Cheapest GPU Cloud Providers Compared
June 24, 2026 (updated) by @anthonynsimon
We currently track 4630 GPU offerings across 66 providers. All normalized to a USD / GPU / hour rate for easier comparison.
GPU cloud pricing summary
The tables below update automatically as pricing changes. Each segment shows the cheapest on-demand price for each GPU model, along with the provider offering that price and the total number of providers offering that GPU.
High Performance
Large-scale model training, multi-node distributed workloads, and high-throughput inference.
| GPU Model | VRAM | Cheapest / GPU / hr | Cheapest Provider | Total Providers |
|---|---|---|---|---|
H100
|
80 GB | $1.38 |
Thunder Compute
|
47 |
|
A100
|
80 GB | $0.52 |
Vast.ai
|
40 |
|
H200
|
141 GB | $1.98 |
Vast.ai
|
32 |
|
L40S
|
48 GB | $0.55 |
Novita
|
29 |
|
B200
|
192 GB | $3.75 |
Packet·ai
|
26 |
|
B300
|
288 GB | $6.06 |
Vast.ai
|
13 |
|
L40
|
48 GB | $0.47 |
Vast.ai
|
11 |
MI300X
|
192 GB | $1.99 |
Hot Aisle
|
10 |
|
GB200
|
192 GB | $10.50 |
CoreWeave
|
8 |
|
MI355X
|
2300 GB | $8.60 |
Oracle Cloud
|
5 |
|
MI325X
|
256 GB | $3.21 |
Cyfuture AI
|
4 |
|
GH200
|
96 GB | $1.99 |
Vultr
|
3 |
|
A800
|
40 GB | $0.93 |
GPUhub
|
1 |
Gaudi2
|
96 GB | $1.09 |
Cyfuture AI
|
1 |
Mid-Range
Fine-tuning, single-node training, development, and production inference.
| GPU Model | VRAM | Cheapest / GPU / hr | Cheapest Provider | Total Providers |
|---|---|---|---|---|
|
RTX Pro 6000
|
96 GB | $0.80 |
Vast.ai
|
24 |
|
V100
|
32 GB | $0.10 |
Vast.ai
|
18 |
|
RTX A6000
|
48 GB | $0.33 |
Runpod
|
16 |
|
L4
|
24 GB | $0.31 |
Vast.ai
|
15 |
|
RTX 4090
|
24 GB | $0.18 |
Salad
|
14 |
|
RTX 6000 Ada
|
48 GB | $0.54 |
Vast.ai
|
12 |
|
RTX 5090
|
32 GB | $0.27 |
Salad
|
11 |
|
A10
|
24 GB | $0.20 |
Vast.ai
|
10 |
|
A4000
|
16 GB | $0.09 |
Vast.ai
|
9 |
|
A5000
|
24 GB | $0.11 |
Salad
|
9 |
|
T4
|
16 GB | $0.13 |
Vast.ai
|
8 |
|
RTX 3090
|
24 GB | $0.11 |
Salad
|
7 |
|
P100
|
16 GB | $0.08 |
Vast.ai
|
6 |
|
A40
|
48 GB | $0.29 |
Vast.ai
|
6 |
|
RTX 4080
|
16 GB | $0.13 |
Salad
|
5 |
|
RTX 5080
|
16 GB | $0.20 |
Vast.ai
|
4 |
|
A30
|
24 GB | $0.35 |
Massed Compute
|
4 |
|
RTX 5060 Ti
|
16 GB | $0.09 |
Vast.ai
|
3 |
|
RTX 3090 Ti
|
24 GB | $0.12 |
Salad
|
3 |
|
Quadro RTX 6000
|
24 GB | $0.13 |
Vast.ai
|
3 |
|
RTX 5070 Ti
|
16 GB | $0.12 |
Salad
|
2 |
|
RTX PRO 4000
|
24 GB | $0.27 |
Vast.ai
|
2 |
|
RTX Pro 4500
|
32 GB | $0.28 |
Vast.ai
|
2 |
|
RTX 5000
|
16 GB | $0.60 |
OVH
|
2 |
|
RTX Pro 5000
|
48 GB | $0.60 |
Vast.ai
|
2 |
|
K80
|
24 GB | $0.90 |
AWS
|
2 |
|
A10G
|
24 GB | $1.01 |
AWS
|
2 |
|
RTX 4070 Ti Super
|
16 GB | $0.11 |
Salad
|
1 |
|
A4500
|
20 GB | $0.19 |
Runpod
|
1 |
|
T4G
|
16 GB | $0.42 |
AWS
|
1 |
|
RTX 5000 Ada
|
32 GB | $0.43 |
Vast.ai
|
1 |
|
RTX 5880 Ada
|
48 GB | $0.48 |
Vast.ai
|
1 |
|
RTX Pro 6000D
|
84 GB | $0.82 |
GPUhub
|
1 |
|
V100S
|
32 GB | $0.88 |
OVH
|
1 |
Budget
Learning, experimentation, lightweight inference, and small-scale projects.
| GPU Model | VRAM | Cheapest / GPU / hr | Cheapest Provider | Total Providers |
|---|---|---|---|---|
|
RTX 4000 Ada
|
20 GB | $0.20 |
Runpod
|
7 |
|
RTX 3070
|
8 GB | $0.06 |
Salad
|
5 |
|
RTX 3080 Ti
|
12 GB | $0.10 |
Vast.ai
|
5 |
|
RTX 3060 Ti
|
8 GB | $0.05 |
Salad
|
4 |
|
RTX 2060
|
6 GB | $0.04 |
Salad
|
3 |
|
GTX 1070
|
8 GB | $0.04 |
Salad
|
3 |
|
RTX 3060
|
12 GB | $0.05 |
Vast.ai
|
3 |
|
RTX A2000
|
12 GB | $0.05 |
Vast.ai
|
3 |
|
RTX 2080 Ti
|
11 GB | $0.05 |
Vast.ai
|
3 |
|
RTX 4060
|
8 GB | $0.06 |
Vast.ai
|
3 |
|
RTX 5060
|
8 GB | $0.08 |
Vast.ai
|
3 |
|
RTX 4060 Ti
|
16 GB | $0.08 |
Vast.ai
|
3 |
|
RTX 3080
|
10 GB | $0.08 |
Salad
|
3 |
|
RTX 4070
|
12 GB | $0.09 |
Salad
|
3 |
|
RTX 4070 Ti
|
12 GB | $0.10 |
Salad
|
3 |
|
A16
|
16 GB | $0.47 |
Vultr
|
3 |
|
GTX 1080
|
8 GB | $0.04 |
Vast.ai
|
2 |
|
RTX 2070
|
8 GB | $0.04 |
Salad
|
2 |
|
GTX 1660 Super
|
6 GB | $0.04 |
Salad
|
2 |
|
GTX 1660
|
6 GB | $0.04 |
Salad
|
2 |
|
GTX 1650
|
4 GB | $0.04 |
Salad
|
2 |
|
GTX 1060
|
6 GB | $0.04 |
Salad
|
2 |
|
GTX 1080 Ti
|
11 GB | $0.04 |
Salad
|
2 |
|
P4000
|
8 GB | $0.06 |
Vast.ai
|
2 |
|
RTX 2080
|
8 GB | $0.07 |
Salad
|
2 |
|
RTX 3070 Ti
|
8 GB | $0.08 |
Salad
|
2 |
|
RTX 5070
|
12 GB | $0.10 |
Salad
|
2 |
|
M60
|
16 GB | $0.75 |
AWS
|
2 |
|
GTX 1050 Ti
|
4 GB | $0.03 |
Salad
|
1 |
|
GTX 1070 Ti
|
8 GB | $0.05 |
Vast.ai
|
1 |
|
RTX 3050
|
8 GB | $0.05 |
Salad
|
1 |
|
P40
|
24 GB | $0.19 |
Vast.ai
|
1 |
|
Radeon Pro V520
|
8 GB | $0.38 |
AWS
|
1 |
|
M4000
|
8 GB | $0.45 |
Paperspace
|
1 |
|
P5000
|
16 GB | $0.78 |
Paperspace
|
1 |
|
P4
|
8 GB | $0.99 |
Google Cloud
|
1 |
|
P6000
|
24 GB | $1.10 |
Paperspace
|
1 |
Prices are for on-demand billing only. Spot and reservation pricing may be lower. Prices are normalized to per-GPU-per-hour in USD. For full methodology, see our cost estimate assumptions.
How to read this table
Price / GPU / hr is the total instance price divided by the number of GPUs in that instance. For example, an 8x H100 instance at $24.00/hr works out to $3.00 per GPU per hour. This makes it possible to compare single-GPU and multi-GPU offerings on equal footing.
Cheapest Provider shows which provider currently offers the lowest on-demand rate for that GPU model. This can change as providers update pricing.
Providers is the total number of cloud providers offering that GPU model with on-demand billing.
Billing types explained
Cloud GPU providers typically offer several billing models:
- On-demand: Pay by the hour with no commitment. Start and stop at any time. This is what the table above shows.
- Spot / preemptible: Discounted rates (often 50-80% cheaper) on spare capacity, but your instance can be interrupted with short notice. Good for fault-tolerant batch workloads.
- Reservation: Commit to 1-12+ months for a lower hourly rate. Best for steady-state production workloads where you know your capacity needs.
- Custom / contract: Negotiated pricing for large deployments. Typically requires contacting the provider directly.
The table above focuses on on-demand pricing because it is the most universally comparable across providers. For spot and reservation pricing on a specific GPU, visit the individual GPU model pages.
Tips for getting the best GPU cloud price
- Compare per-GPU-hour, not per-instance-hour. A cheaper 8-GPU instance might actually cost more per GPU than a single-GPU offering from another provider.
- Check availability. The cheapest listed price does not help if the GPU is out of stock. Some providers show availability status on their pricing pages.
- Consider total cost. Egress fees, storage, and networking costs add up. A slightly higher GPU rate with free egress may be cheaper overall.
- Match the GPU to the workload. An H100 is not always necessary. For inference workloads, a mid-range GPU like the L40S or A6000 may deliver sufficient performance at a fraction of the cost.
- Look at spot pricing for batch jobs. If your workload can tolerate interruptions, spot instances can reduce costs by 50% or more.
Explore all GPU cloud pricing data
This page shows the single cheapest provider per GPU model. For the full picture, visit our free cloud GPU index. Each GPU model page there has a complete offerings table, spec breakdown, and historical price chart.