How Vast.ai Pricing Actually Works: Sort Order, Bids, and Winning Rentals

8 min readUpdated August 30, 2026
How Vast.ai Pricing Actually Works: Sort Order, Bids, and Winning Rentals

Vast.ai is a search engine for compute, and renters use it the way everyone uses a search engine: they look at the first page. Where your machine lands in that list is decided almost entirely by price. If you understand how the sort works, how on-demand and interruptible pricing interact, and what a renter actually sees in the quote, you understand most of what separates a machine that rents from one that idles.

Renters sort by price, and the sort is ruthless

When a renter searches for an RTX 5090 on Vast.ai, they get a list of offers filtered by their requirements and, in the common case, sorted by effective cost. Two listings with the same GPU, similar bandwidth, and similar reliability are separated by pennies per hour, and the cheaper one sits higher. Renters rarely scroll past the machines that look like good deals near the top.

This has a consequence hosts underestimate: a price that is one cent too high does not earn you one cent less. It can earn you nothing, because the rental goes to the machine listed above yours. Marketplace pricing is closer to an auction for position than to a menu.

The sort is not purely price. Reliability score, verified status, bandwidth, and machine specs all filter and weight results. But among comparable machines, price decides the order, and order decides who gets rented.

On-demand and interruptible: two prices, one machine

Every listing carries two prices that do different jobs:

  • On-demand price. The renter pays this rate and keeps the machine until they stop the instance. This is the price that competes in search results and the one most of your revenue flows through.
  • Interruptible (spot) price. Renters place bids for idle capacity. An interruptible job runs while nothing better is happening on the machine and gets paused when an on-demand rental or a higher bid arrives. It monetizes hours your on-demand price failed to fill.

A common mistake is treating these as one number. They are not. Your on-demand price is a competitive weapon that needs constant adjustment. Your interruptible price is a discount on top of it, and usually works best as a fixed percentage below on-demand. We cover the spot side in detail in the interruptible pricing guide.

A worked example: the RTX 5090 shelf

Say the market for a verified single RTX 5090 looks like this on a Tuesday afternoon:

ListingOn-demand $/hrStatus
Competitor A$0.35Available, top of the comparable sort
Competitor B$0.42Available
Your machine$0.45Available, sorted below both
Median of rented 5090s$0.40What renters are actually paying

At $0.45 you are third in line behind two cheaper machines, and above the price where rented 5090s are actually trading. You will win a rental only when both competitors are taken. To win the sort order you would price at $0.34, one cent under Competitor A. If Competitor A gets rented and disappears from the available list, the right move reverses: raise your price back toward $0.40, because now you are the cheapest and there is no reason to charge $0.34 when the market clears at $0.40.

That last step is where hosts lose the most money. Undercutting is easy. Remembering to raise your price the moment the competition clears out, at 3 AM, is not.

Where hosts lose rentals

  • Pricing against asking prices. The available list is full of machines priced on hope. The number that matters is what rented machines actually go for. See asking price vs rented price for why this distinction drives everything.
  • Stale prices. The market moves every time a machine gets rented, released, or repriced. A price set Monday is wrong by Tuesday, sometimes by Monday evening.
  • Hidden costs in the quote. Renters see storage and bandwidth pricing alongside your hourly rate. An aggressive storage price can quietly disqualify an otherwise cheap machine.
  • Racing to the floor. Blind undercutting scripts drag everyone down. Winning the sort by one cent beats winning it by ten.

The platform sets the guardrails

Vast.ai applies its own platform safeguards to pricing, and hosts set a price floor below which they refuse to rent. Everything a pricing agent does happens inside those bounds. The Vast.ai docs cover the host side of listing management if you want the platform's own description of the machinery.

Price is position, not just revenue

Think of your on-demand price as a coordinate in the search results rather than a revenue dial. The revenue comes from rented hours, and rented hours come from position.

It is also worth internalizing how fast the board changes. Every rental start removes a competitor from the available list, every rental end adds one back, and every host running any kind of automation reshuffles the prices in between. On a popular GPU type, the cheapest comparable competitor can change identity several times an hour. A pricing decision is not a setting; it is a perishable answer to a question the market keeps re-asking.

What to do with this

Three numbers deserve your attention before any strategy discussion: the cheapest comparable competitor, the median price of rented machines like yours, and your own break-even. The first two move all day. The third is yours to calculate once, and it anchors everything else. Start there with our guide on setting your price floor, then look at how the gap to your competitor is managed in bid offset explained.

Put your pricing on autopilot

The pricing agent watches the market around the clock and repositions your machines every few minutes. Setup takes about 3 minutes.

$5 / system / month · no contracts · cancel anytime