GPU models: 8Providers: 12US price points: 244Market: United StatesList prices: 20 Nov 2026, 07:43 UTCMarket sample: 19 Nov, 19:21 UTC

How GPU rental works

Renting a GPU means paying a provider by the hour, or by the year, for a machine in their data centre. The price for the same chip varies by 10× depending on who you rent from and where: an H100 costs $1.73 to $17.57 per hour today. This guide explains the choices with today’s prices from 14 providers.

What happens when you rent a GPU

GPU rental, also called GPU as a service, splits the work in two. The provider buys the servers, installs them in a data centre with power, cooling and fast networking, and keeps them running. You pick what you need, connect remotely and pay for the time. Seven decisions take you from a model on your laptop to a running job.

  1. Pick the GPU

    Match memory to your model first, then speed. A 70-billion-parameter model in 16-bit needs about 140 GB for its weights alone, so it does not fit on one 80 GB H100. See below.

  2. Pick the term

    On-demand while you are still measuring; reserve the baseline once you know it; use spot for work that can restart. See below.

  3. Pick the form

    A single GPU in a virtual machine, a whole 8-GPU bare-metal server, or a cluster of servers on InfiniBand for large training runs.

  4. Pick the provider

    Compare price, region (data residency), interconnect, support and how fast they can deliver. Reserved clusters are usually sold by quote, not by checkout. See below.

  5. Sign up and verify

    Self-serve clouds need a card and an account. Large reservations add a company check (KYC) and a contract with a service level.

  6. Test, then run

    Run an acceptance test on day one, then connect over SSH, Jupyter or Kubernetes and start your job.

  7. Pay for what you use

    On-demand bills the hours you ran. Reserved terms bill every month whether you use the machine or not.

What one GPU-hour costs today

On-demand list prices per GPU-hour, collected today from public price lists and APIs. The spread column shows how far apart the cheapest and the most expensive offer are for the same chip.

GPUMarketplace medianCheapest neocloudCheapest hyperscalerHighest list priceSpreadProvidersDetail
H100 SXMNVIDIA Hopper · 80 GB HBM3$1.73Vast.ai, 18 offers$2.50Core42, UAE, from price$6.88AWS, N. Virginia$17.57Microsoft Azure, Dubai10.1×13H100 SXM prices
H100 NVLNVIDIA Hopper · 94 GB HBM3$3.73Vast.ai, 5 offers$3.19Runpod, US$6.98Microsoft Azure, Virginia$9.98Microsoft Azure, Dubai3.1×3H100 NVL prices
H200NVIDIA Hopper · 141 GB HBM3e$4.61Vast.ai, 16 offers$3.59Runpod, US$7.91AWS, N. Virginia$15.16Microsoft Azure, Dubai4.2×11H200 prices
B200NVIDIA Blackwell · 180 GB HBM3e$8.75Vast.ai, 9 offers$5.00Core42, UAE, from price$14.00Oracle Cloud, All regions$14.24AWS, N. Virginia2.8×9B200 prices
B300NVIDIA Blackwell Ultra · 288 GB HBM3e$11.88Vast.ai, 9 offers$8.96Crusoe, US$15.00Oracle Cloud, All regions$17.80AWS, N. Virginia2.0×8B300 prices
GB200NVIDIA Grace Blackwell · 186 GB HBM3e–$10.50CoreWeave, US$16.00Oracle Cloud, All regions$50.29Microsoft Azure, Abu Dhabi4.8×3GB200 prices
MI300XAMD CDNA 3 · 192 GB HBM3–$2.59DigitalOcean, US$6.00Oracle Cloud, All regions$8.58Microsoft Azure, Dubai3.3×6MI300X prices
A100 80GBNVIDIA Ampere · 80 GB HBM2e$1.12Vast.ai, 33 offers$1.09Thunder Compute, US$3.43AWS, N. Virginia$7.62Microsoft Azure, Abu Dhabi7.0×11A100 80GB prices

Sources: Azure Retail Prices API, AWS Price List, Oracle price list, provider pricing pages, Vast.ai offers API. Per-provider rows with source links on each GPU page. Download: US prices (CSV).

Five ways to rent, and what each saves

The same machine has a different hourly price depending on how long you commit. We measure each discount by matching the term price to the on-demand price of the exact same instance in the same region.

On-demand

Base pricethe list price every discount is measured against

Start a machine, pay by the hour, stop when you are done. No contract, no minimum. It is the most expensive way to rent and the easiest.

Fits: experiments, short fine-tunes, testing a new GPU before committing.

Cheapest H100 today: $2.50/GPU/hr Core42, UAE, from price

Spot (interruptible)

80% median discount vs on-demandrange 38% to 82%, 47 matched prices

Spare capacity sold at a discount. The provider can take the machine back at short notice when a full-price customer needs it, so your job must save checkpoints and restart.

Fits: batch inference, hyperparameter sweeps, training that checkpoints often.

Cheapest H100 today: $1.99/GPU/hr Together AI, US

1-year reserved

36% median discount vs on-demandrange 24% to 45%, 58 matched prices

You commit to a machine for twelve months and pay monthly (some providers also take it upfront). The price per hour drops because the provider can plan around you.

Fits: a steady inference service, a team that trains every week.

Cheapest H100 today: $3.26/GPU/hr DigitalOcean, US

3-year reserved

56% median discount vs on-demandrange 45% to 57%, 57 matched prices

The longest public commitment. The deepest list discount, but you carry the risk that a faster GPU arrives while you are still paying for this one.

Fits: production inference at a known volume, capacity you will use around the clock.

Cheapest H100 today: $2.97/GPU/hr AWS, N. Virginia

Marketplace

38% lower for interruptible bids than for on-demand hostsrange 14% to 49%, 7 matched prices

Independent hosts list their own servers and set their own prices. Prices move daily and quality varies by host, so check reliability scores and network speed before you rent.

Fits: price-sensitive work that can tolerate a less uniform fleet.

H100 median today: $1.73/GPU/hr Vast.ai, 18 offers

What it costs per month

Pick a GPU and a term. The calculator uses the cheapest public price we see today for that combination.

Estimated monthly cost–

List prices only, before tax, storage and data transfer. Large reservations are usually negotiated below list.

Who rents out GPUs

Three kinds of seller. Hyperscalers (3 tracked) sell GPUs next to hundreds of other cloud services. Neoclouds (10 tracked) are built for GPUs only and usually cost less. Marketplaces let independent hosts list their own servers. The table lists everyone in our daily price feed.

ProviderGPUs listedTermsRegions we trackH100 fromSource
AWS GulfHyperscalerH100 SXM, H200, B200, B300, A100 80GBon-demand, 1-year, 3-yearN. Virginia, Ohio, Oregon, UAE$6.88/GPU/hrPrice source
Microsoft Azure GulfHyperscalerH100 SXM, H100 NVL, H200, GB200, MI300X, A100 80GBon-demand, spot, 1-year, 3-yearAbu Dhabi, Arizona, Dubai, Illinois, Texas +2$11.06/GPU/hrPrice source
Oracle CloudHyperscalerH100 SXM, H200, B200, B300, GB200, MI300X, A100 80GBon-demandAll regions$10.00/GPU/hrPrice source
Core42 GulfNeocloudH100 SXM, H200, B200, MI300Xon-demandUAE$2.50/GPU/hrPrice source
CoreWeaveNeocloudH100 SXM, H200, B200, B300, GB200, A100 80GBon-demand, spotUS$6.16/GPU/hrPrice source
CrusoeNeocloudH100 SXM, H200, B300, MI300X, A100 80GBon-demandUS$3.90/GPU/hrPrice source
DigitalOceanNeocloudH100 SXM, H200, B300, MI300Xon-demand, spot, 1-yearUS$4.41/GPU/hrPrice source
LambdaNeocloudH100 SXM, B200, A100 80GBon-demandUS$4.59/GPU/hrPrice source
OblivusNeocloudH200, A100 80GBon-demandUS–Price source
RunpodNeocloudH100 SXM, H100 NVL, H200, B200, B300, A100 80GBon-demandUS$2.69/GPU/hrPrice source
Thunder ComputeNeocloudH100 SXM, A100 80GBon-demandUS$3.20/GPU/hrPrice source
Together AINeocloudH100 SXM, H200, B200, B300on-demand, spotUS$3.99/GPU/hrPrice source
VultrNeocloudH100 SXM, B200, MI300X, A100 80GBon-demandUS$2.99/GPU/hrPrice source
Vast.aiMarketplaceH100 SXM, H100 NVL, H200, B200, B300, A100 80GBon-demand, interruptibleWorldwide hosts$1.72/GPU/hrPrice source

“Gulf” marks providers with capacity in the UAE, Qatar, Bahrain or Saudi Arabia.

What sets the price of a GPU-hour

Who you rent from

+135%

In the US, AWS and Azure charge that much more than the cheapest neocloud for the same GPU (median across GPU models, today).

Where the machine stands

+47%

The same provider and instance in a Gulf region versus the US. Local capacity is scarcer, and some buyers need their data to stay in the country. Gulf prices are tracked on our sister site, the Gulf GPU Index.

How long you commit

−56%

Median 3-year discount against on-demand for the same machine. The longer the term, the more of the hardware risk you take over from the provider.

Which generation

Newer chips cost more per hour but can finish the job in fewer hours. Compare cost per finished training run or per million tokens, not just per hour. Older GPUs such as the A100 stay cheap and useful for smaller models.

Interconnect

A single GPU in a VM is the cheapest form. A full server with NVLink, or a cluster on InfiniBand, costs more and is what large training needs.

Volume

Hundreds of GPUs on a long term are priced by quote. Public list prices are the ceiling for that kind of deal, not the floor.

Need B300 capacity in the UAE? Oasis Compute, the publisher of GPU Ratings, reserves B300 clusters for teams in the UAE and GCC.

Reserve B300 capacity

GPU rental glossary

GPU-hour
One GPU for one hour. The unit every price on this page uses.
On-demand
Pay per hour, no commitment, stop any time.
Spot / interruptible
Discounted spare capacity that the provider can reclaim at short notice.
Reserved instance
A 1- or 3-year commitment in exchange for a lower hourly rate.
Bare metal
A whole physical server for one customer, no virtualisation layer.
Neocloud
A cloud that specialises in GPUs, such as CoreWeave, Lambda or Crusoe.
Hyperscaler
The largest general clouds: AWS, Microsoft Azure, Google Cloud, Oracle Cloud.
Marketplace
A platform where independent hosts list their own GPU servers and prices.
SXM vs PCIe / NVL
SXM GPUs sit on a board with fast NVLink between them; PCIe and NVL cards plug into standard slots. SXM is the training default.
NVIDIA's direct GPU-to-GPU link inside one server.
InfiniBand
The high-speed network that connects servers in a training cluster so many GPUs can work as one.
HBM
High-bandwidth memory on the GPU. More HBM means bigger models fit on one card.
Burn-in
Running the GPUs at full load for hours before go-live to catch faulty hardware.
Take rate
The share of each rental a marketplace keeps as its fee.

Frequently asked questions

How does GPU rental work?

You pick a GPU model, a number of GPUs and a rental term, and the provider gives you a virtual machine or a whole bare-metal server with those GPUs attached. You pay per GPU-hour on demand, or a lower hourly rate if you commit for one or three years. You connect over SSH or a notebook, run your job, and stop the machine when you are done.

How much does it cost to rent an H100?

Today the lowest cloud list price we track for an H100 SXM is $2.50 per GPU-hour (Core42), and the highest is $17.57 (Microsoft Azure, Dubai). The Vast.ai marketplace median is $1.73. Around the clock for a month (730 hours) that is $1,825.00 to $12,829.75 per GPU.

Is it cheaper to rent or buy a GPU?

Renting is cheaper when you use the GPU part of the time, or when you need it for less than about the useful life of the card. Buying only pays off at high, steady use, and then you also pay for power, cooling, networking, a data-centre rack and staff. Many teams rent on demand first, measure how many GPU-hours they really use, then reserve that baseline.

What is the difference between on-demand, spot and reserved GPUs?

On-demand is pay-as-you-go with no commitment. Spot is spare capacity at a discount that can be taken back at short notice. Reserved means you commit for one or three years for a lower hourly rate; across the machines we track, a 3-year term is a median 56% cheaper than on-demand.

What is GPU as a service (GPUaaS)?

GPU as a service means renting GPU computing power from a provider over the internet instead of owning the hardware. The provider buys, hosts and maintains the servers; you pay per hour or per term and get remote access.

What is a bare-metal GPU server?

A bare-metal server is a whole physical machine that only you use, with no virtualisation layer between your software and the GPUs. You get all of the server's GPUs (usually eight), its CPUs, memory and network cards. It suits large training jobs that need full speed and direct access to the interconnect.

Why do GPU prices differ so much between providers?

Hyperscalers bundle enterprise support, compliance and many other services, and price their GPUs well above specialised GPU clouds (neoclouds) and marketplaces. In the US, AWS and Azure list the same GPU at a median +135% above the cheapest neocloud. Region matters too: the same machine costs on average +47% more in Gulf regions than in the US. Interconnect, contract length and how many GPUs you take also move the price.

What is a GPU-hour?

One GPU used for one hour. Providers that sell 8-GPU servers often quote the price for the whole machine; divide by eight to compare it with a per-GPU price. All prices on GPU Ratings are per GPU-hour.

How do I check that a rented GPU performs as promised?

Run an acceptance test before your real work: check the GPU model and memory with nvidia-smi, run a burn-in or diagnostic (for example NVIDIA DCGM) for an hour or more, and for multi-GPU or multi-node jobs test the interconnect bandwidth with NCCL tests. Ask the provider what happens if a GPU fails during your term.