison treats a GPU VM as a complete operating cost, not as a chip rental. Match these dimensions before calculating a winner:
- Region and currency: select the same geographic region, or document why a different region is required. Taxes, local rates, and availability can change the result.
- GPU model, count, and memory: an H100, H200, A100, or another accelerator is not interchangeable, and one GPU is not equivalent to eight.
- Machine shape: record vCPUs and RAM as well as the GPU. A workload can be CPU-, memory-, storage-, or network-bound even when the GPU is the most visible component.
- Billing model: label on-demand, Spot, a savings plan, a reserved commitment, or a capacity reservation separately.
- Cost scope: state whether the number covers only compute, attached GPUs, a complete VM, a managed platform fee, storage, networking, software, or all of those items.
- Utilization: convert hourly prices into the hours you actually expect to run. A low hourly rate can lose to a higher rate if the cheaper option requires idle capacity or creates interruption recovery work.
Google’s pricing guidance is explicit that an attached GPU adds to the machine-type cost. Some accelerator-optimized machine families instead bundle GPU costs into the machine price. Mixing those two pricing structures is a common source of double counting or omission.
How does AWS GPU pricing work?
AWS presents GPU capacity through EC2 accelerated-computing instances and several purchasing paths. The cited instance documentation identifies configurations rather than rates: P5.4xlarge has one H100, P5.48xlarge has eight H100 GPUs, and P5e configurations include eight H200 GPUs. Those specifications help you choose a comparable shape, but they do not establish what the instance costs in a particular region.
Choose the AWS purchase path first
- On-Demand: pay for capacity without a long-term usage commitment. This is the cleanest baseline for a short or uncertain project, but it is not automatically the lowest rate.
- Savings Plans: commit to a level of usage or spend in exchange for a discount. The commitment is financial; it should not be confused with reserving a particular physical configuration.
- Reserved Instances: use a reservation model when the instance family, region, and term fit a predictable requirement.
- Spot: use spare capacity at a variable price when the workload can tolerate interruption and restart. The operational cost of checkpointing and recovery belongs in the comparison.
- Dedicated Hosts or Dedicated Instances: pay for isolation or tenancy requirements rather than treating them as ordinary shared VM capacity.
- Capacity Reservations: reserve capacity in a location when obtaining the instance matters more than receiving the lowest possible unit rate. A capacity reservation is a capacity guarantee, not the same thing as a usage discount.
For AWS, obtain a current full-VM quote in the selected region and record whether the price includes the GPU, the host machine, and any required software. Do not infer an hourly amount from the P5 configuration names.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
Which Azure GPU VM family fits the workload?
Azure’s N-family is divided by workload as well as accelerator. ND-series machines target deep learning, AI, and large-scale work; NC-series target compute-intensive AI and machine learning, high-performance computing, and rendering; NV-series target graphics-intensive workloads and visualization. The available series include devices such as NVIDIA A100, H100, H200, GB200, and AMD MI300X, but a specific device’s regional availability must be checked before you build a budget.
Azure Machine Learning example
The Azure Machine Learning pricing page displays an ND96asr A100 v4 configuration with eight A100 GPUs at $19,853.810 per month Pay-As-You-Go. The same page displays $16,512.42 per month for a one-year savings-plan scenario and $9,946.77 per month for a three-year scenario. Keep the term beside each number: the figures are not interchangeable hourly rates, and the cited material does not establish a matching AWS or Google configuration.
Azure Databricks example
The Azure Databricks pricing page’s US example shows an NC24ads A100 v4 cluster at $7,500.75 per month Pay-As-You-Go. That total includes both VM compute and Databricks unit (DBU) charges. It is useful for estimating a managed analytics or machine-learning platform, but it must not be placed in a table beside bare Azure VM prices as if both numbers covered the same service.
How does Google Cloud charge for GPUs?
Google Cloud uses two different patterns. For a conventional machine with an attached accelerator, the GPU charge is added to the machine-type cost. Accelerator-optimized families can bundle the GPU into the machine price. Your quote must therefore identify which pattern applies before you compare it with an AWS or Azure VM.
Google also states that GPU resources can qualify for resource-based committed-use discounts when the applicable reservation requirements are met. Spot GPU resources receive Spot pricing but do not receive sustained-use discounts. Spot prices can change, so a budget should model more than one utilization and price scenario.
The published Spot example
The Google Cloud Spot VM pricing page displays a whole-VM total of $6.307365068 per hour for the listed one-GPU A3 High H100 Spot configuration. This is not an on-demand rate, and it is not a universal H100 price. Record the machine shape and region shown on the page when you use the example; otherwise the decimal precision creates false confidence.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why can’t these published figures be ranked directly?
They answer different financial questions. Azure Machine Learning presents a monthly eight-GPU machine scenario, Azure Databricks presents a monthly managed-cluster total, and Google presents a Spot whole-VM hourly total. AWS’s cited material supplies instance configurations and purchasing mechanics but no matching price. A direct sort from lowest to highest would compare unlike units, terms, and service layers.
For a defensible three-way quote, create one worksheet with these columns:
Rank #4
| Comparison field | Entry to capture |
|---|---|
| Provider and region | Exact cloud, region, and currency used by the calculator |
| Accelerator | GPU model, number of GPUs, and memory specification |
| VM shape | Instance or machine type, vCPU count, and RAM |
| On-demand total | Complete hourly VM total, with attached-GPU treatment documented |
| Spot total | Whole-VM hourly price, interruption policy, and recovery assumptions |
| Commitment scenario | Term, required reservation, payment assumption, and resulting effective rate |
| Additional charges | Attached storage, network egress, operating system or software, and managed-platform fees |
| Evidence record | Source page name and the date and time the price was checked |
Do Spot prices or commitments make a difference?
Yes, but they change both price and risk. Spot is appropriate only when the workload can be interrupted, checkpointed, and restarted. A commitment can lower the effective unit cost for steady usage, but it creates a term obligation and may require a reservation or minimum spend.
Model Spot as a reliability choice
Estimate the cost of lost progress, checkpoint storage, restart time, and engineering effort. Compare that expected cost with the on-demand alternative rather than treating the Spot discount as free savings. For training jobs, write checkpoints frequently enough that an interruption does not erase an unacceptable amount of work.
Calculate commitment break-even
For each term, multiply the on-demand hourly rate by expected monthly utilization, then compare it with the commitment’s effective monthly cost and any unused-capacity exposure. Run the calculation at low, expected, and high utilization. A three-year figure can look attractive at full utilization while costing more than on-demand when a project ends early.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should you calculate the monthly budget?
Start with the complete VM rate, not the GPU line item. Multiply by planned run hours, then add recurring storage and software charges and expected network costs. For a managed service, add the platform fee shown by that service; for a bare VM, do not silently substitute a managed-service total.
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
Use separate scenarios for:
- Development: intermittent daytime use, with shutdowns when the GPU is idle.
- Production: the availability level and capacity reservation needed to meet a service objective.
- Batch or training: checkpointing, restart time, and the percentage of work suitable for Spot.
- Commitment: the utilization level at which a one- or three-year term remains economical.
Keep one-time setup, recurring compute, and variable egress in separate lines. That makes it possible to change regions or purchase terms without rebuilding the entire estimate.
What mistakes make a GPU comparison misleading?
- Comparing a GPU add-on with a complete VM: add the machine-type charge when the provider prices the GPU separately.
- Mixing managed and unmanaged services: Databricks DBU-inclusive totals cannot be presented as Azure VM rates.
- Calling Spot an on-demand discount: interruption exposure and changing prices are part of the product.
- Ignoring GPU count: an eight-GPU ND96asr scenario is not comparable with a one-GPU A3 High example.
- Leaving out the region: a price without a region and currency cannot be reproduced reliably.
- Using excessive decimal precision as certainty: the Google Spot example’s many decimal places do not remove volatility or configuration dependence.
- Assuming a commitment guarantees capacity: a usage discount, a reserved configuration, and a capacity reservation solve different problems.
Which provider should a buyer choose?
Choose the provider and purchase model that match your workload and financial tolerance, then verify the live quote. AWS is a fit when you need several purchasing mechanisms and a choice between flexible Spot usage, commitments, reservations, or dedicated capacity. Azure is attractive when the workload aligns with ND, NC, or NV families and when a managed Azure Machine Learning or Databricks workflow is part of the requirement. Google Cloud is compelling when its machine families, regional availability, attached-GPU or accelerator-optimized pricing, and Spot or committed-use terms fit the workload.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Do not select on a provider-wide slogan such as “cheapest GPU.” Select a specific region, GPU count, VM shape, term, and cost scope; save the quote; and repeat the check immediately before purchase because these prices and availability can change.
Frequently Asked Questions
How often should I refresh a cloud GPU quote?
Refresh it immediately before procurement and record the access date and time. The official pricing pages used for these examples do not provide a single authoritative publication date for every displayed figure, so a saved quote is more useful than relying on an old screenshot.
Should finance compare monthly or hourly GPU prices?
Use both, but derive them from the same utilization assumption. Hourly figures expose the provider’s unit rate, while a monthly model reveals idle capacity, commitment obligations, and recurring platform charges.
What should I send to an engineering team before requesting quotes?
Provide the GPU model or performance target, number of GPUs, minimum vCPU and RAM, required region, expected run hours, interruption tolerance, storage and network needs, and whether a managed platform is required. Without those inputs, providers can return prices for materially different machines.
Why does the number of GPUs matter so much in a quote?
GPU count changes both capacity and the rest of the machine shape needed to feed those accelerators. An eight-GPU training configuration should therefore be compared with another eight-GPU configuration, not with a single-GPU Spot example.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
