Open any GPU marketplace and you see a wall of asking prices. None of those numbers tell you what a rental is worth. An unsold listing is a wish; a rented listing is a transaction. The price where transactions actually happen is the market clearing price, and it is the only number that should drive your pricing decisions.
Asking price is what sellers hope for
The available listings on Vast.ai include machines priced by hosts who checked the market once in March, machines priced to cover an electricity bill from a country with expensive power, and machines priced high on purpose because the owner only wants them rented at a premium. All of these show up in the list. None of them are evidence of what renters will pay.
If you average the asking prices of available RTX 4090s and price against that average, you are pricing against wishful thinking. The machines that stay in the available list all day are, by definition, the ones priced wrong. Surviving listings are a biased sample: the well-priced ones keep leaving the list because they get rented.
Rented price is what buyers actually pay
The rented side of the market is a record of real decisions. Every rented machine represents a renter who compared offers and committed money at a specific rate. Aggregate those rates and you get the shape of real demand:
- The rented P50 median is the midpoint: half of all active rentals for that GPU type are trading below it, half above. This is the practical market clearing price.
- Prices well below the median mean hosts leaving money on the table, often via blind undercutting scripts.
- Prices well above the median are machines that won a rental on something other than price: a rare configuration, high bandwidth, a renter in a hurry.
The clearing price also moves. Model releases, term-end deadlines for research groups, weekend hobbyist traffic, and new supply from large hosts all shift where rentals clear, week by week and sometimes day by day.
A tale of two data sources
Here is the same market read two different ways:
| Signal | RTX 4090 example | What it tells you |
|---|---|---|
| Average asking price (available) | $0.52/hr | What unsold hosts wish renters would pay |
| Cheapest available competitor | $0.31/hr | The price to beat for sort position |
| Rented P50 median | $0.38/hr | Where the market actually clears |
Price to the asking average and you sit idle at $0.52. Match the cheapest competitor forever and you donate $0.07 per hour to renters who would have happily paid $0.38. The correct behavior uses both numbers: undercut the cheapest competitor while one exists below the median, and drift back toward the clearing price when the cheap competition gets rented and leaves the list. How that gap is managed is the subject of bid offset explained.
Comparable means comparable
One warning before you trust any median: it is only as good as the filter that built it. A rented P50 computed across every listing that says RTX 4090 mixes single cards with 8x rigs, verified datacenter machines with residential hosts, and US listings with every region on earth. Each of those splits moves the clearing price. A 4x rig rented as one instance often clears at a different per-GPU rate than four separate single-card rentals, and a verified machine with strong bandwidth clears above an unverified one with the same silicon.
So before comparing your machine to a median, make sure the median was computed over machines a renter would actually consider interchangeable with yours: same GPU type, similar GPU count, similar verification status, and a region filter that matches where your renters come from. A precise answer to the wrong question is worse than a rough answer to the right one.
Why most hosts never see this data
The asking side is easy to see: it is the public search page. The rented side takes work. You need to pull rented offer data, filter it to machines genuinely comparable to yours, compute the median, and then repeat all of that often enough to catch the moves, because a clearing price from last Tuesday is trivia, not intelligence.
Doing this by hand a few times is a good education. Doing it every few minutes forever is a job.
Clearing prices across platforms
The clearing price for the same GPU differs between marketplaces, because supply and demand differ. An RTX 4090 might be 90% utilized on one platform and oversupplied on another. If you are deciding where to list new hardware, or whether your current platform is underpaying you, compare utilization and pricing across markets before committing. Our GPU IQ guide covers how to read cross-platform supply, demand, and 24 hour utilization across six marketplaces.
Utilization is the other half of the equation
Put the number to work
Once you accept that the rented median is the real price of your hardware, your decisions get simpler. Your floor tells you whether the market is worth being in at all. The clearing price tells you what a rented hour should earn. The gap between your price and the cheapest competitor decides whether you get that hour. For the revenue math on why staying rented at $0.34 beats sitting pretty at $0.50, read utilization beats price. And if you want the platform's own view of the marketplace mechanics, the Vast.ai docs are the place to start.