Cloud GPU pricing has the exact emotional texture of booking a flight at 1:17am.
You open one provider. Then another. Then a marketplace. Then a managed cloud. Then a pricing page with sixteen footnotes and a region selector that has personally decided to become your enemy.
At some point you forget what you were trying to run.
So I built gpu.fund.
It is a simple comparison site for cloud GPU rentals. No login. No procurement ceremony. Just prices, filters, provider links, and a report page for the current GPU rental market.
What it tracks right now
The current launch snapshot has 146 cloud GPU prices across 12 providers.
A few examples from the latest crawl:
- RTX 3090 from $0.11/hr on Vast.ai
- RTX 4090 from $0.34/hr on Vast.ai
- RTX 5090 from $0.74/hr on Vast.ai
- A100 from $0.70/hr on Vast.ai
- H100 from $2.00/hr on Together.ai
- MI300X from $1.99/hr on RunPod
That spread is the whole reason this exists. The same broad class of workload can be either pretty reasonable or seed-round barbecue depending on where you look first.
The boring useful takeaway
For a lot of inference and experiment work, the cheapest answer is not “rent the fanciest GPU”.
It is usually more like:
- RTX 3090 is still weirdly hard to kill for cheap experiments
- RTX 4090 and RTX 5090 rentals are often the practical lane if VRAM fits
- A100 makes sense when 80GB VRAM matters more than raw flexing
- H100 is great, but many people are buying a race car to drive to the grocery store
- MI300X is interesting when the software stack cooperates
The real enemy is not just price. It is comparison friction.
Hourly price is only the first layer. Storage, bandwidth, interruptible terms, regions, availability, multi-GPU bundles, and weird minimums can all change the real bill. The goal is not to pretend GPU pricing is simple. The goal is to make the first pass less miserable.
What I want feedback on
I am looking for useful complaints from people who actually rent GPUs for inference, fine-tuning, training runs, scraping jobs, and weekend experiments that become Monday problems.
Most useful feedback:
- missing providers
- bad GPU normalization
- fields that should exist in filters
- storage and bandwidth costs worth surfacing
- interruptible pricing quirks
- multi-GPU bundle pricing
- region or availability signals that matter
- whether the report page answers the question you actually have
The site is here:
And the live market report is here:
If this saves one person from opening twelve pricing tabs and whispering “what is a compute unit” into the void, the cat has done its job.