NVIDIA GPU Rental Prices in 2026, Compared

NVIDIA GPU rental prices in 2026 split into a floor and a median. On Vast.ai an RTX 4090 is $0.14/hr at the floor and $0.49/hr at the median, about 3.5× apart. H100 SXM hours median $3.06. Budget the median. The floor is one machine, not a plan. Salad lists 4090-class hours from $0.16.

01. Why do NVIDIA GPU rental prices disagree with the screenshot?

The floor is one listing. The median is what people pay.

On a 4090 those two numbers are $0.14/hr and $0.49/hr. The low screenshot is one Vast.ai listing: one host, one card, one disk, one price. The rows above it step up. The middle of the list is the median.

The GPU rental median vs floor is why two screenshots disagree. One post shows $0.14/hr. A colleague paid near $0.49/hr. Same chip name, different row. For a budget you will still accept next month, type the median. Keep the floor as a sighting of a single machine.

A copied floor can be gone by the time you need the hour. Plan on the middle of the board.

The 3.5× gap is a 4090 fact, not a law for every chip.

3.5× is the 4090 median divided by the 4090 floor. It does not travel with the chip name. A 5090, a PRO 6000, a 3090, and an H100 each have their own spread in the table.

RTX 5090: $0.27/hr floor, $0.60/hr median, about 2.2×. RTX PRO 6000: $1.20/hr to $1.49/hr, about 1.2×. H100 SXM: $1.73/hr to $3.06/hr, about 1.8×. RTX 3090: $0.10/hr to $0.17/hr, about 1.7×. Budget the multiple on the row you will actually rent.

Vast.ai’s board is large enough for a median to mean something.

Vast.ai lists about 17,000 GPUs, about 1,400 hosts, and about 700,000 rentals a month. On a board that size, the median is the middle listing. Use it as the month’s planning figure.

Hosts still change rates. The machine you click can sit above or below the middle. What the scale supports is planning on that middle. A floor pulled from a wide board is still one machine.

Vast.ai is the source of every dollar figure in the table. The A100 row stays blank. No verified A100 hourly price belongs in this set. Read floor and median as a pair. The right-hand column is there so 3.5× stays on the 4090 row, whose neighbors sit at 1.2×, 1.7×, 1.8×, or 2.2×.

02. What does a 4090, a 5090, or a 3090 hour cost?

A 3090 hour has a $0.17 median.

The RTX 3090 rental price to plan on is a $0.17/hr median and a $0.10/hr floor. The card is 24GB. A 4090 is 24GB as well. Inside that ceiling you are paying for speed. Take the 3090 when the weights fit and the queue can wait. The hour you skip is the gap between $0.17 and $0.49. If the render is what the day is for, a slower card at 24GB can eat the saving by running longer.

A 4090 hour has a $0.49 median.

The RTX 4090 rental price to write down is $0.49/hr. That is the planning number for image and short video. The $0.14/hr floor is one listing, about 3.5× under the median.

Each rented hour of image work, on the card you actually get, is still a $0.49 hour. A short calendar does not turn the floor into the planning number.

A second 4090 is not a 48GB pool. The 4090 has no NVLink. Each card keeps its own 24GB unless the software splits the work. Price the second card as a second hour at the same median.

A 5090 hour has a $0.60 median.

The RTX 5090 rental price is a $0.60/hr median, from a $0.27/hr floor. The payment is for 32GB. Heavier video, and pipelines that already run out of 24GB, are the jobs that step up. When 24GB finishes the shot, the extra 8GB sits idle at the higher median.

03. What does an H100 hour cost next to a 4090 hour?

About six 4090 medians fit in one H100 median.

$3.06 against $0.49. The H100 rental price per hour on Vast.ai is an H100 SXM median of $3.06/hr, with a $1.73/hr floor, about 1.8× from floor to median. About six hours at the 4090 median line up with one hour at the H100 median.

The six-to-one figure compares invoices. It does not claim a speed multiple. You pay for 80GB on a datacenter training card. A job that fits in 24GB does not need that row. An H100 filter is how a small image team picks up the large bill.

PRO 6000 sits in between, at a $1.49 median, with 96GB.

The RTX PRO 6000 median is $1.49/hr. The floor is $1.20/hr, about 1.2× under it. One card holds 96GB. Use the row when the weights need that much memory on a single workstation-class GPU. A stack of ordinary images does not. Read $1.49 only after the 96GB is the requirement.

A100 prices are absent on purpose.

An A100 is a datacenter card at 40GB or 80GB. Compare it with an H100 SXM on the 80GB figure. This table prints no A100 dollar amount. Use the live listing that day. The H100 median is not a stand-in. Which card the job needs is the comparison in 4090 vs A100 vs H100.

04. How do Vast.ai, Salad, Clore.ai, and China cloud listings compare?

These are different products that share a chip name.

Vast.ai, Salad, Clore.ai, and a China cloud can all print “4090” and still be four products. The host, the payout, and the network differ. One averaged market price would blend a median, an entry price, a commission range, and a domestic cloud menu. Cite the figure each place actually publishes.

Salad’s published entry is $0.16/hr for a 4090-class card.

Salad lists 4090-class hours from $0.16/hr. Salad reports about 450,000 registered suppliers and about 60,000 GPUs online a day. Salad pays suppliers largely in gift-card balance. That balance is not cash against a power bill.

No Salad median appears on this page. $0.16 stays an entry price. The 4090 number to plan with is still Vast.ai’s $0.49/hr median.

Clore.ai lists about 20,000 GPUs, with commission from 1.8% to 10% and token subsidies.

About 20,000 GPUs, commission from 1.8% to 10%, plus token subsidies. That is the Clore.ai figure set. A subsidy denominated in a token is not an hourly cash rate. No Clore.ai hour price is printed here, and none is estimated from the commission band. Read the payout rule before you count a subsidy as electricity money.

One China cloud posts 4090 hours from about ¥1.32, and one offer is ¥1.86/hr with 150 free hours and 200+ images.

One domestic provider lists RTX 4090 hours from about ¥1.32/hr. Another offer is ¥1.86/hr, with 150 free hours and 200+ preinstalled images. Both are China cloud menus on a datacenter network. They are not a price for a home connection rented across a border, and they do not replace the Vast.ai median.

The 150 free hours change the first bill on that one offer. They do not add a median to the card table. After those hours, use the stated rate.

05. Why is the 4090 median so far above China’s power cost?

A US or EU host pays $0.10–$0.22 per hour in electricity for a 600W rig before any margin.

Power only, on about a 600W rig, before the card is recovered. The China figure in the same model is $0.047/hr. Renters pay the listing. The power line is the host’s cost, folded into that listing.

That is the structural reason a $0.14 floor is fragile.

A $0.14/hr floor and a US or EU power cost of $0.10–$0.22/hr leave a high-power host with little left for the card. That host cannot live on the floor. A China host at $0.047/hr has a wider gap under the $0.49 median. The wider gap is a cost fact. It does not put the floor into the budget.

A ¥1.32 China cloud menu and the $0.49 Vast.ai median are separate columns. One is a domestic datacenter price. The other is a marketplace listing whose US or EU host is already paying $0.10–$0.22/hr to power a 600W rig.

Electricity is why $0.49 sits far above $0.047. It still does not set the sticker. The host has to recover the card. Budget the median.

The full power table sits on the rental-vs-buying page.

Region-by-region rates for the same 600W rig are on the page about renting vs buying the card. Here, the power sentences stop at $0.10–$0.22/hr for a US or EU host and $0.047/hr for China. That page is where an owned rig gets compared. This page is the rental board.

06. How should a team write the number into a budget?

Write the median of the card you will actually get.

Name the card, then copy its median. 24GB image or short video: $0.49/hr on a 4090, or $0.17/hr on a 3090 if the job can wait. 32GB video that no longer fits: $0.60/hr on a 5090. 96GB on one card: $1.49/hr on a PRO 6000. 80GB training: $3.06/hr on an H100 SXM. A100: live listing only.

Write the stop next to the rate. The median is for an hour that is doing the job. An hour left running after the files are downloaded bills again at that rate.

Ignore a planning band until it is a listing.

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.

That band is about $0.20–$0.25/hr. List price unset. Rentals are not open. Do not drop it into a budget beside the Vast.ai medians until it is a real listing.

Revisit this table when you reuse the numbers elsewhere.

Other pages on this site that quote a 4090 median or an H100 median are quoting the table in section 01. $0.49 and $3.06 stay tied to those rows. If a median has to change, it changes here first, so the site does not carry two middles for one card.

FAQ

What is a fair NVIDIA GPU rental price for a 4090?

$0.49/hr is the Vast.ai median to plan on. $0.14/hr is the floor, about 3.5× lower, and it is one listing.

Why is an H100 hour about $3?

The Vast.ai H100 SXM median is $3.06/hr, with a $1.73 floor. You pay for 80GB and a training-class card, not a faster 4090.

Is Salad’s $0.16 the same product as Vast’s median?

No. Salad shows 4090-class hours from $0.16, and pays suppliers largely in gift-card balance. Vast’s planning number remains the $0.49 median. This page has no Salad median.

Is SpareHive in this price table?

No. Rentals are not open. The planning band is roughly 50–60% under the $0.49 median, and the list price is unset.