HYPERSCALER LIST PRICES. SPECIALIST RATES.

AWS, Azure, and Google Cloud GPU pricing.
Against the specialists.

Compare published H100 prices from AWS, Microsoft Azure, and Google Cloud with specialist GPU clouds. See what on-demand costs, what commitments save, and when a specialist is the better buy.

Compare the rates Meet with an advisor

01 / ON-DEMAND LIST PRICES

Eight H100 GPUs.
Three clouds.

Each cloud sells H100 capacity as an eight-GPU instance. These are the published on-demand rates for Linux in each cloud’s main US region, with no commitment.

USD · Linux · On-demandChecked September 21, 2026

Scroll the table horizontally to see configurations and source links.

Published on-demand prices for eight-GPU H100 instances on AWS, Google Cloud, and Microsoft Azure, checked September 21, 2026.
Cloud	Instance and region	Per instance-hour	Per GPU-hour	Official source
AWS	p5.48xlarge: 8 × H100, 192 vCPUs, 2,048 GiB memory. US East (N. Virginia).	$55.04	$6.88	On-Demand pricing
Google Cloud	a3-highgpu-8g: 8 × H100, 208 vCPUs, 1,871 GB memory, 6,000 GiB SSD. Iowa (us-central1).	$88.49	$11.06	Accelerator-optimized pricing
Microsoft Azure	ND96isr H100 v5: 8 × H100 80 GB, 96 vCPUs, 1,900 GiB memory. East US.	$98.32	$12.29	Linux VM pricing

Per GPU-hour is our calculation: the instance rate divided by eight. The Azure figure was read from Microsoft’s Retail Prices API, and the AWS figure from AWS’s published on-demand price data. The instances differ in CPU, memory, storage, and networking, so the hourly rate alone is not a like-for-like comparison. Prices vary by region and change over time.

02 / AZURE H100 PRICING

Azure ND H100 v5
price per hour.

Azure sells H100 as the ND96isr H100 v5 virtual machine: eight H100 80 GB GPUs, 96 vCPUs, and 1,900 GiB of memory. Microsoft does not sell a single H100 GPU in this series. The on-demand list price is $98.32 per hour in East US, or $12.29 per GPU-hour.

USD · Linux · East USChecked September 28, 2026

Scroll the table horizontally to see every column.

Azure ND96isr H100 v5 list prices in East US by purchase option, checked September 28, 2026.
Purchase option	Listed price	Per VM-hour	Per GPU-hour
Pay as you go	$98.32 per hour	$98.32	$12.29
1-year reservation	$551,221 for the term	about $62.92	about $7.87
3-year reservation	$1,134,310 for the term	about $43.16	about $5.40
5-year reservation	$1,722,566 for the term	about $39.33	about $4.92
Spot, can be evicted	$18.17 per hour	$18.17	about $2.27

Lowest listed regions

East US, East US 2, West US 2, and West US 3 all list $98.32 per hour, or $12.29 per GPU-hour.

Other US and Europe

South Central US and Canada Central list about $118, or $14.75 per GPU-hour. West Europe, Sweden Central, and West US list $127.82, or $15.98.

Asia Pacific and Africa

Japan East lists $142.56, or $17.82 per GPU-hour. South Africa West is the highest at $186.61, or $23.33.

Prices are from Microsoft’s Retail Prices API, Linux, before any negotiated discount or Azure savings plan. Per-hour and per-GPU figures for reservations are our calculation, spreading the term price over 8,760 hours a year. A listed price is not a promise of available capacity in that region. Ask Microsoft or your reseller to confirm quota before you plan around it.

03 / WHAT COMMITMENTS SAVE

Lower rates,
with strings attached.

Each cloud publishes cheaper rates for the same instance, in exchange for a commitment, a fixed window, or the risk of interruption. Per GPU-hour figures are our calculation.

AWS

EC2 Capacity Blocks reserve GPU instances for a set future window. The p5.48xlarge lists at $41.528 per instance-hour, about $5.19 per GPU-hour, in US East (N. Virginia). AWS says these prices update regularly with supply and demand, next in October 2026.

Google Cloud

For a3-highgpu-8g in Iowa, a 3-year committed use discount lists at $38.86 per hour, about $4.86 per GPU-hour, and a 1-year commitment at $61.38, about $7.67. Flex-start, where the job waits for capacity, lists at $38.32, about $4.79. Spot lists at $52.96.

Microsoft Azure

For ND96isr H100 v5 in East US, a 1-year reservation lists at $551,221 for the term, about $62.92 per hour or $7.87 per GPU-hour. A 3-year reservation lists at $1,134,310, about $5.40 per GPU-hour. Spot lists at $18.17 per hour, and can be evicted.

Read the fine print on every discount.

A multi-year commitment locks in spend for years. Spot capacity can be taken back mid-job. A queued or scheduled start means you wait for the GPUs. Compare the discounted rate with the terms you would actually accept, and ask whether the commitment guarantees capacity or only lowers the bill.

04 / THE SPECIALIST SIDE

What specialist
GPU clouds publish.

Specialist providers, sometimes called neoclouds, focus on GPU capacity. These on-demand H100 rates come from our GPU cloud pricing comparison, checked September 18, 2026.

Published per GPU-hour

Runpod: $3.49 for an H100 SXM GPU Pod listing
Nebius: $3.85 for an HGX H100 on-demand instance, rising to $4.50 on October 1, 2026
Crusoe: $3.90 for an H100 HGX on-demand instance
Lambda: $4.29 for a single H100 SXM instance
CoreWeave: about $6.16, from $49.24 for an eight-GPU instance

How to read the gap

On demand, every specialist rate above sits below all three hyperscalers. Hyperscaler commitment and reservation rates of about $4.79 to $5.40 per GPU-hour reach specialist on-demand territory, but only with a multi-year commitment, a fixed reservation window, or a queued start. Azure Spot lists lower, at about $2.27 per GPU-hour, but can be evicted.

Configurations and billing units differ between providers. Compare the whole configuration and total cost, not the GPU line alone.

05 / CHOOSING BETWEEN THEM

Price is one reason.
Here are the others.

The cheapest GPU-hour is not always the cheapest project. These factors often decide which side of the table fits.

A hyperscaler may fit when

Your data, applications, and security tooling already live in that cloud
You already hold a spending commitment or credits you need to use
Procurement and compliance have approved that provider
You need GPUs in many regions alongside other managed services

A specialist may fit when

GPU compute is the main cost and you want the lowest published rates
You need multi-node clusters with a fast network between servers
You want shorter terms than a multi-year commitment
Your team can run workloads outside your primary cloud

Moving data between clouds can add transfer charges and time. Check each provider’s data transfer pricing before splitting training from the rest of your stack. Planning multi-node training? See bare metal GPU clusters.

06 / COMPARE QUOTES FAIRLY

Put every quote
on the same basis.

Ask each provider, hyperscaler or specialist, for the same details before comparing totals.

Configuration

GPU model, variant, and memory
Billing unit: per GPU or per instance
CPU, system memory, and local storage included
Region and network between servers

Total cost and terms

Storage, data transfer, and support charges
Commitment length, and whether capacity is guaranteed
Quota or approval needed before launch
Interruption terms for spot or queued capacity

A dated snapshot of published prices, not a live quote. Rates, regions, and availability change. Confirm current pricing and terms with each provider before budgeting.

WHY WORK WITH US

Powered by
Bridgepointe Technologies.

GPU Cloud Advisors is powered by Bridgepointe Technologies, a technology advisory firm since 2002. You get one advisor, the research on this site, and the provider relationships behind it.

Bridgepointe clients include

NVIDIA
Salesforce
Adobe
Uber
Netflix
Airbnb
Roblox
LinkedIn
CrowdStrike
Sony
PayPal
Mastercard

75+ Fortune 500 clients

Including more than 20 of the Fortune 100, with 97% client satisfaction.

1,200+ data center and cloud projects

Requirements assessed, capacity sourced and providers selected.

Terms, not just the rate

We negotiate ramp schedules, early billing and expansion terms, so you aren’t paying for capacity before your workload can use it.

One requirement, every fit

Bridgepointe partners with 380+ technology providers. We take one requirement to the providers that fit it and bring back a side-by-side comparison you can actually decide from.

Case study: Bridgepointe helped Hive, an AI company, choose data center and connectivity providers across 15 projects, saving its team at least 180 hours. Ask us for the case study

LET’S PLAN YOUR GPU PROJECT

Tell us what
you need.

Two short steps. Tell us what you’re building and an advisor will help you evaluate providers, confirm capacity requirements, and plan the support you need to get into production.

Not sure which GPU or how many you need? Start with your workload and what could change.

Our advisory services are free to you. We’re compensated by whichever provider you choose through us.

No obligationSizing help welcome

What we’ll work through with you

A provider shortlist suited to your workload and location.
Initial capacity, expansion needs and delivery timing to confirm.
Total cost, commitment terms and deployment support to compare.

Prefer to talk it through? Book a free 30-minute consultation or call 844-506-2299.

HYPERSCALER GPU PRICING QUESTIONS

Before you
compare bills.

Bring your current bill or an existing quote.
Meet with an advisor

Which big cloud has the lowest on-demand H100 price?

In our September 21, 2026 check, AWS listed the lowest on-demand rate of the three at $6.88 per GPU-hour for p5.48xlarge in US East (N. Virginia). Google Cloud listed $11.06 in Iowa and Azure $12.29 in East US. Rates vary by region and change over time.

How much does an Azure H100 cost per hour?

Azure’s ND96isr H100 v5, with eight H100 GPUs, lists at $98.32 per hour on demand in East US, about $12.29 per GPU-hour. Reservations bring that to about $7.87 per GPU-hour for one year or $5.40 for three years. Prices were checked September 28, 2026, and are higher in most regions outside the US East and West.

Are specialist GPU clouds cheaper than AWS, Azure, or Google Cloud?

On published on-demand rates, usually. The specialists in our comparison list $3.49 to $6.16 per GPU-hour for H100. Hyperscaler discounts can come close, but with commitments, fixed reservation windows, or interruptible capacity. Compare total cost, including storage, data transfer, and support.

What is the difference between a commitment and a capacity guarantee?

A spending commitment lowers your rate in exchange for committed spend. It does not always reserve GPUs for you. A capacity reservation, like an AWS Capacity Block, holds specific capacity for a set window. Ask exactly what each offer secures.

Can I keep my main cloud and rent GPUs elsewhere?

Yes. Many teams keep data and applications in their main cloud and run GPU work with a specialist. Account for data transfer charges and the time needed to move data before you split the workload.

What does your advisory service cost?

Our advisory services cost you nothing. We’re compensated by whichever provider you choose through us. You pay the provider for your infrastructure and services, with no obligation to choose a provider.