Skip to content
AI Compute Radar
IndexCollected 1 h ago

All GPU classes · GPU rental price

L40S rental price per hour

What an hour of L40S costs to rent right now on three rental sources, the list price at AWS, Oracle Cloud, Nebius and CoreWeave, what the price did over the last weeks, and when renting beats buying. Read from public listings; described, never forecast.

$0.80 typical on the Vast marketplace right now. From $0.79 at RunPod. About the same as last week.

Show the numbers

Vast marketplace, middle price of the last seven days: $0.80 per hour. Against the week before: 0 %. Against 30 days ago: 0 %.

Cloud list prices

L40S at the clouds: list price per hour

AWS, Azure, Oracle Cloud and Google Cloud, and the GPU clouds Lambda, Nebius, CoreWeave and Hyperstack, publish their prices without an account, so they are read here four times a day. A pay-as-you-go list price per GPU-hour: not a market median, and no statement about free capacity. The hyperscalers sell these GPUs as whole instances and Lambda in clusters of one to eight, so each row names the unit the price sits in.

L40S at the clouds: list price per hour
Cloud$/GPU-hSold as
AWS from$1.86g6e.xlarge1 × 48 GB per instanceUS East (N. Virginia) +2 regions
Oracle Cloud from$3.50BM.GPU.L40S.44 × 48 GB per instance$14.00/h for the whole instanceone list, no regions
Nebius from$1.55L40S with Intel CPU1 × 48 GB per instanceone list, no regions
CoreWeave from$2.25L40S8 × 48 GB per instance$18.00/h for the whole instanceNorth America +1 region

How much more does a big cloud cost?

On 22 Nov the cheapest big cloud for one L40S was AWS at $1.86 per GPU-hour: 2.3× the Vast.ai median of $0.80.

In plain words: for every dollar one L40S cost at the Vast.ai median that day, the cheapest big cloud listed $2.30.

How to read this
  • We call this the cloud premium.
  • What it is: the lowest list price per GPU-hour among AWS, Azure and Oracle Cloud, divided by the same day's Vast.ai median for the same card.
  • Where the numbers come from: both sides are read by this site, the clouds four times a day from their public price lists, the marketplace every hour; the day is a UTC day.
  • What it says: how much more a reserved, supported cloud instance cost than one GPU from a marketplace host on that day. The two are different products: the cloud comes with capacity guarantees, networking and support; the marketplace host is one machine somewhere.
  • What it does not say: nothing about speed, reliability or availability, and nothing about tomorrow. The factor can rise because the marketplace got cheaper, not because the cloud got dearer.
  • Example from this page: AWS listed $1.86, the Vast.ai median was $0.80, so the factor is 2.3×.

GPU price indices explained →

Times the Vast.ai price a big cloud listed for one L40S2.3× the market price
Range 2.3–2.5Oct 8 – Nov 22 · 46 readings

1.0× would be the same price. Higher means the cheapest big cloud cost more than the Vast.ai median that day.

The cloud premium divides the cheapest cloud list price per GPU-hour by the same UTC day's Vast.ai median, both measured here; a day without both readings has no factor. A list price buys a whole instance with reserved capacity and support, a marketplace median buys one GPU from one host, so the factor states what that difference cost on the day, not why, and never where it is going.

AWS: the lowest Linux on-demand price per GPU across the instance types in AWS's accelerator documentation and its public regions — an instance's hourly price divided by its documented GPU count. Read from the price files behind AWS's own on-demand pricing page, which carry the same publication as its bulk price list. GovCloud and the European Sovereign Cloud are left out.

Azure: the lowest Linux pay-as-you-go price per GPU across Microsoft's documented VM sizes and its public regions — a VM's hourly price divided by its GPU count from Microsoft's size documentation. Spot and low-priority prices and the US government regions are left out. The region shown is the first of those that list this price.

Oracle Cloud: the lowest pay-as-you-go price per GPU-hour among the class's bare-metal shapes on Oracle's public price list, which has no regions. The whole machine costs that price times its GPU count.

Azure and Oracle Cloud collected since 2 October 2026, AWS since 5 October, Lambda and Nebius since 19 October, CoreWeave since 24 October, Hyperstack since 27 October, Google Cloud since 3 November. A list price is its own kind of number: it stands beside the marketplace medians and is never blended with them. The cloud premium compares AWS, Azure and Oracle Cloud only.

AWS EC2 on-demand pricing ↗ (External link)Oracle Cloud price list ↗ (External link)Nebius pricing ↗ (External link)CoreWeave pricing ↗ (External link)

Window

Price history

L40S · Vast.ai0.80$/h
Range 0.80–0.80Nov 9 – Nov 22 · 14 readings

How many offers stand behind this price?

How many offers stood behind each day's median. A median over fewer than 5 offers is a reading, not a market; such days are flagged here and in the table above.

In plain words: how many hosts were actually offering this card that day. The fewer there were, the less a median tells you.

How to read this
  • We call this market depth.
  • What it is: the number of offers a renter could take at the top of the hour, as a median over the day's readings, for the same marketplace and card as the price chart above.
  • Why it is here: a median over two offers is one host's asking price, not a market price. When the count drops, the price chart can jump without anyone having paid more.
  • What the mark means: a reading under 5 offers is marked "few offers" next to the price. Treat its median as an example, not a rate.
  • What it does not say: nothing about how many GPUs are rented out, only how many were on offer at that moment.

How the Rent Index is computed →

Offers for one L40S on Vast.ai, per day15offers
Range 6–16Nov 9 – Nov 22 · 14 readings

Below 5 offers a day's price is one host's asking price, not a market price.

No day under 5 offers in this window.

Radar Rent Index

What the L40S rental price did

One line per source from the same snapshots as the tables above: the current window against the one before it, and against the first week collected. A source with too few days says so instead of showing a change.

What the L40S rental price did
SourceThis weekvs. last weekvs. 30 days agoSince start
Vast marketplacelatest day Nov 22$0.80median of 7 days0 %flat0 %0 %since Oct 8 · 46 days on record
Clore marketplacecollecting—collecting—collecting——collecting
RunPodlatest day Nov 22$0.79median of 7 days0 %flat0 %0 %since Oct 8 · 46 days on record
AWS fromlatest day Nov 22$1.86median of 7 days0 %flat0 %0 %since Oct 8 · 46 days on record
Oracle Cloud fromlatest day Nov 22$3.50median of 7 days0 %flat0 %0 %since Oct 8 · 46 days on record
Nebius fromlatest day Nov 22$1.55median of 7 days0 %flat0 %0 %since Oct 19 · 35 days on record
CoreWeave fromlatest day Nov 22$2.25median of 7 days0 %flat0 %0 %since Oct 24 · 30 days on record

No day was excluded.

A day is the median of its hourly medians, counting only hours with at least five offers (verified hosts on Vast.ai, free hosts on Clore.ai). A week is the median of its days and needs at least four of them. Words follow fixed thresholds: up or down from ±8 %, sharply from ±20 %; offers up or down from ±15 %. How the index is made →

What did the L40S price do over 30 days?

Per source: the lowest and the highest daily median of the last 30 days, the days they fell on, the average, and the swing between low and high as a share of that average. Glitch days are left out by the same rule as the Rent Index.

In plain words: over the last 30 days one L40S-hour at the Vast.ai median never cost less than $0.73 (Oct 28) or more than $0.80 (Nov 10); the gap between the two is 8 % of the average of $0.79.

How to read this
  • What it is: the lowest and the highest cleaned daily median within the window, and the swing, which is high minus low divided by the average of all days in the window.
  • Where the numbers come from: the same daily series as the Rent Index, read by this site; a day counts only when enough offers stood behind it, and a day far off its neighbours is left out as a glitch.
  • What it says: how wide the price moved in the window. A swing of 10 % means the dearest day cost a tenth of the average more than the cheapest day did.
  • What it does not say: nothing about where the price is going, and nothing about why it moved; a wide swing can come from hosts arriving and leaving as much as from anyone repricing.

How the daily series is made →

What did the L40S price do over 30 days?
SourceLowHighAverageSwing computed — Computed: arithmetic over published inputs, which are shown with it.Days
Vast.ai median$0.73Oct 28$0.80Nov 10$0.798 %30
RunPod list price$0.79Oct 24$0.79Oct 24$0.79unchanged30
AWS list price$1.86Oct 24$1.86Oct 24$1.86unchanged30
Oracle Cloud list price$3.50Oct 24$3.50Oct 24$3.50unchanged30
Nebius list price$1.55Oct 24$1.55Oct 24$1.55unchanged30
CoreWeave list price$2.25Oct 24$2.25Oct 24$2.25unchanged30

A list price with a swing of zero did not change in the window; that is a reading too. The marketplaces' rows are medians over the offers of each day, so a day with few offers can set a low or a high that one host asked for; the depth chart on this page shows how many offers stood behind each day.

The daily series behind this table, one row per day and source: 30 days as CSV · every day since 26 Aug 2026 as CSV · JSON · how to read it. CC BY 4.0, name the source.

Reading the L40S price

As of today's readingsOn November 22, 2026 the cheapest reading for one L40S-hour was $0.79 at RunPod (list price); the dearest listed was $3.50 at Oracle Cloud, 4.4 times as much. The Vast.ai median of $0.80 stood on 15 verified offers. 4 cloud list prices cover the class: AWS $1.86, Oracle Cloud $3.50, Nebius $1.55, CoreWeave $2.25 per GPU-hour. The Vast.ai week median stands unchanged against the week before and unchanged against 30 days ago. 31 of 49 tracked open models run comfortably in the card's 42.2 GB of usable memory at 8K context; the largest is Qwen AgentWorld 35B-A3B.

The L40S is the datacenter card for people who need 48 GB but not the price of an H100: an Ada-generation GPU that clouds offer in single-card instances, which is rare above the consumer tier. It is rented for inference of mid-sized models, for fine-tuning jobs that fit in one card, and for anything that needs more memory than a 4090 and a datacenter host's networking and uptime. The marketplaces carry it too, but in small numbers.

This is the class with the widest mix of source kinds: two marketplaces, RunPod's posted price, and the clouds that sell the card per GPU or in four-card machines. The cloud list prices differ by what they include: a single-card instance at one cloud, a bare-metal machine with four cards at another, so the row names the unit behind each price. Clore.ai often has only one or two hosts with the card; those readings are shown and marked, and they never set the row's cheapest price.

The fit table on this page uses the card's 48 GB less the usual runtime margin, which puts it between the consumer classes and the 80 GB datacenter cards: the 30-billion-parameter models that are tight on a 24 GB card run comfortably here, the largest open models do not. There is no break-even block, because the card has no consumer launch price to compare an hourly rate against. Clore.ai's reading for this class is often a single host; the page marks it and keeps it out of the cheapest-source highlight.

Price moves, last 30 days

Day-over-day moves of 10 % or more on a marketplace, and any change of a list price (RunPod, AWS, Azure, Oracle Cloud), between complete days.

No move of 10 % or more in the last 30 days.

All classes as an RSS feed →

Rent or buy

A single L40S has no sourced purchase price: it is a data-center part sold inside systems, so there is no honest break-even to print.

The guide: buy or rent a GPU? →

What 48 GB holds

31 of 49 tracked open models fit a single L40S at 8K context, measured file by file; the largest comfortable one is Qwen AgentWorld 35B-A3B.

Check a model in the finder →

Other classes

Prices by source

Vast.ai vs RunPod: the measured comparison →

How the index is made →

GPU price indices explained →

Keep reading