NVIDIA L40S cloud prices

The cheapest L40S hour you could actually launch on 16 October 2026 was GPUBank's interruptible tier at $0.33, which includes 16 vCPU, 94 GB of RAM, a 1 TB NVMe disk, an IPv4 address and all traffic, with five minutes of notice before the card is taken back 1. Novita lists spot at $0.28, but it was out of stock in every size 3. For capacity that cannot be reclaimed, GPUBank and Novita tie at $0.55 on demand, followed by Beam at $0.76 and Runpod's Community Cloud at $0.79, the only peer-hosted L40S in stock that week 10 9. Vast.ai's own L40S page showed no current offers and Lium no longer lists the card, so the lower marketplace figures quoted in September were not available 7 4. GPUBank's top-up bonus lowers its effective on-demand hour to about $0.34 with a $1,000 top-up 1.

An L40S costs between about $0.28 and $3.77 per GPU-hour in October 2026, depending on the mode and on what the price includes. GetDeploying put the median on-demand price at $1.54 across 133 configurations on 15 October, with reserved offers at a median of $1.09 and spot at $1.11 3. The cheapest whole-machine on-demand price was $0.55 at GPUBank and Novita, Runpod's Secure Cloud costs $1.09, Civo $1.29, Nebius about $1.55 once CPU and RAM are added, and AWS $1.861 for a g6e.xlarge with only 4 vCPU, or $3.77 per GPU on its eight-card instance 1 9 15 16 22 23. Before comparing two prices, check whether each covers vCPU, RAM, disk, a public IP and egress, and whether the card is sold alone or only as part of an eight-card node.

A full 730-hour month of one L40S costs $401.50 on GPUBank on demand, $321.20 on its monthly reservation and $240.90 on its interruptible tier, plus $0.03 per GB-month for any persistent disk 1. Elsewhere the same month costs $394.20 to $416.10 on a CloudRift reservation of three months or one, $576.70 on Runpod's Community Cloud, $795.70 on its Secure Cloud and $1,358.53 on an AWS g6e.xlarge before EBS and egress 14 9 22. At GetDeploying's $1.54 median, a month comes to about $1,124 3. Whole-server deals only pay off if you fill them: LeaderGPU's eight-card server at €3,290 a month is about €411 per card, more than GPUBank's single-card reservation 18. Paying GPUBank's on-demand hours from a balance built with $1,000 top-ups brings the cash cost of a month to about $251.

The L40S is good for serving 7B to 30B models to many users and weak at single-user token speed. Its 864 GB/s of GDDR6 caps generation: Runpod measured 48 tokens per second per user on Qwen2.5-7B with vLLM, against 61 on an RTX 4090, 96 on an A100 SXM and 101 on an RTX 5090 27, and Koyeb measured 46 on Llama 3.1 8B at batch 1 29. Batching changes the picture: Koyeb's L40S reached 1,153 tokens per second at batch 32, and in a vLLM test of a 27B model at 4-bit with 32 concurrent requests it delivered 554 tokens per second where the 24 GB RTX 4090 stalled at 300 for lack of KV cache 28. Its 48 GB and FP8 support let it hold a 30B model in FP8 with room for context. For a chat product where one user waits on every token, an RTX 5090 or an A100 is faster for the money.

An L40S can run a 70B model at 4-bit, slowly and with a short context. XiongjieDai's llama.cpp tables measured 15.31 tokens per second on Llama 3 70B in Q4_K_M on a single L40S, against 22.11 on an 80 GB A100 PCIe, while the 16-bit version did not fit on either card and the RTX 4090 could not load the 4-bit one 30. A 4-bit 70B model takes about 40 GB for its weights, which leaves only a few gigabytes of the 48 GB for the KV cache, so long prompts or several users at once will not fit. Two L40S cards share the weights and leave room for a longer context, but they talk over PCIe without NVLink, so tensor-parallel scaling is imperfect 5. For regular 70B serving, a 96 GB RTX PRO 6000 at $1.15 or an 80 GB A100 SXM at $0.95 on GPUBank is simpler 1; for occasional experiments one L40S at $0.55 is enough.

Rent the RTX 4090 if your model fits in 24 GB with room to spare, and the L40S if it does not. On GPUBank the 4090 costs $0.23 an hour against $0.55, and in Runpod's tests it generated tokens faster for a single user, 61 per second against 48, while transcribing Whisper turbo at about $6.53 per 1,000 audio hours against $9.43 on the L40S 1 27 25. The L40S earns its premium when memory is the limit: Qwen-Image or FLUX in full bf16, large ControlNet stacks, 14B to 30B models in FP8, or batched serving, where a vLLM test showed the 4090 capped at nine concurrent requests by its KV cache while the L40S kept 32 in flight 26 28. The L40S also has ECC memory and a datacenter licence, whereas NVIDIA's GeForce driver licence excludes datacenter deployment 6.

The L40S is better than the A100 for image generation and transcription, and worse for LLM token generation and training. In Runpod's tests it rendered SDXL in 4.55 seconds against 7.65 on an A100 SXM and transcribed Whisper turbo at 58.3 times realtime against 34.0 26 25, which on GPUBank's prices makes about $0.70 per 1,000 SDXL images against $2.02 1. The A100's 80 GB of HBM2e at about 2 TB/s more than doubles the L40S's bandwidth, which is why it generated 96 tokens per second per user on Qwen2.5-7B against 48 and reached 2,083 tokens per second at batch 32 on Llama 3.1 8B against 1,153 27 29. It also has NVLink, which the L40S lacks 5. Pick the A100 at $0.95 for LLM serving, fine-tuning and anything that needs 80 GB; pick the L40S at $0.55 for diffusion, video, audio and embeddings.

The L40 and the L40S are both Ada Lovelace datacenter cards with 48 GB of GDDR6, but the L40S is the later, AI-oriented version with higher Tensor Core throughput, rated at 733 dense FP8 TFLOPS 5, and it is the one providers market for inference and generative work. On bandwidth-bound work the two behave alike: Runpod's vLLM test measured both at 48 tokens per second per user on Qwen2.5-7B 27. On compute-bound work the S model pulls ahead: in Runpod's Whisper test the L40 transcribed turbo at 51.7 times realtime against the L40S's 58.3 25. The L40 usually rents for less, with GetDeploying's cheapest in-stock L40 at $0.78 on 15 October 34, and some listings and aggregators mix the two names up, so check the exact model on the provider's page before you compare a price with the figures on this page, which all refer to the L40S.

Interruptible L40S capacity is worth it for any job that checkpoints, such as batch transcription, embedding runs, image batches or training that saves its state, and not for an interactive session or a live endpoint. GPUBank's interruptible tier costs $0.33 against $0.55 on demand, gives five minutes of notice through the metadata endpoint, a webhook and the console, runs a checkpoint hook, stops the machine rather than deleting it and credits the last 30 minutes if the notice is missed 1. Nebius preemptible costs about $0.75 with its CPU and RAM, Seeweb's spot gives two minutes of notice, and AWS spot for a g6e.xlarge stood at $1.829, barely below on-demand 16 12 22. In our transcription scenario, interruptible capacity cut the bill from $97.33 to $59.60, and a reclaim costs only the work done since the last checkpoint.

An L40S costs $0.79 an hour on Runpod's Community Cloud and $1.09 on its Secure Cloud, both with 16 vCPU and 94 GB of RAM per card and billed per second, and $1.75 an hour as a serverless worker in the 48 GB class 9. Community Cloud machines belong to independent hosts and Secure Cloud machines sit in Runpod's partner datacenters; GetDeploying showed the first in stock in 1 and 2-card sizes and the second in 1, 2 and 4-card sizes on 16 October 3. Storage is billed on top: $0.07 per GB-month for a network volume under 1 TB, $0.10 for a volume disk while the pod runs and $0.20 once it is stopped, so 200 GB of models kept on an idle pod adds $40.00 a month. A full 730-hour month comes to $576.70 on Community Cloud and $795.70 on Secure Cloud before storage, against $401.50 on GPUBank's on-demand L40S with 16 vCPU, 94 GB of RAM, a 1 TB NVMe disk and traffic included 1.

Vast.ai can be a cheap place to rent an L40S when it has listings, but on 16 October 2026 its own L40S page showed no current offers at all 7. It is a marketplace: each card belongs to a host who sets the price, the CPU, the disk, the bandwidth rate and the uptime, so availability for a less common card like the L40S comes and goes with a few hosts. Its billing documentation says storage is charged every second an instance exists, including while it is stopped, and bandwidth at each host's rate in both directions 8. When listings return, filter for verified datacenter hosts, check the reliability score and the bandwidth price, and keep data you cannot lose on storage you control. We found no published automatic credit for a host that fails mid-job, whereas GPUBank credits the interrupted session in full 1.

AWS and Oracle offer the L40S, and we found no public L40S instance on Google Cloud or Microsoft Azure as of 16 October 2026. AWS has sold it since 15 August 2024 in the G6e family, from the g6e.xlarge at $1.861 an hour in us-east-1 with one card, 4 vCPU and 32 GiB, to the g6e.48xlarge with eight cards at $30.131 an hour, or $3.77 per GPU; a three-year reservation brings the g6e.xlarge down to $0.804 22 23. Oracle sells the BM.GPU.L40S.4 bare-metal shape with four cards at $14.00 per node-hour, $3.50 per GPU, according to the trackers we could read 24. These clouds make sense when your data and services already live there or when you need RDMA networking between L40S nodes. On price per GPU-hour they cost three to seven times what a specialised cloud charges for the same card.

A new L40S costs about $9,700 to $11,000 from resellers in October 2026, with occasional sales near $7,700,, and used cards trade at $6,750 to $8,750. SabrePC listed the PNY card at $9,701.76 in stock on 16 October, Cloud Ninjas at $10,584 with ten in stock, CDW at $10,497.99 and Server Supply at $11,000, and clustermelt.com's tracker put the used range at $6,750 to $8,750 on 1 October 31 32 35 36 33. The card is passively cooled and needs a server chassis with forced airflow, so budget for the machine around it as well. With an assumed $5,000 server, two years of use and a resale of $6,350, the hardware costs about $348 a month before power, which means one card does not pay for itself against GPUBank's $0.55 rental even at full-time use 1. Buying makes sense for teams that will run it around the clock for four years or more, or that must keep the hardware on their own premises.

The L40S is a strong ComfyUI card, especially when models outgrow 24 GB. In Runpod's ComfyUI tests it rendered Juggernaut XL, an SDXL checkpoint, in 4.55 seconds, FLUX.2 Klein 9B in bf16 in 3.31 seconds, and Qwen-Image 2512 in bf16 in 38.91 seconds while using 41.4 GB, a model that runs from system memory on cards under about 46 GB 26. Under shared load it sustained 694 SDXL images an hour. On GPUBank's $0.55 that comes to about $0.70 per 1,000 SDXL images and $0.51 per 1,000 FLUX.2 Klein images 1. The RTX PRO 6000 Blackwell Server Edition is about twice as fast at roughly the same cost per image, and for plain SDXL a cheaper 24 GB card is usually enough. GPUBank's ComfyUI templates, including one for Wan 2.2 video, launch on the L40S directly.

Whisper runs at about 58 times realtime on an L40S for a single stream of large-v3-turbo in int8, the fastest of 23 cards in Runpod's 2026 faster-whisper test, narrowly ahead of the H100 NVL at 57.8 and the RTX 5090 at 53.0 25. One hour of audio takes about 62 seconds. The full large-v3 model in int8 ran at 23.1 times realtime, and peak memory stayed under 2.5 GB in int8 and 4.5 GB in fp16, so Whisper never fills the card. On GPUBank's on-demand price that comes to about $9.43 per 1,000 audio hours for turbo and $23.81 for large-v3, and about 40 percent less on interruptible capacity 1. Batching several streams on one card raises throughput well beyond these single-stream figures. If cost per audio hour is all that matters, an RTX 4090 at $0.23 is cheaper, at about $6.53 per 1,000 hours.