Every hour your machine sits available at your on-demand price is an hour that earned nothing. Interruptible pricing exists to sell those hours to renters who will trade reliability for a discount. Priced well, spot rentals turn idle time into revenue without touching your on-demand strategy. Priced badly, they either never fire or quietly undercut your own listing.
How interruptible rentals work on Vast.ai
Alongside your on-demand price, your listing accepts interruptible bids. A renter names an hourly bid, and their job runs while the machine has nothing better to do. Two things can stop it: an on-demand rental arrives, or another interruptible renter outbids them. The interrupted job pauses rather than dies, and can resume when capacity frees up. The Vast.ai docs describe the bidding mechanics from the renter's side, which is worth reading once, because your customers on this tier think in bids, not list prices.
The renters are a distinct crowd: batch training runs, rendering queues, checkpoint-friendly workloads, and price-sensitive experimenters. They tolerate interruption because their jobs can survive it. What they will not tolerate is a spot price close to on-demand, because then the discount does not pay for the risk.
Why hosts should care about the second price
The economics are simple and lopsided:
- An idle hour earns $0.00 while still costing you electricity for a machine at the wall and a slot in your attention.
- A spot hour at a 10% discount earns 90% of your on-demand rate, on capacity that on-demand renters were not buying anyway.
- Spot renters never block on-demand revenue, because on-demand rentals preempt them by design.
The catch is that the spot price needs to track your on-demand price. If your on-demand price moves every few minutes with the market and your interruptible price was set once in January, the relationship between them drifts: some weeks the discount is so small no spot renter bites, other weeks it is so large you sell hours at a needless markdown.
Auto mode: a discount that follows your price
GPU Autopilot manages the interruptible price in one of two modes:
- Auto. The spot price is held at a percentage discount below your current on-demand price, 10% by default. When the agent repositions your on-demand price, the spot price follows. The relationship you chose stays true through every market move.
- Fixed manual price. You set an exact spot price and it stays put. Useful when you have a firm number in mind for the machine's idle hours and want it independent of on-demand swings.
Note that your bid offset applies to the on-demand price only. The interruptible side is governed by the discount or your manual number, not by competitor gaps.
Worked example: what the two tiers earn
A single RTX 4090 with an on-demand price of $0.38 per hour and Auto mode at the default 10% discount, over a 720 hour month:
| Scenario | Hours | Rate | Revenue |
|---|---|---|---|
| On-demand rentals | 430 hrs | $0.38/hr | $163.40 |
| Idle, no spot price set | 290 hrs | $0.00 | $0.00 |
| Same idle hours, 60% filled by spot | 174 hrs | $0.342/hr | $59.51 |
The spot fill rate is the market's call, not yours, and 60% is an illustration rather than a promise. But the structure of the trade is fixed: whatever fraction of idle hours spot renters take, it is revenue added on top of an on-demand number that did not change.
Who takes these hours? Workloads that checkpoint. A training run that saves state every 15 minutes loses almost nothing to an interruption; it resumes from the last checkpoint when capacity frees up. Render farms, hyperparameter sweeps, batch inference over large datasets, and any queue-shaped job fit the same profile. Interactive sessions and jobs with hard deadlines do not, which is why those renters pay your on-demand rate instead. The two tiers segment your customers by urgency, and you collect from both.
Keep the discount honest
One piece of arithmetic worth checking before you pick a discount: how close your on-demand price can get to your break-even. A price floor of $0.13 with an on-demand price at $0.14 during a slow week means a 10% discount implies roughly $0.126 for spot hours, right at the edge of costing you money. If your floor and your typical slow-week price sit close together, choose a smaller discount, or set a fixed manual spot price you have verified against your own break-even math.
The bigger picture: filling hours beats raising rates
Interruptible pricing is one instrument in a theme that runs through all of GPU hosting: rented hours are worth more than impressive listing prices. A machine earning $0.34 around the clock beats one listed at $0.50 that rents half the time, and spot revenue is the cheapest utilization you will ever buy, because it costs you hours you were not selling. For the full argument with the monthly math, read utilization beats price. And if the two-price system is still new to you, start with how Vast.ai pricing actually works to see how on-demand and interruptible fit together in the search results renters see.