AIMultipleAIMultiple
No results found.

Cloud GPUs for Deep Learning: Availability& Price / Performance

Cem Dilmegani
Cem Dilmegani
updated on Aug 21, 2025

If you are flexible about the GPU model, identify the most cost-effective cloud GPU based on our benchmark of 10 GPU models in image and text generation & finetuning scenarios.

If you prefer a specific model (e.g. A100), identify the lowest-cost GPU cloud provider offering it.

If undecided between on-prem and the cloud, explore whether to buy or rent GPUs on the cloud.

Or learn our cloud GPU benchmark methodology to identify the most cost-efficient GPU

Cloud GPU price per throughput

Two common pricing models for GPUs are “on-demand” and “spot” instances. See the most cost effective GPU for your workload based on on-demand prices from the top 3 hyperscalers:

Cloud GPU Throughput & Prices

Updated on September 1, 2025

Showing 10 of 199
Show the lowest cost region
Amazon Web Services

Amazon Web Services

Code
inf1.2xlarge
Region
North America
GPU
1 x AWS Inferentia
Images/s
35
Price/h
$ 0.38
331,579Images / $
Lambda

Lambda

Code
gpu_1x_gh200
Region
North America
GPU
1 x NVIDIA H200 96 GB
Images/s
96
Price/h
$ 1.49
231,946Images / $
Amazon Web Services

Amazon Web Services

Code
g5g.xlarge
Region
North America
GPU
1 x NVIDIA A10G 16 GB
Images/s
27
Price/h
$ 0.43
226,047Images / $
Latitude.sh

Latitude.sh

Code
g3.a100.large
Region
North America
GPU
8 x NVIDIA A100 80 GB
Images/s
469
Price/h
$ 8.28
203,913Images / $
Latitude.sh

Latitude.sh

Code
g3.a100.large
Region
West Europe
GPU
8 x NVIDIA A100 80 GB
Images/s
469
Price/h
$ 8.28
203,913Images / $
Runpod

Runpod

Code
NVIDIA L4_1_US-GA-2
Region
North America
GPU
1 x NVIDIA L4 24 GB
Images/s
23
Price/h
$ 0.43
192,558Images / $
Visit Website
Runpod

Runpod

Code
NVIDIA L4_1_EUR-IS-1
Region
North Europe
GPU
1 x NVIDIA L4 24 GB
Images/s
23
Price/h
$ 0.43
192,558Images / $
Visit Website
Runpod

Runpod

Code
NVIDIA L4_2_EUR-IS-1
Region
North Europe
GPU
2 x NVIDIA L4 24 GB
Images/s
46
Price/h
$ 0.86
192,558Images / $
Visit Website
Runpod

Runpod

Code
NVIDIA L4_4_EUR-IS-1
Region
North Europe
GPU
4 x NVIDIA L4 24 GB
Images/s
91
Price/h
$ 1.72
190,465Images / $
Visit Website
Runpod

Runpod

Code
NVIDIA L4_3_EUR-IS-1
Region
North Europe
GPU
3 x NVIDIA L4 24 GB
Images/s
68
Price/h
$ 1.29
189,767Images / $
Visit Website
Filters
GPU Name
AWS Trainium 512 GB
AMD MI300X 192 GB
NVIDIA B200 180 GB
NVIDIA H100 141 GB
NVIDIA H200 141 GB
NVIDIA H200 96 GB
NVIDIA A100 80 GB
NVIDIA H100 80 GB
NVIDIA L4 48 GB
NVIDIA A100 40 GB
AWS Inferentia2 32 GB
AWS Trainium 32 GB
NVIDIA V100 32 GB
NVIDIA A10G 24 GB
NVIDIA L4 24 GB
NVIDIA A10G 16 GB
NVIDIA T4 16 GB
NVIDIA V100 16 GB
NVIDIA M60 8 GB
NVIDIA T4 8 GB
NVIDIA T4 4 GB
NVIDIA T4 2 GB
AWS Inferentia
NVIDIA K80
Cloud
Acecloud
Alibaba Cloud
Amazon Web Services
CoreWeave
Digital Ocean
Lambda
Latitude.sh
Runpod
Region
North America
Latin America
East Europe
North Europe
South Europe
West Europe
East Asia
North Asia
South Asia
South East Asia
West Asia
Australia & New Zealand
Africa
Middle East
Not Specified

See cloud GPU benchmark methodology for details.

On-demand is the most straightforward pricing model where you pay for the compute capacity by the hour or second, depending on what you use with no long-term commitments or upfront payments. These instances are recommended for users who prefer the flexibility of a cloud GPU platform without any up-front payment or long-term commitment. On-demand instances are usually more expensive than spot instances, but they provide guaranteed uninterrupted capacity.

On-demand GPUs from other cloud providers

Cloud
GPU / Memory*
# of GPUs
On-demand $
Throughput**
Throughput/$***
A100 / 80 GB
8
15.12
1,362
90
TensorDock
A100 / 80 GB
4
1.200
821
684
Vast.ai
V100 / 16 GB
8
0.944
289
306
Vast.ai
V100 / 16 GB
2
0.358
77
215
TensorDock
A100 / 80 GB
1
1.200
232
193
TensorDock
V100 / 16 GB
1
0.220
42
191
TensorDock
A100 / 80 GB
1
1.400
232
165
Jarvislabs
A100 / 40 GB
1
1.1
179
163
Lambda
A100 / 40 GB
1
1.1
179
163
Lambda
H100 / 80 GB
1
1.99
322
162

* Memory and GPU model are not the only parameters. CPUs and RAM can also be important, however, they are not the primary criteria that shape cloud GPU performance. Therefore, for simplicity, we have not included number of CPUs or RAM in these tables.

** Training throughput is a good metric to measure relative GPU effectiveness. It measures the number of tokens processed per second by the GPU for a language model (i.e. bert_base_squad).1 Please note that these throughput values should serve as high level guidelines. The same hardware would have a significantly different throughput for your workload since there is significant throughput difference even between LLMs running on the same hardware.2

*** Excludes cost of storage, network performance, ingress/egress etc. This is only the GPU cost.3

Spot GPUs

Cloud
GPU / Memory*
# of GPUs
Spot
Throughput**
Throughput/$***
Azure
A100 / 80 GB
1
0.76
232
303
Azure
A100 / 80 GB
4
3.05
821
269
Azure
A100 / 80 GB
2
1.53
406
266
Jarvislabs
A100 / 40 GB
1
0.79
179
227
GCP
A100 / 40 GB
1
1.62
179
111
AWS
V100 / 16 GB
1
0.92
42
46
AWS
V100 / 16 GB
4
3.67
153
42
Azure
V100 / 16 GB
1
1.04
42
40
AWS
V100 / 16 GB
8
7.34
289
39
Azure
V100 / 16 GB
2
2.08
77
37

In all these throughput per dollar tables:

  • Not all possible configurations are listed, more commonly used, deep learning focused configurations are included.
  • West or Central US regions were used where possible
  • These are the list prices for each category, high volume buyers can possibly get better pricing

Finally, it is good to clarify what “spot” means. Spot resources are:

Interruptible so users need to keep on recording their progress. For example, Amazon EC2 P3, which provides V100 32 GB, is one of the most frequently interrupted Amazon spot services.4

Offered on a dynamic, market-driven basis. The price for these GPU resources can fluctuate based on supply and demand, and users typically bid on the available spot capacity. If a user’s bid is higher than the current spot price, their requested instances will run.

Cloud GPU costs & availability

GPU
Lowest price (USD/hr)
Vendor with the lowest price
Nvidia L4
$0.38
Seeweb
Nvidia RTX4000
$0.38
Hetzner, Paperspace by DigitalOcean
AMD 7900XTX
$0.39
DataCrunch
Nvidia T4G
$0.42
AWS
Nvidia M4000
$0.45
Paperspace by DigitalOcean
Nvidia RTX6000
$0.50
Lambda Labs
Nvidia T4
$0.53
Azure
Nvidia V100
$0.62
DataCrunch
Nvidia H100
$2.49
Lambda Labs
Nvidia A100
$1.29
DataCrunch

Sorting by lowest price. For other low cost options, you can check out cloud GPU marketplaces.

GPU availability

Input the model that you want in the search box to identify all cloud providers that offer it:

Provider
GPU
Multi-GPU
$/hour***
AWS
M60 8 GB
1, 2, 4x
1.14
AWS
T4 16 GB
1, 2, 4, 8x
1.20
AWS
A10G 24 GB
1, 4, 8x
1.62
AWS
V100 16 GB
1, 4, 8x
3.06
AWS
V100 32 GB
8x
3.90****
AWS
A100 40 GB
8x
4.10****
AWS
A100 80 GB
8x
5.12****
Azure
K80 12 GB
1, 2, 4x
0.90
Azure
T4 16 GB
1, 4x
1.20
Azure
P40 24 GB
1, 2, 4x
2.07

*** On-demand price *($) per single GPU. Excludes cost of storage, network performance, ingress/egress etc. This is only the GPU cost.

**** Computed values. This was needed when single GPU instances were not available.5 6

Other cloud GPU considerations

Availability: Not all GPUs listed above may be available due to capacity constraints of the cloud providers and increasing demand for generative AI.

Data security: For example, cloud GPU marketplaces like Vast.ai offer significantly lower prices but depending on the specific resource requested, the data security of the workload could be impacted, givings hosts the capability to access workloads. Since we prioritized enterprise GPU needs, Vast.ai wasn’t included in this benchmark.

Ease of use: Documentation quality is a subjective metric but developers prefer some cloud providers’ documentation over others. In this discussion, GCP’s documentation was mentioned as lower quality than other tech giants’.7

Familiarity: Even though cloud providers put significant effort into making their services easy-to-use, there is a learning curve. That is why major cloud providers have certifications systems in place. Therefore, for small workloads, the cost savings of using a low cost provider may be less than the opportunity cost of the time it takes a developer to learn how to use their cloud GPU offering.

Buy GPUs or rent cloud GPUs

Buying makes sense

– If your company has the know-how and preference to host the servers or manage colocated servers.

– For uninterruptible workloads: For the volume of GPUs for which you can ensure a high utilization (e.g. more than 80%) for a year or more.8

– For interruptible workloads: The high utilization period quoted above needs to be a few times longer since on-demand (uninterruptible computing) prices tends to be a few times more expensive than spot (interruptible computing) prices.

Our recommendation for businesses with heavy GPU workloads is a mix of owned and rented GPUs where guaranteed demand runs on owned GPUs and variable demand runs on the cloud. This is why tech giants like Facebook are building their own GPU clusters including hundreds of GPUs.9

Buyers may be tempted to consider consumer GPUs which offer a better price/performance ratio however, the EULA of their software prohibits their use in data centers.10 Therefore, they are not a good fit for machine learning except for minor testing workloads on data scientists’ machines.

Cloud GPU benchmark methodology

Prices: Cloud GPU prices are crawled monthly.

Configuration:

  • 4 bit FP quantization was used throughout the benchmark.

Frameworks:

  • Finetuning: Unsloth
  • Inference: llama-cpp-python

Performance:

  • Performance of all GPU models except NVIDIA H100 and AMD GPUs were measured on AWS.
  • It is assumed that the same GPU provides the same performance in any cloud.

Performance on:

  • Text finetuning was measured by finetuning Llama 3.2 with the first 5k conversations on FineTome using 1M tokens. Finetuning was carried over 5 epochs. Number of tokens times number of epochs was divided by finetuning time to identify number of tokens finetuned per second.
  • Text inference was measured during inference of 1 million tokens including both input and output tokens. We divided number of tokens by the total duration to calculate the average number of tokens per second during inference.
  • Image operations was measured by finetuning YOLOv9 with 100 images from SkyFusion for 4 epochs and then by inferencing the finetuned model with ~500 640×640 images.

Next steps:

  • Data collection frequency will be increased
  • We will increase GPU coverage, include more metrics and refresh our performance measurement over time.

What are the top cloud GPU hardware?

Almost all cloud GPUs use Nvidia GPU instances. AMD and other providers also offer GPUs however due to various reasons (e.g. limited developer adoption, lower price per performance etc.), their GPUs are not as widely demanded as Nvidia GPUs.

To see cloud GPU providers that offer non-Nvidia GPUs, please check out the comprehensive list of cloud GPU providers.

Read about all AI chips / hardware.

What are cloud GPU marketplaces?

Distributed cloud marketplaces like Salad, Vast.ai, and Clore.ai provide access to decentralized GPU computing power through a marketplace model. Users with idle hardware can list their GPUs for rent, while those needing GPU power can select from available resources at different price points. These platforms facilitate the connection between supply and demand without relying on centralized cloud providers. They offer a cost-effective and flexible solutions for GPU-intensive tasks.

Salad: decentralized network for tasks like AI training or crypto mining, with a focus on user rewards and ease of use.

Vast.ai: Connects GPU providers with users in need of affordable and scalable computational resources. Focus is on AI and machine learning workloads.

Clore.ai: A distributed marketplace for cloud GPUs. Focus is mostly on: AI, and other HPC needs.

Kryptex: A platform that enables users to earn cryptocurrency by renting out their GPUs. Main focus is to perform tasks like crypto mining or processing complex calculations.

Nvidia DGX Cloud Lepton: Connects AI developers with Nvidia’s network of cloud providers, including CoreWeave, Lambda, and Crusoe, to access GPU resources. Focus is on providing a flexible marketplace for training and deploying AI models across diverse cloud platforms.11

What are the top cloud GPU platforms?

Top cloud GPU providers are:

  • AWS
  • Microsoft Azure
  • CoreWeave
  • Google Cloud Platform (GCP)
  • IBM Cloud
  • Jarvis Labs
  • Lambda Labs
  • NVIDIA DGX Cloud
  • Oracle Cloud Infrastructure (OCI)
  • Paperspace CORE by DigitalOcean
  • Runpod.io

For more on these providers, check out cloud gpu providers.

If you are not sure about cloud GPUs, explore other options like serverless GPU.

For tools allow users to collaborate on, check out cloud collaboration tools.

If you are unclear about what cloud GPUs are, here is more context:

What is a cloud GPU?

Unlike a CPU, which may have a relatively small number of cores optimized for sequential serial processing, a GPU can have hundreds or even thousands of smaller cores designed for multi-threading and handling parallel processing workloads.

A cloud GPU refers to a certain way of GPU usage that’s provided as a service through cloud computing platforms. Much like traditional cloud services, a cloud gpu allows you to access high-performance computing resources spot or on-demand, without the need for upfront capital investment in hardware.

What are the functions/application areas of cloud GPUs?

Cloud GPUs are primarily used for processing tasks that require high computational power. Here are some of the primary uses for cloud GPUs:

Machine Learning and AI

GPUs are particularly effective at handling the complex calculations required for machine learning (ML) and artificial intelligence (AI) models. They can process multiple computations in parallel, making them suitable for training large neural networks and algorithms.

Deep learning

Deep learning is a sub-field of machine learning. Deep learning algorithms greatly benefit from the parallel processing capabilities of GPUs, making training and inference faster and more efficient.

Data processing

Data analysis

GPUs are used to accelerate computing and data processing tasks, such as Big Data analysis and real-time analytics. They can handle high-throughput, parallel processing tasks more efficiently than CPUs.

Scientific computing

In scientific research, cloud GPUs can handle computations for simulations, bioinformatics, quantum chemistry, weather modeling, and more.

Simulations

Certain complex simulations can run more efficiently on GPUs.

Gaming & entertainment

Cloud GPUs are used to provide cloud gaming services, such as Google’s Stadia or NVIDIA’s GeForce Now, where the game runs on a server in the cloud, and the rendered frames are streamed to the player’s device. This allows high-quality gaming without the need for a powerful local machine.

Graphics rendering

GPUs were initially designed to handle computer graphics, and they still excel in this area. Cloud GPUs are used for 3D modeling and rendering, 3D visualizations, virtual reality (VR), computer-aided design (CAD), and computer-generated imagery (CGI).

Video processing

They’re used in video encoding and decoding, video editing, color correction, effects rendering, and other video processing tasks.

Cryptocurrency mining

GPUs are also used in tasks like cryptocurrency mining. However application-specific integrated circuits (ASICs) are offering better economics for more commonly mined crypto currencies.

Notes

Cloud providers are constantly updating their offering, this research will be constantly updated.

Principal Analyst
Cem Dilmegani
Cem Dilmegani
Principal Analyst
Cem has been the principal analyst at AIMultiple since 2017. AIMultiple informs hundreds of thousands of businesses (as per similarWeb) including 55% of Fortune 500 every month.

Cem's work has been cited by leading global publications including Business Insider, Forbes, Washington Post, global firms like Deloitte, HPE and NGOs like World Economic Forum and supranational organizations like European Commission. You can see more reputable companies and resources that referenced AIMultiple.

Throughout his career, Cem served as a tech consultant, tech buyer and tech entrepreneur. He advised enterprises on their technology decisions at McKinsey & Company and Altman Solon for more than a decade. He also published a McKinsey report on digitalization.

He led technology strategy and procurement of a telco while reporting to the CEO. He has also led commercial growth of deep tech company Hypatos that reached a 7 digit annual recurring revenue and a 9 digit valuation from 0 within 2 years. Cem's work in Hypatos was covered by leading technology publications like TechCrunch and Business Insider.

Cem regularly speaks at international technology conferences. He graduated from Bogazici University as a computer engineer and holds an MBA from Columbia Business School.
View Full Profile

Comments 2

Share Your Thoughts

Your email address will not be published. All fields are required.

0/450
Ashley Jenkinson
Ashley Jenkinson
Oct 31, 2024 at 08:54

Cem – great article, I’d love to pick your brain on private networking or direct connects to these GPU instances.

Cem Dilmegani
Cem Dilmegani
Nov 10, 2024 at 06:58

Hi Ashley, thank you! Sure, happy to chat.

Harsh Sharma
Harsh Sharma
Oct 06, 2024 at 02:19

Hi there, fantastic article and very well-researched. Would you mind checking out Dataoorts at https://dataoorts.com

Cem Dilmegani
Cem Dilmegani
Oct 22, 2024 at 03:18

Sure, we’ll review to see if we can include Dataoorts in the next edit.