One GPU running for one hour. Data-centre GPUs usually come eight to a server, so an hour on a full server is eight GPU-hours. A month of 730 hours on one 8-GPU server is 5,840 GPU-hours, and a year of 8,760 hours is 70,080.

Is a GPU price quoted per GPU or per server?

Both are common, so always check. A server price is usually eight times the per-GPU price, because most data-centre GPU servers have eight GPUs. A server at $36 an hour is $4.50 per GPU-hour. Quoting per GPU-hour makes servers of different sizes easy to compare.

Why is a three-year GPU contract cheaper per hour than one year?

Because the provider knows it will earn back the cost of the server, which Reuters put at about $550,000 in the US for an 8-GPU B300 in April 2026. A longer contract lowers its risk and its cost of borrowing. On AWS, an 8-GPU H100 server costs $2.97 per GPU-hour on a three-year plan against $4.33 on a one-year plan, both paid monthly: about 31% less.

Why do the big clouds charge more for the same GPU?

They bundle the GPU with the rest of their cloud, global regions and enterprise contracts. Silicon Data found that in the third quarter of 2026 the big clouds charged about two to two and a half times what other providers charged for an H100 in the same region. For a team that only needs GPU servers, specialist providers usually cost less.

Are GPU rental prices going up or down?

Up, for now. SemiAnalysis’s index of one-year H100 contracts rose almost 40% between October 2025 and March 2026, and Silicon Data’s H100 index went from $2.00 in December 2025 to $3.10 in November 2026. NVIDIA expects supply to stay tight into 2028, so check how long any quote is valid.

How many hours are in a month for GPU billing?

Price calculators, including AWS’s, use 730: 8,760 hours a year divided by twelve. Use that to compare quotes, then check what each provider bills: some charge for the actual hours in each month, up to 744 in a 31-day month, and some bill a flat 720 hours (30 days). At $36 an hour, the gap between 720 and 744 hours is $864 a month.

Does FlexiCloud rent GPUs by the hour?

No. We quote per GPU-hour so you can compare prices, but we rent whole 8-GPU servers on a term of at least one year, billed upfront for the term. Our floor rates on our longest term start at $2.95 per GPU-hour for the H100 and $4.50 for the B300, before tax; a one-year term costs more.

GPU cost per hour: why quotes for the same GPU differ, and how to compare them

Q: How much does a GPU cost per hour in 2026?

As of 14 November 2026, an NVIDIA H100 costs about $3.10 per GPU-hour at specialist GPU providers, by Silicon Data’s index, and $6.88 to $12.29 on demand at the big clouds. A B300 costs about $7.67 at specialists and $17.80 on demand at AWS. Long contracts cost much less: AWS’s H100 falls to $2.59 if you pay three years upfront, before storage, data transfer and support.

Q: Do GPU rental prices include tax?

Often not, so compare quotes before tax. When the provider is in India, 18% GST is added to the invoice. In countries with VAT or GST, a business buying from a provider abroad usually accounts for the tax itself under the reverse charge and can often reclaim it. FlexiCloud’s prices are in US dollars before tax.

14 November 2026Vinod Chacko

GPU cost per hour: why quotes for the same GPU differ, and how to compare them

Short answer: as of 14 November 2026, an NVIDIA H100 costs about $3.10 per GPU-hour at specialist GPU providers, by Silicon Data's index, and $6.88 to $12.29 on demand at the big clouds. A B300 costs about $7.67 at specialists and $17.80 on demand at AWS. Long contracts cost much less than on demand: paying for three years upfront brings AWS's H100 down to $2.59, before storage, data transfer and support, which AWS charges separately.

A GPU-hour is one GPU running for one hour. The GPU sets the starting point; most of the gap comes from how long you commit, how you pay, who you rent from, where the server sits and what the price includes.

This guide explains each of those, then shows how to turn two very different quotes into numbers you can compare. If you already know what you need, our GPU server prices per hour and per month are on one page.

What is a GPU-hour?

One GPU, running for one hour. Data-centre GPUs usually come eight to a server, so one hour on a full server is eight GPU-hours. Over a month of 730 hours (the figure the big clouds use in their price calculators), one 8-GPU server is 5,840 GPU-hours. Over a year of 8,760 hours, it is 70,080.

Quoting per GPU-hour makes servers of different sizes easy to compare. It does not always mean you can rent for one hour. Many providers only rent whole servers, for months or years at a time, and quote per GPU-hour so buyers can compare like with like.

So the first question to ask about any quote: is this price per GPU, or per server? A server price of $36 an hour is $4.50 per GPU-hour, not $36.

What a GPU-hour costs in November 2026

Two public reference points, checked on 14 November 2026. Big clouds publish their list prices. Analysts track what specialist GPU providers charge.

Price per GPU-hour, USD	H100	H200	B200	B300
Big cloud, on demand (AWS list price, US East, 8-GPU server)	$6.88	$7.91	$14.24	$17.80
Specialist GPU providers (Silicon Data index, 13 November 2026)	$3.10	$3.60	$6.87	$7.67

Neither row is what you would pay on a long contract. On a three-year term paid upfront, the same AWS servers drop to $2.59 per GPU-hour for the H100, as the next section shows, and $6.69 for the B300, before AWS's separate charges for storage, data transfer and support. At the other end, Microsoft Azure lists its 8-GPU H100 server at $12.29 per GPU-hour on demand in East US.

The Blackwell GPUs (B200 and B300) cost roughly twice as much per hour as the Hopper ones (H100 and H200), because they are newer, faster, carry more memory and are in shorter supply.

Why quotes for the same GPU differ

Eight things move the price. Two usually move it most: how long you commit and who you rent from.

1. How long you commit

The longer the term, the lower the hourly rate. The big clouds show this most clearly, because they publish every option for the same server. Here is one 8-GPU H100 server on AWS in US East:

How you buy the same 8x H100 server	Per GPU-hour	Saving against on demand
On demand, stop any time	$6.88	None
One-year plan, paid monthly	$4.33	37%
Three-year plan, paid monthly	$2.97	57%
Three-year plan, all paid upfront	$2.59	62%

Same server, same region, and the dearest option costs 2.7 times the cheapest.

The lowest rate comes with two catches. First, you pay for all three years on day one: about $544,000 for one 8-GPU server. Second, it covers the server only. AWS charges separately for storage volumes (EBS), data sent out to the internet, public IP addresses and support. Say you keep 20 TB on its standard SSD storage, about $1,600 a month, and send 10 TB a month out to the internet, about $900. That adds about 43 cents per GPU-hour, taking the $2.59 to about $3.02 before support. Point 7 below covers these extras.

Microsoft Azure and Google Cloud follow the same pattern: roughly a third off for one year and more than half off for three.

The reason is the cost of the machine. Reuters reported in April 2026 that an 8-GPU B300 server costs about $550,000 in the US. Spread over three years, that is about $2.62 per GPU-hour for the hardware alone. Over five years it is about $1.57. That is before power, space, networking, staff and the cost of borrowing. A provider with a signed three-year contract knows it will earn that money back. A provider renting by the hour has to price in the weeks the server sits idle.

2. How and when you pay

Paying upfront can lower the rate further. On AWS, paying the whole term in advance cuts another 7% on a one-year plan and 13% on a three-year plan. Not everyone works this way: Azure charges the same total whether you pay upfront or monthly.

For dedicated GPU servers, paying in advance is normal. Some providers ask for part of the contract upfront; others bill the whole term upfront, as AWS's all-upfront plans do. An index run by the analyst firm SemiAnalysis assumes a 25% prepayment on contracts of three months or more, and it has reported smaller AI labs prepaying more than 20% on contracts of more than four years. A lower rate can reflect a larger payment upfront, so compare quotes on the same payment schedule. Before you sign, ask who owns and runs the hardware, and get the answer in writing. Our questions to ask a GPU provider before you sign cover that.

3. Who you rent from

The big clouds usually charge the most for the same GPU, and the gap is widest on demand. Silicon Data found that in the third quarter of 2026 they charged about two to two and a half times what other providers charged for an H100 in the same region. The gap ran from about 1.6 times in the Nordics to about 3 times in the western US and the Middle East.

That premium buys things some teams need: the rest of the cloud next door, global regions and existing enterprise contracts. Specialist GPU providers focus on GPU servers and price them lower. At the bottom of the market, online marketplaces let many owners list spare GPUs, often one at a time, with less certainty about the hardware and how long it stays available.

4. Where the server sits

Power and space cost very different amounts around the world. An NVIDIA DGX B300 server draws about 14.5 kW. Add the cooling and power losses of an efficient modern data centre, and each B300 GPU uses about 2.5 kWh every hour. Here is what that electricity costs at each region's industrial or large-business prices:

Where	Electricity price	Power cost per B300 GPU-hour
Saudi Arabia, industrial	SAR 0.20 per kWh	about $0.14
Texas, industrial	6.8 US cents per kWh	about $0.17
Virginia, industrial	10.1 US cents per kWh	about $0.26
Maharashtra, India, high-voltage industrial	about Rs 10.7 per kWh	about $0.29
Germany, very large users	EUR 0.138 per kWh	about $0.39
Singapore, regulated tariff (large users usually pay less)	SGD 0.286 per kWh	about $0.56
United Kingdom, very large users	22.7 pence per kWh	about $0.76

Our arithmetic: 14.5 kW, plus 40% for cooling and losses, divided by eight GPUs, times each region's 2026 price. Power is a small part of the hourly rate, under $1 per GPU-hour even in the UK, but it adds up: run one B300 server flat out for a year and the electricity bill ranges from about $9,500 in Saudi Arabia to about $53,000 in the UK. Rent for data-centre space varies too, though less: CBRE's figures for early 2026 put it at about $190 to $235 per kW a month in Northern Virginia, about $188 in Sydney and about $403 in Singapore.

Supply and demand matter too. Silicon Data found H100 rental prices differed by more than six times between regions in 2025. And the law can decide the region before price does: India requires payment data to stay in India, and the UAE keeps health data in the country. If your data is regulated, the cheapest region may not be an option.

5. What you actually get

Two quotes for "an H100" can describe very different machines:

Bare metal, a virtual machine, a container or a slice. Bare metal is the whole physical server, with nothing between your software and the hardware. A virtual machine runs on top of the provider's software layer. A container or Kubernetes pod shares the server's operating system. A slice is part of one GPU: NVIDIA's Multi-Instance GPU feature splits an H100, H200, B200 or B300 into up to seven pieces, and the smallest gets one-seventh of the compute and one-eighth of the memory. Seven slices do not add up to one whole GPU, so a slice price times seven is not a whole-GPU price.
SXM or PCIe. The same GPU name comes in two physical forms. SXM GPUs sit on a shared board, usually eight to a server, joined by NVLink at 900 GB/s per GPU. PCIe versions are plug-in cards: the H100 NVL uses about half the power of the H100 SXM, links to other cards through a slower NVLink bridge (600 GB/s) and has about 16% less peak compute, though a little more memory (94 GB against 80 GB). Our guide to PCIe, SXM, HGX and DGX explains the difference.
The rest of the server. Processors, system memory and local disks vary. NVIDIA's own 8-GPU systems come with eight 3.84 TB NVMe drives, about 30 TB in all. Check what yours has.
Who else is on it. A dedicated server is yours alone. A shared one is cheaper and splits the machine with other customers. See bare metal vs cloud GPUs for what that means for your data.

6. The network between servers

If your job runs on more than one server, the servers need a fast network between them, usually InfiniBand or RoCE (a fast form of Ethernet). It carries 400 gigabits a second per GPU on H100 servers and up to 800 on B300 servers. That network is expensive: in its quarter to April 2026, NVIDIA sold $15 billion of data-centre networking against $60 billion of computing, about $1 of network for every $4 of compute.

So a quote that includes a full cluster network cannot be compared with one that has a basic internet connection, even if the GPUs match. For one server on its own, the GPUs already talk to each other inside the box, and this item does not apply. Our guide to renting a multi-node B300 cluster goes further.

7. What sits outside the hourly rate

The GPU-hour is rarely the whole bill. Ask about each of these in writing:

Shared storage. The server's own disks are usually included, but they are scratch space. Shared or durable storage is often extra. At AWS's list price, 20 TB of fast shared file storage adds about $2,900 a month.
Data transfer. Bringing data in is usually free. Sending it out can be metered: AWS, for example, charges about $0.09 per GB, so moving 10 TB out in a month costs about $900. Some providers include it.
Support. Sometimes included, sometimes a percentage on top. AWS's Business Support+ plan, for example, costs 9% of the first $10,000 of monthly spend and 7% of the next $70,000.
Repairs. GPUs fail more often than people expect. During 54 days of training Llama 3 on up to 16,000 H100s, Meta recorded 419 unexpected interruptions; 148 were faulty GPUs and 72 were GPU memory failures. By our arithmetic, that is roughly one GPU fault every 16 months for each 8-GPU server running flat out, so on a multi-year term expect at least one. Ask how fast a failed GPU is replaced and whether a spare server is available.
Uptime promises. Read what the service level agreement counts as downtime and what you get back. AWS's promise for a single server is 99.5% a month, which allows about three hours and 40 minutes of downtime before any credit, and the first credit is 10% of that server's monthly bill. AWS counts a server as down only when it has no outside connection, so a failed GPU on a server you can still reach may not count.
Set-up fees and software. Some providers charge a one-off set-up fee; some operating systems carry licence fees.

8. When you ask

GPU rental prices are rising, not falling. SemiAnalysis's index of one-year H100 contracts rose almost 40% between October 2025 and March 2026. Silicon Data's H100 index went from $2.00 in December 2025 to $3.10 on 13 November 2026. NVIDIA told investors in August 2026 that supply will stay tight at least to the end of its 2028 financial year, about January 2028. Reuters reported in October that AWS's capacity is largely sold out through 2027.

So a price you saw last year is probably out of date, and every quote should say how long it is valid. For the newest GPUs, the start date matters as much as the rate: a server ready in a week and one ready in three months are not the same offer.

How to compare two quotes

Put every quote on the same basis. Work out everything you will pay over the whole term, then divide by the GPU-hours you get:

Effective cost per GPU-hour = total cost over the term ÷ (number of GPUs × hours in the term)

The total should include the GPU rate, set-up fees, storage, data transfer, support, and any tax you cannot reclaim. Use 730 hours for a month and 8,760 for a year, and check which figure each provider bills on. At $36 an hour, a month of 720 hours costs $25,920, a month of 730 hours $26,280 and a month of 744 hours $26,784. That is $864 a month between the first and the last.

Tax needs care too. Compare prices before tax, then add what you will actually pay. When the provider is in India, 18% GST is added to the invoice, so a quote that already includes GST looks 18% dearer than one that does not. In countries with VAT or GST, a business buying from a provider abroad usually accounts for the tax itself under the reverse charge and can often reclaim it. Your tax adviser can confirm what applies to you.

Then check each quote for the same things:

GPU, form factor (SXM or PCIe) and GPUs per server
Price per GPU-hour, and hours per month billed
Term, and what happens if you need to stop early
Upfront payment, and the payment schedule after it
Region and data centre
Bare metal, virtual machine or container; dedicated or shared
Processors, memory and local disks
Network between servers, if you need more than one
Storage, data transfer and support: included or extra?
Repair time and uptime promise
Start date, and how long the quote is valid
Tax: included or not?

A worked example

Two made-up quotes for one 8-GPU B300 server, to show the method. The numbers are illustrations, not offers.

	Quote A	Quote B
Rate	$6.60 per GPU-hour	$5.90 per GPU-hour
Term	One year	Three years
Payment	Monthly	50% of the server cost upfront, the rest monthly
Extras	Shared storage, $1,500 a month	Data transfer, about $800 a month
Total over the term	$480,528	$1,269,216
Paid on day one	$40,044 (one month)	$620,208
Effective cost per GPU-hour	$6.86	$6.04

Quote B is cheaper per hour, even with its extras. But it commits you to 2.6 times as much money, and $620,208 of it is due on the first day. It is also only cheaper if you use the server for all three years. If the contract has no early exit, stopping after 18 months still costs the full $1,240,416 for the server, which works out to $11.80 for every GPU-hour you used.

Neither quote is wrong. B suits a team with three years of steady training or inference ahead. A suits a team that is still finding out how much compute it needs. The method makes the trade-off visible.

How FlexiCloud helps

We quote per GPU-hour so you can compare like with like, but we rent whole 8-GPU servers on a term of at least one year, billed upfront for the term, with the details agreed case by case. Our floor rates on our longest term are $2.95 per GPU-hour for the H100, $3.35 for the H200, $3.75 for the B200 and $4.50 for the B300, in US dollars before tax. Each is for a dedicated, single-tenant 8-GPU server, with monitoring, failure handling and support included. A one-year term costs more. See what our "from" price means.

We work with more than ten partner providers across six regions: North America, the UK, Europe, India, Asia Pacific and the Gulf. We check every operator before we quote them. Tell us the GPU, the region and the term, and we will come back with quotes on the same basis, with each one saying in writing whether the server is bare metal, a virtual machine or a container, what hardware it has, and what is included.

Get a quote, or see our GPU server prices first.

Where these figures come from

All prices were checked on 14 November 2026 unless a date is given. Prices change often; check the sources for current figures.

AWS on-demand and Savings Plans prices for p5.48xlarge (H100), p5en.48xlarge (H200), p6-b200.48xlarge (B200) and p6-b300.48xlarge (B300), US East (N. Virginia): EC2 on-demand pricing and Savings Plans pricing.
Azure on-demand and reservation prices for the ND96isr H100 v5 (8x H100), East US: Azure Linux virtual machine pricing. Upfront and monthly payment: Microsoft Learn, Buy an Azure reservation. Google Cloud commitments: accelerator-optimized VM pricing.
Specialist GPU provider indices, values as of 13 November 2026: Silicon Data daily index for H100, H200, B200 and B300 (see the dated table on each page). H100 index in December 2025: Silicon Data, H100 price spike (9 January 2026). Big-cloud vs specialist and regional gaps: Silicon Data, H100 rental prices by region, Q3 2026 and The geography of GPU pricing.
H100 one-year contract index and prepayment terms: SemiAnalysis, The great GPU shortage (2 April 2026) and the SemiAnalysis GPU rental index.
B300 server price: Reuters, via Yahoo Finance (30 April 2026). The per-GPU-hour figures are our arithmetic.
AWS capacity: Reuters, via Yahoo Finance (24 October 2026). NVIDIA supply outlook and networking revenue: NVIDIA earnings calls of 20 May and 26 August 2026, published on NVIDIA Investor Relations.
Server power: NVIDIA DGX B300 User Guide. Data-centre overhead: Uptime Institute Global Data Center Survey 2025 (average PUE 1.54; we used 1.4, typical of newer sites).
Electricity prices: US EIA (January to August 2026), Eurostat (first half of 2026), UK DESNZ (second quarter of 2026), Maharashtra Electricity Regulatory Commission (FY 2026-27), SP Group Singapore (fourth quarter of 2026) and Saudi industrial tariffs as reported by MEED (26 October 2026: SAR 0.20 per kWh now, rising to SAR 0.23 from 1 January 2027). Exchange rates: European Central Bank, 13 November 2026.
Data-centre rents: CBRE Global Data Center Trends 2026.
Multi-Instance GPU: NVIDIA MIG User Guide. H100 SXM and NVL: NVIDIA H100.
Storage, data transfer, support and uptime: AWS EBS pricing (gp3 SSD, $0.08 per GB a month), AWS FSx for Lustre pricing, AWS data transfer pricing, AWS Support pricing and the Amazon Compute SLA.
GPU failures: Meta, The Llama 3 Herd of Models, section 3.3.4.
Hours per month in price calculators: AWS Pricing Calculator assumptions.
Data rules: Reserve Bank of India, storage of payment system data; UAE Federal Law No. 2 of 2019 on health data.