GPU Cloud Price Comparison Guide

A practical checklist for estimating cloud GPU rental cost before you open twelve pricing tabs.

Live price board: https://gpu.fund

Quick snapshot

Latest local gpu.fund crawl: 2026-05-27 01:00 UTC, 140 prices across 12 providers.

Cheapest observed single-GPU hourly prices in this snapshot:

GPU	Lowest observed price	Provider	Notes
RTX 3090	$0.1489/hr	Vast.ai	Cheap experiments when 24GB VRAM is enough
RTX 4090	$0.6900/hr	RunPod	Good practical lane for inference and small fine-tunes
RTX 5090	$0.9081/hr	Vast.ai	Newer consumer GPU, availability can swing
A100 80GB	$0.7348/hr	Vast.ai	Useful when VRAM matters more than raw H100 speed
H100	$2.0000/hr	Together.ai	Strong training/inference option, but not always needed
H100 PCIe/NVL class	$2.3881/hr	Vast.ai	Check exact SKU and interconnect before comparing
MI300X	$1.9900/hr	RunPod	Interesting for AMD-friendly stacks
B200	$4.2329/hr	Vast.ai	Blackwell supply is early and noisy

Prices move. Treat this as a format example, not a quote. Use gpu.fund for the current crawl.

How to estimate the real cost

Hourly GPU price is only the first line item. The useful estimate is:

estimated run cost = GPU hourly rate * runtime hours * GPU count
                   + persistent storage
                   + bandwidth or egress
                   + idle setup/debug time
                   + retries, preemptions, or failed runs

For short experiments, idle setup time often matters more than a five-cent hourly difference. For production inference, utilization matters more than the sticker price.

Fields to check before renting

1. GPU model and VRAM

   • Make sure the SKU is actually the model you expect.
   • H100 SXM, H100 PCIe, H100 NVL, and H200 are not interchangeable for serious comparisons.

2. Price per GPU vs price per machine

   • Some providers show a node price with multiple GPUs.
   • Normalize to price per GPU-hour before comparing.

3. Availability and queue time

   • The cheapest H100 is fake if it sits pending forever.
   • Prefer providers that show inventory or launch success clearly.

4. Storage and snapshots

   • Persistent volumes, images, and snapshots can outlive the GPU rental.
   • Delete what you do not need after the run.

5. Bandwidth and egress

   • Moving models, datasets, checkpoints, and outputs can change the bill.
   • This is especially easy to miss when training data sits in another cloud.

6. Interruptible terms

   • Spot and interruptible instances can be cheap.
   • They are only cheap if your job can resume cleanly.

7. Region and compliance needs

   • Region affects latency, availability, and sometimes data handling.
   • Do not ignore it for user-facing inference.

8. Multi-GPU topology

   • For training, interconnect matters.
   • Eight GPUs on a weak topology can be worse than fewer GPUs with the right fabric.

Practical shortcuts

• If the model fits in 24GB VRAM, test on RTX 3090/4090/5090 before paying H100 money.
• If you need 80GB VRAM, compare A100 80GB, H100, H200, and MI300X rather than defaulting to the newest NVIDIA card.
• If you are debugging code, rent the cheapest compatible GPU first, then scale.
• If you are running production inference, measure throughput per dollar, not price per hour.
• If your job can be interrupted, spot-style pricing can be worth it. If not, include failure cost.

Useful links

• Live GPU rental price board: https://gpu.fund
• Current market report: https://gpu.fund/report
• Hidden GPU rental costs checklist: https://gpu.fund/blog/hidden-costs-cloud-gpu-rentals

Disclaimer

This is infrastructure research and cost estimation, not financial advice, investment advice, or a guarantee of provider pricing. Always verify live provider terms before renting.