L40S cloud GPU pricing

Q: What is the cheapest L40S cloud GPU right now?

The cheapest published L40S rate on the panel is $0.550 per GPU-hour at Novita AI (L40S x 1, Not published). Every price links to the provider's own published rate.

Q: How much does a L40S cost per hour?

Across 35 providers tracked today, published per-GPU hourly rates for L40S run from $0.55 to $4.68, with a median of $1.39. Committed-use and reserved rates are excluded; everything shown is on-demand.

Q: How many providers rent L40S GPUs?

35 providers publish 58 distinct L40S configurations on Compute Cafe. 11 of those rows were reported in stock by the provider's live endpoint at last observation.

Q: Where is L40S cheapest?

By published per-GPU hourly rate, the cheapest geography for L40S today is india at $0.58/hr (AIC Cloud). The spread across 6 tracked geographies runs to $4.68/hr in china.

Q: Are L40S prices going up or down?

Over the 53 days Compute Cafe has tracked the panel, the L40S floor is rising: $0.380 to $0.550/GPU/hr (+44.7%). The median continuously-tracked listing moved from $1.29 to $1.29 (+0.0%). Daily snapshots are public in the repo, so the trend is auditable.

Q: What does a L40S cost per month?

At the current floor of $0.550/GPU/hr, a L40S running 24/7 costs about $402 per month (730 hours of on-demand arithmetic on the cheapest published rate). Committed-use contracts can price lower; the panel tracks on-demand rates only.

Q: Is a full 8-GPU L40S node cheaper than a single card?

On today's panel: the single-GPU floor is $0.550/GPU/hr and the 8-GPU node floor is $0.650/GPU/hr - more expensive per GPU on a full node (+18.2%). Check the table for the exact configs behind each floor.

Q: Is the L40S good for LLM inference?

It is the panel's default mid-market serving chip: models up to ~34B quantized run well, FP8 support keeps throughput high, and hourly rates are far below Hopper. Above that size, memory bandwidth becomes the constraint.

Q: L40S vs A100 for fine-tuning?

For 7B-13B LoRA work the L40S is typically cheaper per hour and fast enough. For 70B-class or memory-bandwidth-bound jobs, the A100 80GB's HBM2e wins regardless of list compute.

66 listings across 42 providers. Cheapest published rate: $0.550/GPU/hr at Novita AI. Sorted by per-GPU hourly price. Every row links to the provider's own published price.

WHAT IS THE L40S

The L40S is the workhorse of the inference-and-fine-tuning middle: Ada tensor cores with FP8, 48 GB of memory, and a price well below the H100 line. It has no NVLink, so multi-GPU scaling is PCIe-limited - but for single-card and small-fleet serving, it is the most common answer on the panel, and often the best $/token in the catalog.

MAKER

NVIDIA

ARCHITECTURE

Ada Lovelace

LAUNCHED

2023

MEMORY

48 GB GDDR6 ECC

MEMORY BANDWIDTH

864 GB/s

TENSOR COMPUTE

362 TFLOPS FP16/BF16 tensor dense (733 sparse); 733 TFLOPS FP8 dense

POWER DRAW

350 W

INTERCONNECT

PCIe 4.0 x16 (no NVLink)

WHAT IT IS GOOD AT

LLM inference up to ~34B

48 GB serves quantized 34B or BF16 13B per card with FP8 throughput to spare.

Fine-tuning 7B-13B

LoRA and full fine-tunes of small models run economically; 864 GB/s GDDR6 is the limit for memory-hungry jobs.

Image and video generation

Ada RT and tensor cores suit diffusion workloads; 48 GB handles large batch image pipelines.

WHO SHOULD NOT RENT THE L40S

70B-class training or serving - 48 GB GDDR6 and no NVLink are hard limits.

Pure raster/RT graphics pipelines - the L40 is tuned for that and usually costs less.

Provider	Config	VRAM	Location	$/hr	$/GPU/hr	Source
SpotGPUs	L40S × 1	48 GB	Not published	$0.250	$0.250STALE	receipt
Novita AI	L40S × 1	48 GB	Not published	$0.550	$0.550	receipt
AIC Cloud	L40S × 1	48 GB	Mumbai	$0.575	$0.575	receipt
Lium	L40S × 1	48 GB	Not published	$0.590	$0.590	receipt
Baku Compute	L40S (from) × 8	48 GB	Azerbaijan (Baku)	$5.200	$0.650	receipt
CloudRift	L40S × 1	48 GB	Not published	$0.690	$0.690	receipt
HydraHost	L40S × 1	48 GB	Not published	$0.700	$0.700	receipt
Beam	L40S (from) × 1	48 GB	Not published	$0.760	$0.760	receipt
GPUBrazil	L40S × 1	45 GB	North America	$0.782	$0.782	receipt
RunPod	L40S × 1	48 GB	Global	$0.790	$0.790	receipt
IO.NET	L40S × 1	48 GB	Not published	$0.790	$0.790	receipt
Applications.kz	L40S (from) × 1	48 GB	Kazakhstan	$0.790	$0.790	receipt
Vultr	L40S (bare metal) × 8	48 GB	Not published	$6.784	$0.848STALE	receipt
InHosted	L40S × 1	48 GB	India	$0.883	$0.883	receipt
packet.ai	L40S × 1	48 GB	US East · Virginia	$0.920	$0.920	receipt
Seeweb	L40S × 1	48 GB	Italy	$0.945	$0.945STALE	receipt
gpu.ai	L40S × 1	48 GB	Not published	$0.950	$0.950	receipt
Spheron	L40S × 1	48 GB	Not published	$0.960	$0.960	receipt
Massed Compute	L40S × 1	48 GB	Not published	$0.970	$0.970	receipt
Massed Compute	L40S × 8	48 GB	Not published	$7.760	$0.970	receipt
Hostnot GPU	L40S × 1	48 GB	US Central	$1.050	$1.050	receipt
Hostnot GPU	L40S × 1	48 GB	Asia-Pacific	$1.060	$1.060	receipt
Hostnot GPU	L40S × 1	48 GB	US Central	$1.060	$1.060	receipt
Hostnot GPU	L40S × 2	48 GB	Asia-Pacific	$2.120	$1.060	receipt
E2E Networks	L40S × 1	48 GB	India	$1.087	$1.087	receipt
Koyeb	L40S × 1	48 GB	Not published	$1.200	$1.200	receipt
UpCloud	L40S × 1	48 GB	Not published	$1.234	$1.234STALE	receipt
GPUBrazil	L40S × 1	45 GB	Europe	$1.266	$1.266	receipt
Civo	L40S × 1	48 GB	Not published	$1.290	$1.290	receipt
Civo	L40S × 2	48 GB	Not published	$2.580	$1.290	receipt
Civo	L40S × 4	48 GB	Not published	$5.160	$1.290	receipt
Civo	L40S × 8	48 GB	Not published	$10.320	$1.290	receipt
GPUBrazil	L40S × 1	45 GB	Asia-Pacific	$1.300	$1.300	receipt
Cyfuture	L40S × 1	48 GB	India	$1.321	$1.321STALE	receipt
DCP	L40S (from) × 1	48 GB	Saudi Arabia	$1.387	$1.387	receipt
Gcore	L40S (from) × 1	48 GB	Not published	$1.411	$1.411	receipt
GPUBrazil	L40S × 1	48 GB	Global	$1.444	$1.444	receipt
Hostnot GPU	L40S × 1	48 GB	Asia-Pacific	$1.450	$1.450	receipt
Hostnot GPU	L40S × 1	48 GB	US East	$1.450	$1.450	receipt
DigitalOcean	L40S × 1	48 GB	Not published	$1.570	$1.570	receipt
Vultr	L40S × 1	48 GB	Not published	$1.671	$1.671STALE	receipt
Machine0	L40S × 1	48 GB	Not published	$1.727	$1.727	receipt
OVHcloud	L40S (from) × 1	48 GB	Not published	$1.800	$1.800	receipt
AWS	L40S × 1	48 GB	US East (N. Virginia)	$1.861	$1.861	receipt
Neysa	L40S (from) × 1	48 GB	India	$1.950	$1.950	receipt
Cerebrium	L40S × 1	48 GB	Not published	$1.951	$1.951	receipt
Verda (formerly DataCrunch)	L40S × 8	48 GB	Europe (Finland / Iceland)	$15.980	$1.998	receipt
Verda (formerly DataCrunch)	L40S × 1	48 GB	Europe (Finland / Iceland)	$1.998	$1.998	receipt
Verda (formerly DataCrunch)	L40S × 2	48 GB	Europe (Finland / Iceland)	$3.996	$1.998	receipt
Verda (formerly DataCrunch)	L40S × 4	48 GB	Europe (Finland / Iceland)	$7.992	$1.998	receipt
Lightning AI	L40S × 1	48 GB	US	$2.140	$2.140STALE	receipt
CoreWeave	L40S × 8	48 GB	North America / Europe	$18.000	$2.250	receipt
Catalyst Cloud	L40S × 1	48 GB	New Zealand	$2.299	$2.299	receipt
Catalyst Cloud	L40S × 4	48 GB	New Zealand	$9.197	$2.299	receipt
Catalyst Cloud	L40S × 2	48 GB	New Zealand	$4.598	$2.299	receipt
Shakti Cloud	L40S (bare metal) × 4	192 GB	India (Navi Mumbai)	$10.144	$2.536	receipt
AWS	L40S × 4	48 GB	US East (N. Virginia)	$10.493	$2.623	receipt
Shakti Cloud	L40S × 2	96 GB	India (Navi Mumbai)	$6.319	$3.159	receipt
Shakti Cloud	L40S × 1	48 GB	India (Navi Mumbai)	$3.218	$3.218	receipt
Hinode	L40S (from) × 1	48 GB	Not published	$3.300	$3.300STALE	receipt
Oracle OCI	L40S × 1	48 GB	Not published (global list)	$3.500	$3.500	receipt
Replicate	L40S × 1	48 GB	Global	$3.510	$3.510	receipt
Replicate	L40S × 2	48 GB	Global	$7.020	$3.510	receipt
Replicate	L40S × 4	48 GB	Global	$14.040	$3.510	receipt
Replicate	L40S × 8	48 GB	Global	$28.080	$3.510	receipt
CTYun	L40S × 1	48 GB	China	$4.679	$4.679	receipt

THE L40S MARKET, BY THE NUMBERS

PROVIDERS

42

CONFIGURATIONS

66

MEDIAN $/GPU/HR

$1.39

SPREAD

8.5x

42 providers publish 66 L40S configurations on the panel today. Published per-GPU hourly rates run from $0.550 (Novita AI) to $4.68 (CTYun), a 9x spread between the cheapest and the most expensive published rate for the same chip. The median listing sits at $1.39/GPU/hr. 11 rows were reported in stock by the provider at last observation; the rest are published catalog rates. All rates are on-demand - committed-use and reserved pricing is excluded.

CHEAPEST L40S BY GEOGRAPHY

indiaAIC Cloud$0.575/hr united statesGPUBrazil$0.782/hr finlandVerda (formerly DataCrunch)$1.998/hr icelandVerda (formerly DataCrunch)$1.998/hr new zealandCatalyst Cloud$2.299/hr chinaCTYun$4.679/hr

L40S PRICE TREND, DAILY PANEL SNAPSHOTS

Across every L40S row on the panel, the daily floor moved from $0.380 to $0.550/GPU/hr over 53 days of tracking (+44.7%). The median continuously-tracked listing held flat at $1.29. Snapshots run daily; the full history is public in the repo.

L40S PROVIDER NOTES

Novita AI · 1 config

Sets the panel floor for this model.

from $0.550/hr

AIC Cloud · 1 config · 1 region

Within 15% of the panel floor.

from $0.575/hr

Lium · 1 config

Within 15% of the panel floor.

from $0.590/hr

Baku Compute · 1 config · 1 region

Published catalog rates.

from $0.650/hr

CloudRift · 1 config

Published catalog rates.

from $0.690/hr

HydraHost · 1 config

Published catalog rates.

from $0.700/hr

Beam · 1 config

Published catalog rates.

from $0.760/hr

GPUBrazil · 4 configs · 4 regions · 4 in stock

Widest regional footprint here - 4 regions.

from $0.782/hr

BUYING L40S: WHAT THE PANEL SAYS

At the current floor, one L40S running around the clock costs about $402 a month (730 hours of on-demand arithmetic). Node shape matters: the single-GPU floor is $0.550/GPU/hr against $0.650/GPU/hr on an 8-GPU node (+18.2%). Per-GPU floors by config size: 1x $0.55 · 2x $1.06 · 4x $1.29 · 8x $0.65. Compare the alternatives below before committing - the same budget often buys more than one chip class.

L40S VS THE ALTERNATIVES

L40S vs A100 80GB

L40S has more FP16 compute on paper, but GDDR6 bandwidth is less than half the A100's HBM2e and 48 GB caps model size. For 70B-class work the A100 wins; for smaller models the L40S is usually cheaper per token.

L40S vs L40

Same card with graphics-first tuning: the L40S trades some RT/raster throughput for roughly double the tensor compute. For AI work, always the L40S; for rendering, the L40.

L40S vs H100 PCIe

H100 PCIe adds HBM bandwidth, FP8 maturity and 80 GB - at 2-4x the hourly rate. L40S owns the middle market.

Compare live rates: L40S vs A100 80GB SXM · L40S vs A100 80GB · L40S vs L40 · L40S vs H100 PCIe · L40S vs L4 · L40S vs A40 · L40S vs RTX A6000 · L40S vs RTX 6000 Ada · L40S vs RTX 4090 · L40S vs RTX 5000 Ada · L40S vs L20 · L40S vs H100 SXM

L40S PRICING FAQ

What is the cheapest L40S cloud GPU right now?

The cheapest published L40S rate on the panel is $0.550 per GPU-hour at Novita AI (L40S x 1, Not published). Every price links to the provider's own published rate.

How much does a L40S cost per hour?

Across 42 providers tracked today, published per-GPU hourly rates for L40S run from $0.55 to $4.68, with a median of $1.39. Committed-use and reserved rates are excluded; everything shown is on-demand.

How many providers rent L40S GPUs?

42 providers publish 66 distinct L40S configurations on Compute Cafe. 11 of those rows were reported in stock by the provider's live endpoint at last observation.

Where is L40S cheapest?

By published per-GPU hourly rate, the cheapest geography for L40S today is india at $0.58/hr (AIC Cloud). The spread across 6 tracked geographies runs to $4.68/hr in china.

Are L40S prices going up or down?

Over the 53 days Compute Cafe has tracked the panel, the L40S floor is rising: $0.380 to $0.550/GPU/hr (+44.7%). The median continuously-tracked listing moved from $1.29 to $1.29 (+0.0%). Daily snapshots are public in the repo, so the trend is auditable.

What does a L40S cost per month?

At the current floor of $0.550/GPU/hr, a L40S running 24/7 costs about $402 per month (730 hours of on-demand arithmetic on the cheapest published rate). Committed-use contracts can price lower; the panel tracks on-demand rates only.

Is a full 8-GPU L40S node cheaper than a single card?

On today's panel: the single-GPU floor is $0.550/GPU/hr and the 8-GPU node floor is $0.650/GPU/hr - more expensive per GPU on a full node (+18.2%). Check the table for the exact configs behind each floor.

Is the L40S good for LLM inference?

It is the panel's default mid-market serving chip: models up to ~34B quantized run well, FP8 support keeps throughput high, and hourly rates are far below Hopper. Above that size, memory bandwidth becomes the constraint.

L40S vs A100 for fine-tuning?

For 7B-13B LoRA work the L40S is typically cheaper per hour and fast enough. For 70B-class or memory-bandwidth-bound jobs, the A100 80GB's HBM2e wins regardless of list compute.