Cloud bills do not read the way marketing pages promise. A company picks a GPU instance because the sticker price looks reasonable, then the invoice shows egress charges, storage fees, and a minimum node size nobody mentioned at signup. That gap between advertised rate and real bill is why the GPU cloud providers debate has become the most contested pricing fight in cloud computing heading into late 2026.
Amazon Web Services, Microsoft Azure, and Google Cloud still control the bulk of enterprise infrastructure spending, but the rules around what each charges for compute, storage, Kubernetes, and GPU access have shifted hard this year. Synergy Research Group’s Q2 2026 tracking puts AWS at 28% of worldwide cloud infrastructure revenue, Microsoft at 20%, and Google at 15%, with the overall market hitting roughly $143.4 billion for the quarter, up 43% year over year. Growth that fast usually means pricing experiments, and 2026 has delivered plenty: new per-GPU reservation tiers, fixed-rate Direct Connect pricing, and a widening spread between hyperscaler and specialist-cloud GPU rates.
This comparison walks through what AWS, Azure, and Google Cloud actually charge in October 2026, using public list prices checked this month, independent pricing trackers, and the hyperscalers’ own rate cards. It covers compute, storage, GPU rental, Kubernetes, serverless, egress, and the discount programs that change the final number by 25% to 65%. It also covers five real deployment scenarios, a migration checklist for moving a workload between clouds, and a verdict based on the numbers rather than marketing copy.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The 2026 Cloud Market at a Glance
Before comparing line-item pricing, it helps to see how the three platforms differ structurally. Each one packages compute, storage, and managed services differently, and those packaging choices explain a lot of the price gaps that show up later in this piece. AWS leans on breadth and a mature marketplace. Azure leans on enterprise licensing and hybrid tooling. Google Cloud leans on data infrastructure and per-second billing granularity.
| Attribute | AWS | Microsoft Azure | Google Cloud |
|---|---|---|---|
| Q2 2026 market share | 28% | 20% | 15% |
| Flagship compute | Amazon EC2 | Azure Virtual Machines | Compute Engine |
| Object storage | Amazon S3 | Azure Blob Storage | Cloud Storage |
| Managed Kubernetes | Amazon EKS | Azure Kubernetes Service (AKS) | Google Kubernetes Engine (GKE) |
| Serverless functions | AWS Lambda | Azure Functions | Cloud Run functions |
| Managed PostgreSQL | Amazon RDS for PostgreSQL | Azure Database for PostgreSQL | Cloud SQL for PostgreSQL |
| Generative AI platform | Amazon Bedrock | Microsoft Foundry | Gemini platform |
| Flagship 8-GPU H100 node | p5.48xlarge | ND96isr_H100_v5 | a3-highgpu-8g |
| H100 on-demand rate (per GPU) | $6.88 | $12.29 | from $11.06 |
| Baseline 4-vCPU/16GB VM (hourly) | $0.192 | $0.192 | $0.218756 |
| Internet egress (first tier) | $0.09/GB after 100GB free | ~$0.087/GB after 5TB free (East US) | $0.08-$0.12/GB |
| Managed Kubernetes control-plane fee | Varies by cluster version/support | Free tier available | Flat $0.10/cluster-hour |
Two things stand out immediately. First, AWS and Azure price their baseline general-purpose VM identically at $0.192 per hour for a 4-vCPU, 16GB configuration, while Google Cloud’s comparable n2-standard-4 runs about 14% higher at $0.218756 per hour, based on public list rates checked October 16, 2026. Second, the GPU story flips that pattern: AWS is the cheapest hyperscaler for single-GPU H100 access, while Azure charges nearly double for the same chip. The rest of this piece unpacks why.
Compute Pricing Compared: The Baseline VM Bill
Every cloud comparison should start with the boring workload: a general-purpose virtual machine running Linux, no GPU, no premium networking. That baseline reveals how close AWS and Azure keep their general compute pricing, and how a monthly bill grows once storage, networking, and a public IP get added.
| Provider and instance | Region | List rate / hour | 730-hour monthly compute |
|---|---|---|---|
| AWS m6i.xlarge | us-east-1 | $0.192 | $140.16 |
| Azure D4s v5 | East US | $0.192 | $140.16 |
| Google Cloud n2-standard-4 | us-east4 | $0.218756 | $159.69 |
Compute alone rarely equals the final invoice. Adding a 100 GiB gp3 volume, a public IPv4 address, and 500 GB of outbound traffic to the AWS m6i.xlarge example turns a $140.16 compute line into a $187.81 total, a jump of roughly 34%, according to a usage.ai breakdown of AWS’s own published rates. That gap is the reason FinOps teams treat the advertised hourly rate as a starting point rather than a forecast. Storage, IP allocation, and outbound data transfer each add their own line item, and none of them show up in the headline per-hour number vendors publish.
| Charge | Calculation | Estimate |
|---|---|---|
| EC2 compute | $0.192 x 730 hours | $140.16 |
| EBS gp3 storage | 100 GiB x $0.08/GiB-month | $8.00 |
| Public IPv4 address | $0.005 x 730 hours | $3.65 |
| Outbound internet transfer | (500GB – 100GB free) x $0.09 | $36.00 |
| Total | $187.81 |
Azure and Google Cloud build nearly identical bills around their own equivalents of EBS, static IP charges, and metered egress, just with different per-unit rates. The takeaway for budgeting purposes: never price a workload off compute alone. A 100GB database volume and a few hundred gigabytes of monthly egress commonly add 25% to 40% on top of the advertised VM rate on every major cloud, a pattern that holds whether the workload lives on AWS, Azure, or Google Cloud.
GPU Cloud Pricing Compared: H100, A100, and B200 Rates
GPU rental is where AWS, Azure, and Google Cloud diverge the most, and where the broader GPU cloud providers market, including specialist clouds like RunPod, Lambda, and Vast.ai, undercuts the hyperscalers by a wide margin. As of September 2026, cloud GPU rates span from $0.14 per hour for an entry-level card on a marketplace platform to $12.29 per hour for a reserved Azure H100, according to a Hostinger pricing comparison covering seven GPU models across seven providers.
| GPU | VRAM | AWS | Azure | Google Cloud | Cheapest specialist cloud |
|---|---|---|---|---|---|
| RTX PRO 6000 | 96 GB | – | – | ~$4.50 (G4, 8-vGPU bundle) | $0.60 (Hostinger) |
| A100 80GB | 80 GB | ~$3.43 (p4de, per GPU) | $3.67 (NC24ads A100 v4) | $3.67 (a2-highgpu-1g, 40GB) | $1.39 (RunPod) |
| H100 80GB | 80 GB | $6.88 (p5.4xlarge) | $6.98 single GPU / $12.29 per GPU on ND H100 v5 node | from $11.06 (a3-highgpu-8g, 8-GPU minimum) | $2.89 (RunPod) |
| B200 192GB | 192 GB | – | – | ~$16.11 (A4, reservation required) | $4.50 (Hostinger) |
Within the three hyperscalers alone, AWS is consistently the cheapest on-demand H100 option at $6.88 per hour for a single GPU on a p5.4xlarge instance. Azure’s single-GPU NC H100 v5 rate runs slightly higher at $6.98, but its eight-GPU ND H100 v5 node prices each GPU at $12.29, or $98.32 for the full node. Google Cloud’s A3 instance (a3-highgpu-8g) starts around $11.06 per GPU but forces an eight-GPU minimum on-demand, so a team that needs two H100s still pays for eight. That minimum-node requirement, not the per-GPU rate itself, is often the real cost driver for small AI teams evaluating GPU cloud providers.
The A100 80GB tells a similar story. AWS’s p4de instance works out to roughly $3.43 per GPU, Azure’s NC24ads A100 v4 runs $3.67 for a single GPU, and Google Cloud’s a2-highgpu-1g, which actually ships the 40GB variant rather than the 80GB card, lists at $3.67 as well. Specialist clouds price the identical 80GB A100 as low as $1.39 per hour on RunPod, a gap of roughly 2.6x versus the cheapest hyperscaler rate. For teams comparing Blackwell-generation pricing against the Hopper-generation H100, the B200 shows the same hyperscaler premium: Google Cloud’s A4 series starts around $16.11 per GPU-hour but requires a reservation, with no standard on-demand tier available.
Why Hyperscaler GPUs Cost More Than Specialist Clouds
The price gap between AWS, Azure, Google Cloud, and specialist GPU clouds is not really about the chip. It comes down to three structural differences: forced bundling, quota lead time, and what the premium actually buys.
Forced Bundling
Hyperscalers rent a whole virtual machine, not a bare GPU. AWS’s p5.4xlarge pairs one H100 with 16 vCPUs and 256 GB of system RAM. Azure’s ND H100 v5 bundles eight GPUs with 96 vCPUs and 1.9 TB of RAM. Neither AWS nor Azure lets a customer rent just the accelerator; the processors and memory arrive attached in fixed ratios whether the workload needs them or not. Specialist clouds strip that bundle down, which is a core reason their effective per-GPU rate runs lower even on identical silicon.
Quota Lead Time
AWS and Google Cloud cap how many high-end GPU instances a new account can launch, and that quota defaults to zero. Getting AWS P5 access approved typically requires a written justification and takes three to seven business days; Google Cloud’s A3 H100 quota approval can take longer still. That delay carries its own cost: teams blocked on quota often rent available capacity elsewhere rather than wait, even when the available option costs more per hour.
What the Premium Buys
The hyperscaler rate is not pure markup. AWS’s P5 instances support networking up to 3.2 terabits per second between GPU nodes, which matters for distributed training jobs spanning dozens of GPUs. All three platforms bundle enterprise compliance certifications, global regional coverage, and support tiers that most inference or fine-tuning jobs never touch. For a 64-plus GPU training cluster or a regulated workload, that infrastructure is difficult to assemble anywhere else. For a single GPU running inference or image generation, most of that bundle goes unused, and it shows up in the final bill anyway.
GPU Generations Compared: Hopper vs Blackwell Pricing Trajectory
The H100 and the newer B200 sit on different NVIDIA architectures, Hopper and Blackwell respectively, and the price gap between them tells a story about where GPU cloud providers are headed next. The B200 carries 192 GB of memory against the H100’s 80 GB, more than double the capacity, yet Hostinger’s September 2026 comparison shows it pricing at roughly the same multiple over the H100 across most providers rather than scaling linearly with memory. That suggests the market is pricing Blackwell GPUs on throughput and availability constraints rather than a simple per-gigabyte memory premium.
| Generation | Chip | VRAM | Cheapest specialist rate | Cheapest hyperscaler rate |
|---|---|---|---|---|
| Hopper | H100 | 80 GB | $2.89/hr (RunPod) | $6.88/hr (AWS) |
| Hopper | H200 | 141 GB | $3.87/hr (Vast.ai) | Reservation-only on Azure/Google Cloud |
| Blackwell | B200 | 192 GB | $4.38/hr (Vast.ai) | ~$16.11/hr (Google Cloud A4, reservation required) |
The H200, a memory-upgraded version of the H100 with 141 GB of HBM3e versus the H100’s 80 GB, sits between the two generations. It carries roughly 75% more VRAM than the H100 but keeps the same underlying compute throughput, making it a fit for larger model contexts rather than faster training. Neither the H200 nor the B200 is available as a standard on-demand option on Google Cloud or Azure as of this comparison; both require a reservation, which pushes smaller teams back toward specialist clouds or the older H100 if they need immediate access without a committed term. As B200 and H200 supply widens through 2027, the expectation across pricing trackers is that H100 on-demand rates keep drifting down, following the same pattern A100 pricing showed after H100 capacity became widely available, when A100 rates dropped below $1 per hour on several marketplace platforms.
Storage Pricing Compared: S3 vs Blob vs Cloud Storage
Object storage pricing stays close across the big three, but the gap widens once block storage for GPU and compute instances enters the picture. On the object storage side, AWS S3 Standard lists at approximately $0.023 per GB-month for the first 50 TB, versus roughly $0.020 for Google Cloud Storage Standard and about $0.0184 for Azure Blob Hot storage, putting Azure as the cheapest of the three for straightforward object storage at small scale.
| Storage type | AWS | Azure | Google Cloud |
|---|---|---|---|
| Object storage (standard tier, first 50TB) | $0.023 | $0.0184 (Blob Hot) | $0.020 |
| Block storage attached to a GPU/compute VM | $0.08 (EBS gp3) | $0.17 (Premium SSD) | $0.04-$0.17 (standard to SSD persistent disk) |
Block storage attached to a running GPU or compute instance is where the numbers spread further apart. AWS’s EBS gp3 volumes run about $0.08 per GB-month, Google Cloud’s persistent disks range from $0.04 for standard disks up to $0.17 for SSD-backed volumes, and Azure’s Premium SSD sits at roughly $0.17 per GB-month. On GPU-specific specialist clouds, storage billing behaves differently again: RunPod charges $0.10/GB-month while an instance runs and $0.20/GB-month once stopped, and Vast.ai keeps billing storage continuously even when an instance is stopped, with deleting the instance as the only way to stop the charge. A 200GB volume left idle for a month on RunPod costs $40, a detail that catches teams off guard when they pause a GPU job without decommissioning the attached storage.
Data Egress Fees Compared: The Hidden Line Item
Egress, the cost of moving data out of a provider’s network, is the line item that catches most teams off guard, because it rarely appears in the headline comparison between AWS, Azure, and Google Cloud. AWS charges $0.09 per GB for the first 10 TB after a 100 GB monthly free allowance. Google Cloud’s egress runs $0.08 to $0.12 per GB depending on destination and volume. Azure charges roughly $0.087 per GB, with the first 5 TB free when transferring out of East US specifically.
| Provider | Egress rate | Free allowance |
|---|---|---|
| AWS | $0.09/GB | First 100 GB/month free |
| Google Cloud | $0.08-$0.12/GB | Varies by destination/volume |
| Azure | ~$0.087/GB | First 5TB free from East US |
| Specialist GPU clouds (Hostinger, RunPod Pods, Lambda) | $0 | No egress fees on these platforms |
The scale matters more than the rate. Moving a 140 GB Llama-70B model checkpoint out of AWS costs roughly $12.60 at list price. Moving a 10 TB training dataset costs around $900. Teams that routinely move large model checkpoints or training sets between a cloud and local infrastructure, or between two different clouds, should treat egress as a recurring operating cost rather than a one-time fee. AWS introduced fixed monthly pricing for dedicated 10 Gbps and 100 Gbps Direct Connect links across five geographic tiers this year, replacing the per-gigabyte model for those specific dedicated connections, a change aimed directly at large enterprises moving consistent, predictable volumes of data.
Kubernetes Pricing Compared: EKS vs AKS vs GKE
All three hyperscalers run a managed Kubernetes service: Amazon EKS, Azure Kubernetes Service, and Google Kubernetes Engine. Pricing structures differ in ways that matter for teams running many small clusters versus a handful of large ones. Google Kubernetes Engine charges a flat $0.10 per cluster-hour for both its Standard and Autopilot cluster management, a predictable rate regardless of cluster size. Azure offers a free tier for AKS cluster management on the Free pricing tier, with Standard tier pricing applying once uptime and support guarantees matter. AWS prices EKS control-plane fees per cluster hour with variation tied to cluster version support and add-ons.
| Service | AWS EKS | Azure AKS | Google GKE |
|---|---|---|---|
| Control plane fee | Per cluster-hour, varies by tier | Free tier available; Standard tier adds SLA/uptime | Flat $0.10/cluster-hour (Standard and Autopilot) |
| Node management | Self-managed or managed node groups | Self-managed or AKS-managed node pools | Standard (self-managed nodes) or Autopilot (fully managed) |
| Billing granularity | Per-second, 1-minute minimum | Per-second, 1-minute minimum | Per-second, 1-minute minimum |
For teams already running clusters on AWS, a step-by-step EKS setup walkthrough is worth reviewing before estimating a multi-cluster bill, since node group configuration choices affect the compute cost far more than the EKS control-plane fee itself. The same applies on Google Cloud, where an Autopilot cluster bundles per-pod billing into the node cost, versus a Standard GKE cluster where the team manages node pools directly; a GKE setup guide walks through that choice in more detail. The practical difference between the three services matters less for the control-plane fee itself, which is small relative to compute, and more for how each handles node autoscaling, spot or preemptible node pools, and multi-zone resilience, all of which affect the real monthly compute bill sitting underneath the cluster.
Serverless Pricing Compared: Lambda, Functions, and Cloud Run
The serverless comparison set remains consistent across providers: AWS Lambda and Fargate, Azure Functions and Container Apps, and Google Cloud Run and Cloud Functions. AWS has kept its serverless announcements active through the back half of 2026, with regular updates covering Lambda runtime improvements and event-source integrations. All three platforms bill serverless compute on a per-invocation-plus-duration model, typically metered in milliseconds, which makes serverless the cheapest entry point for spiky or low-traffic workloads across any of the three clouds.
The practical difference between the three shows up in cold-start behavior and maximum execution time rather than headline pricing. Cloud Run functions, built on Google’s container-first Cloud Run product, allow longer execution windows than a typical function-as-a-service platform, which suits occasional batch jobs better than AWS Lambda’s shorter default timeout. Azure Functions integrates tightly with the rest of the Azure ecosystem, including Azure AD-based identity and Logic Apps, which matters for teams already standardized on Microsoft tooling. For a pure cost comparison, all three remain priced close enough that the deciding factor is usually integration depth with the rest of a team’s existing cloud footprint rather than the per-million-invocations rate.
Reserved, Spot, and Committed-Use Discounts Compared
List prices are the ceiling, not the floor. Every major cloud offers a discount program for workloads that can tolerate either a long-term commitment or the risk of interruption, and the size of the discount varies more than most teams expect.
| Provider | Spot/interruptible model | Typical discount | Eviction notice |
|---|---|---|---|
| AWS | Spot Instances | ~57% | 2 minutes |
| Google Cloud | Spot VMs | 60-91% | 30 seconds |
| Azure | Spot VMs | ~80% | 30 seconds |
| Provider | Discount range | Term options |
|---|---|---|
| AWS | 25%-45% | EC2 Capacity Blocks (variable length), Compute/EC2 Instance Savings Plans |
| Google Cloud | Up to 65% | 1 or 3-year Committed Use Discounts |
| Azure | Up to 63% | 1 or 3-year Reserved Instances; 5-year available on ND H100 v5 |
Spot pricing is a market rate, not a guaranteed discount. On bid-based capacity, a supply crunch can push spot rates close to or above on-demand pricing, so teams should check live rates rather than assume the published discount holds at the moment of launch. Spot capacity also only makes financial sense with checkpointing built into the job, since an evicted run with no saved progress can end up costing more than an on-demand run would have, regardless of which cloud it ran on. Google Cloud’s 65% ceiling on committed-use discounts is the largest of the three for predictable, long-running workloads, but it requires locking in either a specific resource configuration or a spend commitment for one to three years, which only pays off if usage actually holds steady across the term.
Billing Granularity: Why the Same Job Costs Different Amounts
One detail that rarely appears in pricing pages but materially affects short or interrupted jobs is billing granularity. AWS, Azure, and Google Cloud all bill compute per second with a one-minute minimum, which means a short job pays close to its actual runtime on any of the three hyperscalers. Specialist GPU clouds vary more widely: RunPod and Vast.ai bill per second with no minimum, while Lambda bills per minute. A 16-minute job on RunPod costs roughly 27% of the hourly rate; the identical job on Lambda costs the full hour, since Lambda’s per-minute billing still rounds up to the nearest full-minute unit at the end of a session.
Idle time is the other trap. A GPU instance left running after a job finishes bills at the full hourly rate on every provider in this comparison, hyperscaler or specialist. There is no discount for an idle accelerator anywhere in the market as of late 2026. A session forgotten overnight costs a full night’s compute regardless of which of the three major clouds is hosting it; billing granularity only changes how the edges get rounded, not whether idle time gets charged at all.
Real-World Monthly Bill: From VM Rate to Final Invoice
To make the pricing concrete, consider a mid-sized workload: a web application backend running on a single general-purpose VM, with a database volume and moderate outbound traffic, no GPU involved. This is the kind of workload most small and mid-sized engineering teams actually run most of the time.
A quick way to pull current list prices without relying on marketing pages is the AWS Price List API, which returns machine-readable rate data directly from AWS rather than a cached comparison page:
aws pricing get-products \
--service-code AmazonEC2 \
--filters "Type=TERM_MATCH,Field=instanceType,Value=m6i.xlarge" \
"Type=TERM_MATCH,Field=location,Value=US East (N. Virginia)" \
"Type=TERM_MATCH,Field=operatingSystem,Value=Linux" \
--region us-east-1 \
--max-items 1Running that query, or its Azure Retail Prices API and Google Cloud Billing Catalog API equivalents, is the only reliable way to confirm a rate before committing a budget, since GPU cloud pricing updates monthly and sometimes weekly, and marketplace-style specialist clouds can change hourly. For the baseline VM workload described above, the fully loaded AWS estimate (compute, storage, a public IP, and 500 GB of egress) landed at $187.81 per month against a $140.16 compute-only quote, a 34% increase once the secondary charges are included. Azure and Google Cloud build comparably structured bills; the exact multiplier shifts with regional egress allowances and storage tier choices, but the pattern of compute plus 25% to 40% in secondary charges holds across all three platforms.
Benchmarks: What Three Independent Trackers Found
Three separate pricing trackers published data in the September-to-October 2026 window, and their findings largely agree on direction even where exact figures differ slightly by snapshot date.
- Cloud World Model’s September 2026 snapshot tracked 33 rates across AWS, Google Cloud, Azure, Oracle Cloud Infrastructure, and DigitalOcean, and found no broad list-price increases among any of them that month, meaning the rising bills many teams reported came from usage growth and secondary charges rather than rate hikes.
- Hostinger’s GPU pricing comparison, covering seven GPU models across seven providers including all three hyperscalers, confirmed the 4x-plus spread between the cheapest specialist cloud and the most expensive hyperscaler rate for an identical H100 chip, and documented the spot-discount and reserved-discount ranges used throughout this piece.
- Usage.ai’s cloud provider comparison, checked October 16, 2026, supplied the baseline VM pricing, the $187.81 monthly bill walkthrough, and the Q2 2026 market-share figures sourced from Synergy Research Group.
Taken together, these three trackers point to the same conclusion from different angles: headline list prices on AWS, Azure, and Google Cloud have been largely stable in 2026, but real bills keep climbing because of usage growth, GPU demand pulling in higher-tier hardware, and secondary fees like egress and storage that a flat-rate comparison misses entirely. That pattern echoes commentary from cloud cost analysts tracking the same trend, who have described bills increasing even without a formal price hike on any published rate card.
5 Real-World Use Cases and Which Cloud Fits
Pricing alone rarely settles a cloud decision. Here is how the numbers above play out across five common deployment scenarios.
1. An AI Startup Fine-Tuning a 70B Model
A small team fine-tuning a 70-billion-parameter open model needs A100 or H100 access for days or weeks, not months, and doesn’t need enterprise compliance tooling. At $1.39-$1.43 per hour for an A100 80GB on a specialist cloud, versus $3.43-$3.67 per hour across AWS, Azure, and Google Cloud, the specialist-cloud route cuts training cost by more than half. The tradeoff is losing hyperscaler-grade networking and support, which rarely matters below a handful of GPUs.
2. An Enterprise Running Windows Server and SQL Server Workloads
A company with existing Microsoft licensing, Active Directory, and SQL Server deployments gets the most value from Azure, where license mobility and hybrid-benefit pricing can offset the platform’s higher GPU rates. The decision here is rarely about raw compute price; it is about whether existing Microsoft enterprise agreements translate into Azure credits or discounted licensing, which they do not do automatically and need to be checked case by case.
3. A Data Analytics Team Built Around BigQuery
A team whose core workload already lives in BigQuery gains the most by keeping compute on Google Cloud, since moving data between BigQuery and an external compute cluster on AWS or Azure adds its own egress cost on top of the compute bill. Google Cloud’s flat $0.10/cluster-hour GKE pricing also keeps container orchestration costs predictable for teams running analytics pipelines alongside BigQuery.
4. A Gaming Studio Running Dedicated Game Servers
A studio running dedicated multiplayer game servers at global scale typically defaults to AWS, given its combination of EC2 Spot pricing (roughly 57% off on-demand), broad regional coverage, and EKS-based orchestration tooling. The 2-minute eviction notice on AWS Spot Instances, longer than Azure’s or Google Cloud’s 30-second window, also gives game server fleets more buffer time to drain player sessions before an interruption.
5. A Regulated Fintech Needing Multi-Region Compliance
A fintech company subject to regional data-residency rules needs the regional coverage and compliance certifications that only come bundled into the hyperscaler premium discussed earlier in this piece. For this use case, the lower specialist-cloud GPU rate is irrelevant, since the workload needs the certifications, audit tooling, and regional guarantees that AWS, Azure, and Google Cloud build into their enterprise tiers, and that most specialist GPU clouds simply do not offer.
Migration Guide: Moving a Workload Between Clouds
Switching a workload from one cloud to another, or splitting it across two, is rarely a weekend project. Here is the sequence that keeps a migration from blowing past its budget or its timeline.
- Inventory the current bill line by line. Separate compute, storage, egress, and managed-service fees so the comparison target is the full bill, not just the compute rate.
- Match service equivalents. Use the core-services table earlier in this piece to map EC2 to Azure VMs or Compute Engine, S3 to Blob Storage or Cloud Storage, and EKS to AKS or GKE.
- Check regional availability for every service needed, not just the primary compute instance. A provider can operate in a region without offering every service there.
- Benchmark a representative workload on the target cloud before committing, using matched vCPU and memory configurations, since equal specs do not guarantee equal performance.
- Price egress for the migration itself. Moving a large dataset off the current provider incurs the same per-GB egress charges discussed earlier, and that one-time cost belongs in the migration budget.
- Review quota requirements early if the workload needs GPU instances, since AWS P5 and Google Cloud A3 quota approvals can take three to seven business days or longer.
- Run both environments in parallel for a defined cutover window rather than switching all traffic at once, to catch configuration drift before it affects production.
- Re-evaluate reserved or committed-use pricing only after usage stabilizes on the new platform, since locking in a 1-to-3-year discount before confirming real usage patterns can erase the savings it was meant to deliver.
- Decommission the old environment’s storage explicitly, since stopped instances on several providers continue billing for attached storage until it is deleted, not just stopped.
Teams running Kubernetes-based workloads specifically should treat the migration as two separate projects: moving the cluster control plane and moving the underlying node pools, since node pool pricing, especially for GPU-backed node pools, dominates the bill far more than the EKS, AKS, or GKE control-plane fee itself.
Pros and Cons of AWS, Azure, and Google Cloud
AWS
- Largest market share at 28%, meaning the widest ecosystem of third-party tools and documented integration patterns.
- Cheapest hyperscaler on-demand rate for single-GPU H100 access among the three platforms covered here.
- Fixed-rate Direct Connect pricing introduced in 2026 simplifies budgeting for large, predictable data-transfer needs.
- Reserved and Savings Plan discounts top out lower than Google Cloud or Azure, at 25%-45% versus up to 63%-65% elsewhere.
- GPU quota approval for P5 instances takes three to seven business days, slower than simply renting from a specialist cloud.
Microsoft Azure
- Best fit for organizations with existing Microsoft identity, Windows, or SQL Server licensing.
- Cheapest Blob Hot storage tier among the three platforms at roughly $0.0184/GB-month.
- Highest spot-instance discount of the three hyperscalers at roughly 80% off on-demand.
- Most expensive single-GPU H100 rate among the three at $12.29 per GPU on the ND H100 v5 node.
- Azure Arc extends hybrid management to resources outside Azure, useful for mixed on-premises and cloud estates.
Google Cloud
- Highest committed-use discount ceiling of the three at up to 65% for 1-to-3-year terms.
- Flat, predictable $0.10/cluster-hour GKE pricing regardless of cluster size, simpler to forecast than per-tier competitors.
- Best fit for teams already standardized on BigQuery or other Google Cloud data services.
- Highest baseline general-purpose VM rate of the three at $0.218756/hour versus $0.192 for AWS and Azure.
- A3 H100 instances carry an eight-GPU on-demand minimum, forcing small teams to pay for capacity they don’t need.
Full Pricing Comparison Table
This table consolidates the key figures from every section above into a single reference, using list prices checked between September and October 2026.
| Category | AWS | Azure | Google Cloud |
|---|---|---|---|
| Market share (Q2 2026) | 28% | 20% | 15% |
| Baseline 4vCPU/16GB VM | $0.192/hr | $0.192/hr | $0.218756/hr |
| H100 on-demand (single GPU) | $6.88/hr | $6.98/hr single / $12.29/hr per GPU on 8-GPU node | from $11.06/hr (8-GPU minimum) |
| A100 80GB on-demand | ~$3.43/hr | $3.67/hr | $3.67/hr (40GB variant) |
| Object storage (standard) | $0.023/GB-mo | $0.0184/GB-mo | $0.020/GB-mo |
| Block storage | $0.08/GB-mo | $0.17/GB-mo | $0.04-$0.17/GB-mo |
| Egress (first tier) | $0.09/GB, 100GB free | ~$0.087/GB, 5TB free (East US) | $0.08-$0.12/GB |
| Spot/preemptible discount | ~57% | ~80% | 60%-91% |
| Reserved/committed discount ceiling | 25%-45% | up to 63% | up to 65% |
| Managed Kubernetes control plane | Per cluster-hour, tier-based | Free tier available | Flat $0.10/cluster-hour |
| GPU quota approval time | 3-7 business days (P5) | Not typically quota-gated the same way | Often longer than AWS for A3 H100 |
The Verdict: Which Cloud Wins on Price in 2026
There is no single winner across every category, and that is the honest conclusion the numbers support. AWS wins on GPU on-demand pricing among the three hyperscalers, holding the cheapest single-GPU H100 rate at $6.88 per hour, and it wins on market share and ecosystem breadth with 28% of worldwide cloud infrastructure revenue. Azure wins on storage pricing and spot discounts, with the cheapest Blob Hot tier and the highest spot discount of the three at roughly 80%, but it carries the highest GPU rate per node at $12.29 per GPU. Google Cloud wins on long-term discount ceilings, topping out at 65% for committed-use agreements, and on Kubernetes pricing predictability with its flat $0.10/cluster-hour GKE rate, but its baseline VM pricing and 8-GPU minimum on A3 instances work against teams running small or variable workloads.
For a pure cost-minimization strategy on GPU-heavy workloads, the data points toward a hybrid approach: use AWS, Azure, or Google Cloud for the parts of a stack that need compliance, regional reach, or tight integration with existing managed services, and route GPU-intensive training or inference to specialist GPU clouds where the per-hour rate runs 2x to 4x lower for identical silicon. Teams that need everything under one roof, with compliance tooling and global reach included, will still find AWS the cheapest of the three hyperscalers for most general compute and GPU workloads, with Azure and Google Cloud each winning specific categories worth checking against the exact workload in question before committing a budget.
Frequently Asked Questions
Which cloud is cheapest for GPU compute in 2026?
Among AWS, Azure, and Google Cloud, AWS offers the cheapest on-demand H100 rate at $6.88 per hour for a single GPU. Specialist GPU clouds like RunPod and Vast.ai undercut all three hyperscalers, with H100 rates starting around $2.89 per hour, though they lack the compliance certifications and regional reach the hyperscalers include.
How much does Azure charge for H100 GPUs per hour?
Azure’s NC H100 v5 lists at $6.98 per hour for a single GPU. Its eight-GPU ND H100 v5 node prices each GPU at $12.29, totaling $98.32 for the full node, making it the most expensive per-GPU hyperscaler rate covered in this comparison.
Is Google Cloud cheaper than AWS for Kubernetes?
Google Kubernetes Engine charges a flat $0.10 per cluster-hour for both Standard and Autopilot clusters, which is simpler to forecast than AWS EKS’s tier-based control-plane pricing. The total Kubernetes bill depends far more on node pool compute costs than on the control-plane fee itself on any of the three platforms.
What is the cheapest way to avoid cloud egress fees?
Several specialist GPU clouds, including Hostinger, RunPod’s Pod workloads, and Lambda, charge zero egress fees. Among the three hyperscalers, no platform offers free unlimited egress; AWS, Azure, and Google Cloud each include a limited monthly free allowance before per-GB charges apply.
Do spot instances always cost less than on-demand pricing?
Not guaranteed. Spot and preemptible pricing reflects live market demand for spare capacity, so a supply crunch can push spot rates close to or above the on-demand rate on bid-based platforms. Typical discounts run 57% on AWS, up to 80% on Azure, and 60%-91% on Google Cloud, but the live rate should be checked before assuming the discount holds.
How long does GPU quota approval take on AWS or Google Cloud?
AWS P5 instance access typically requires a written justification and takes three to seven business days for approval. Google Cloud’s A3 H100 quota approval can take longer. Azure does not gate GPU access through the same quota-request process in most regions.
Which cloud has the largest market share in 2026?
AWS leads with 28% of worldwide cloud infrastructure services revenue in Q2 2026, according to Synergy Research Group. Microsoft Azure follows at 20%, and Google Cloud holds 15%, with the remainder split among Oracle Cloud Infrastructure, Alibaba Cloud, IBM Cloud, and smaller regional providers.
Should a startup pick one cloud or go multi-cloud?
Most startups are better served by standardizing on one hyperscaler to avoid the operational overhead of managing two sets of IAM policies, networking rules, and billing dashboards. Multi-cloud setups tend to make more sense once a company has distinct regulatory, data-residency, or GPU-cost reasons to split specific workloads, such as routing GPU training to a specialist cloud while keeping the application layer on a single hyperscaler.
