A rented GPU costs whatever the hours you hold it cost, so the honest answer comes in three lenses: the hour on the rate card, the day as a stack of held hours, and the month as a 7×24 commitment. The 2026 anchors: an RTX 4090 rents at a $0.42/hr median across the platforms Coloprice tracks (Vast.ai's tracked median for the same card is $0.49/hr), an RTX 3090 medians at $0.17/hr on Vast.ai, and an H100 medians near $3.39/hr across the whole market on the GetDeploying index. Hold a median-priced 4090 around the clock and Coloprice's published monthly figure is about $307.
This article gives you the three tiers with their sources, explains why the lowest number you saw is probably not the number to budget, and separates the workloads that burn seconds from the ones that burn days.
01. What are you paying for — the hour, the day, or the month?
The hourly rate is the rate card; the day and the month are held time.
Marketplace listings meter by the hour, some by the second. A day has no separate price on those listings: a day is the hours you hold the machine, awake or asleep, at the hourly rate. A month is the same arithmetic held to 7×24, which is why the monthly figure describes a committed workload rather than a booking.
A committed month at the 4090 median is about $307.
That is Coloprice's own monthly figure for the RTX 4090: the $0.42 median hour held around the clock comes to roughly $307 for the month. Read it as a ceiling, not a typical invoice. Round-the-clock is the most a median-priced 4090 can bill; a job that holds the meter a few hours a day lands far under it, and a forgotten idle instance is the usual way a bill creeps toward that ceiling.
The three lenses answer three different questions.
The hour answers "what am I paying while this runs," and it is the number every rate card leads with. The day answers "what did this job actually cost me," because a job's bill is the hours it held the machine, including the ones after you stopped watching. The month answers "what does it cost to keep this card busy around the clock," which is the right question only for always-on workloads. Quotes that answer one question with a number from a different lens are where most sticker confusion starts.
02. What does an RTX 4090 cost to rent?
Coloprice's tracked 4090 range runs $0.13–$0.74/hr, with the median at $0.42.
Coloprice's RTX 4090 board (October 2026 snapshot) covers the on-demand listings across the platforms it tracks:
| Number | What it is |
|---|---|
| $0.13/hr | Low end of the tracked on-demand range — one listing |
| $0.42/hr | Median across the tracked platforms |
| $0.74/hr | High end of the tracked range |
| ≈$307/month | Coloprice's 7×24 figure at the median hour |
A range this wide means sorting listings by price alone is sorting by luck. The same tracker's vendor notes put RunPod's Community tier at $0.34/hr and its Secure tier at $0.74/hr for a 4090 in mid-October — named vendor prices you can find on a rate card, rather than the edges of the tracked set.
A second tracker prints a different median, and that is normal.
Vast.ai's tracked 4090 median is $0.49/hr. $0.42 and $0.49 are both defensible, because a median counts whatever listings that tracker counts. Pick one source, name it, and budget from it — then stay with that source when you compare cards, because a cross-tracker comparison inherits exactly the gap you are trying to avoid.
SpareHive is not in the live price table.
SpareHive is not in the live price table. The planning band sits roughly 50–60% under the $0.49 RTX 4090 median, list price unset, rentals not open. Join the waitlist at sparehive.com.
03. What does an RTX 3090 cost to rent?
Vast.ai tracks the 3090 at a $0.10 floor and a $0.17 median.
| Vast.ai RTX 3090 | Price |
|---|---|
| Floor | $0.10/hr |
| Median | $0.17/hr |
At the median, the 3090 runs about a third of Vast.ai's $0.49 4090 median while carrying the same 24GB of memory. The $0.10 floor is a single listing, and it obeys the floor rules described in section 05.
The 3090 is the 24GB tier for jobs that can wait.
The card handles image workloads and small fine-tunes; what it gives up against the 4090 is speed, not memory. Both cards carry the same 24GB ceiling, so a workflow that fits a 4090 usually fits a 3090 — the queue just moves slower. If the queue tolerates an overnight run, the $0.17 median is the lowest planning number in this article. If the job is interactive, the slower hour can cost more than it saves, and the 4090's median is the better fit despite the higher rate.
04. What does an H100 cost to rent?
The lowest tracked H100 hours start at $1.69.
gputracker's April 2026 sweep of the "cheapest GPU cloud" rankings — 54 providers, 5,269 instances — put the low H100 hours at:
| Provider (April 2026 sweep) | H100 hour |
|---|---|
| RunPod (spot) | $1.69 |
| Vast.ai | $1.85 |
| RunPod (on-demand) | $2.49 |
| Lambda | $2.49 |
| CoreWeave | $2.89 |
| AWS p5 (per-GPU equivalent) | $4.08 |
Treat the top rows as market floors from a single sweep: real prices at collection time, liable to be gone at booking time. How each provider's full tier list stacks up is our GPU rental pricing comparison.
The all-market H100 median sits near $3.39, in a $2–$12 market.
GetDeploying's index (quoted by Akash, September 2026) puts on-demand H100 between $2 and $12 across the market, with single-card listings mostly $2–$7, and the median near $3.39. The pattern is the one you saw on the 4090: a floor you might catch and a median you can plan on, at datacenter prices.
05. Why is the cheapest number you saw different?
Full-market medians, vendor from-prices, and marketplace floors are three different numbers.
A search result usually mixes them. The full-market median is every listing a tracker counts — Coloprice's $0.42 4090, GetDeploying's $3.39 H100. The vendor from-price is a named provider's own tier — RunPod's Community 4090 at $0.34. The marketplace floor is one machine at one moment — $0.13 on Coloprice's 4090 range, $0.10 for a Vast.ai 3090, $1.69 for a spot H100. For the same card, these numbers sit two to three times apart, which is how two honest articles can disagree.
The floor is one machine's luck; the median is what a budget can hold.
A floor listing depends on one host keeping one machine available at that price. It can be gone by the time you book, and nothing about it generalizes to next week. A median is the middle of what people actually pay across listings; it wobbles, but it survives well enough to plan a month on. Budget medians, and treat floors as a discount you might get.
Every price you repeat should carry a source and a date.
That is the discipline this article follows: each number above names its tracker and its month. A practical test before you bookmark a low figure: ask what it would take to reproduce it — a tracker's live board, a provider's rate card, or a screenshot from a forum post. Only the first two can be checked again a week later, which is the property a number needs before a budget leans on it.
06. What does the workload do to the bill?
Image jobs are metered in seconds; training runs are metered in hours and days.
A single image generation takes seconds of GPU time, so for image work the hourly rate is rarely what decides the bill — whether the meter gets stopped is. A training run is the opposite shape: the meter runs for hours or days while the job grinds, so the hourly rate, uptime discipline, and checkpointing dominate. Same card, same rate card, completely different invoice.
Match the billing lens to the shape of the job.
Bursty image work thinks in hours: book, run, stop at the download. Multi-day runs think in days, and the month lens matters only when a workload approaches 7×24 — the one case where the ≈$307 figure is the right comparison. When you find yourself renting the same tier month after month, that is where the rent-versus-buy arithmetic starts; we ran it in cloud GPU rental vs buying.
FAQ
How much does it cost to rent a GPU per hour?
$0.42/hr at the RTX 4090 median on Coloprice's October 2026 board ($0.49 on Vast.ai's), $0.17/hr for a 3090 on Vast.ai, and near $3.39/hr for an H100 across the market on GetDeploying's index.
How much does it cost to rent a GPU for a month?
About $307 for a 4090 held 7×24 at the $0.42 median — Coloprice's published monthly figure. Treat it as a ceiling: workloads that hold the meter a few hours a day bill far less.
Is the lowest 4090 price I saw the real price?
No. $0.13/hr is the floor of Coloprice's tracked range — one listing, one machine. Budget the $0.42 median and treat the floor as a possible discount.
Why do two sites quote different medians for the same card?
Because a median counts whatever listings that tracker counts, and no two trackers count the same set. Coloprice prints $0.42 for the 4090, Vast.ai's board prints $0.49 — name the source and month with any number you repeat.