Try
10 GPU-hours
$8.20
$0.82/hour
- Test your setup end to end
- Same GPU and image as a long run
- Unused hours never expire
Quick answer
How much does an L40S cost per hour?
As of 29 October 2026, an NVIDIA L40S 48GB costs from $0.97 to $3.15 per GPU-hour on demand across 11 providers, with a median of $1.67. On GPUwithCrypto it costs $0.82/hour, billed per minute and paid in crypto with no KYC. Marketplace L40S offers go lower: $0.33 in Vast.ai's public price feed.
How much does an L40S cost per hour at each provider?
One row per provider at its cheapest on-demand L40S configuration, sorted by price. Seven of the 11 providers charge between $1.50 and $1.89; the most expensive costs 3.2 times the cheapest.
| Provider | Configuration | On-demand $/GPU-hour | Spot / marketplace $/GPU-hour | Billing increment | Accepts crypto | ID / KYC |
|---|---|---|---|---|---|---|
| GPUwithCrypto | L40S 48GB, 1 to 8 GPUs | $0.82 | — (on-demand only) | Per minute | Yes | No |
| Massed Compute | L40S, 1 GPU | $0.97 | — | Hourly rate (increment not stated) | Yes (BTC) | None documented |
| RunPod | L40S, 1 GPU | $1.09 | $0.79 (Community Cloud) | Per second | Yes | KYC before first crypto payment |
| Crusoe | L40S | $1.50 | Contact sales | Per hour or per second | No | None documented |
| Nebius | L40S, Intel host ('from' price; AMD host $1.82) | $1.55 | From $0.74 (preemptible) | — | No | None documented |
| DigitalOcean (Paperspace) | L40S, 1 GPU | $1.57 | — | Per second, 5-minute minimum | Yes (stablecoins, limited) | None documented |
| Vultr | L40S, 1 GPU (2 and 4 GPUs at the same rate) | $1.67 | — | Hourly | Yes (BitPay) | None documented |
| OVHcloud | L40S, 1 GPU | $1.80 | — | Per second, 1-minute minimum | No | Sometimes |
| AWS | L40S, 1 GPU (g6e.xlarge) | $1.86 | — | — | No | None documented |
| Verda (formerly DataCrunch) | L40S, 1 GPU | $1.88 | $0.94 (spot) | — | No | None documented |
| CoreWeave | L40S, 8-GPU instance ÷ 8 | $2.25 | $0.99 (spot) | — | Unknown | Unknown |
| Spheron | L40S | $3.15 | $1.18 (spot) | Per minute, 20-minute minimum | Yes (USDC) | None documented |
| Vast.ai | L40S, marketplace | — | $0.33 (Vast.ai's public price feed, cheapest rentable offer) | Per second | Yes (BitPay, Crypto.com) | None documented |
| Shadeform | L40S, 2 GPUs, resold from other clouds | — | $1.50 (cheapest available listing) | — | Yes (USDC, per Stripe) | None documented |
Prices checked on
USD per GPU-hour, taxes excluded, rounded to the cent. The AWS price covers the whole instance (US East). Vultr's figure comes from its public plans API, which did not confirm availability. 'None documented' means no ID step in the official pages we read, not a promise that a provider never asks.
L40S vs A100: which is the better deal per hour?
For FP8 inference and image generation, the L40S. According to GPUwithCrypto's price check of 29 October 2026, a median on-demand L40S costs 25% less than a median A100 80GB ($1.67 against $2.23). It also has FP8 Tensor Cores, which the A100 lacks.
The A100 keeps two advantages: 80 GB of memory against 48 GB, and 2,039 GB/s of bandwidth on the SXM version against 864 GB/s. A job that needs more than 48 GB, such as a 70B model in 4-bit with a 32K-token context (about 53 GB), needs the A100. On GPUwithCrypto the two cost $0.82/hour and $0.92/hour. The L40S vs A100 comparison goes further.
Why are marketplace and spot L40S cheaper than ours?
Because a host sets the marketplace price, and the provider can stop a spot GPU. On 29 October 2026, Vast.ai's public price feed listed 61 rentable L40S offers, from $0.33 with a median of $0.73. Nebius listed preemptible L40S from $0.74, and RunPod's community-hosted Community Cloud charged $0.79.
Those three prices were below our $0.82 that day. They fit batch jobs that can restart. For a server that must stay up, compare on-demand prices: ours is 51% below the on-demand median. Instances are on-demand, not spot: nobody can take your GPU back while your balance covers it, and the instance runs until you stop it.
How much does an L40S cost per day, week or month?
Our price for one GPU next to the lowest and the median self-serve on-demand price of 29 October 2026, for the same number of hours.
| Period | GPU-hours | GPUwithCrypto | Lowest elsewhere (Massed Compute) | Median elsewhere |
|---|---|---|---|---|
| 1 hour | 1 | $0.82 | $0.97 | $1.67 |
| 1 day | 24 | $19.68 | $23.28 | $40.10 |
| 1 week | 168 | $137.76 | $162.96 | $280.73 |
| 30 days | 720 | $590.40 | $698.40 | $1,203.12 |
Prices checked on
The median is $1.671 per GPU-hour before rounding. Our day, week and month match the 24, 168 and 720 GPU-hour blocks.
What are the L40S's specs?
| Architecture | Ada Lovelace, PCIe Gen4, dual slot |
|---|---|
| Memory | 48 GB GDDR6 |
| Memory bandwidth | 864 GB/s |
| FP8 Tensor | 733 TFLOPS dense, 1,466 with sparsity |
| FP16 / BF16 Tensor | 362.05 TFLOPS dense, 733 with sparsity |
| FP32 | 91.6 TFLOPS |
| NVLink | No |
| Max power | 350 W |
What fits on one L40S?
Memory needed for common open models, using weights × 1.2 for inference (EleutherAI's rule of thumb) and weights plus adapters for LoRA. The VRAM guide has the formulas.
| Model and job | Memory needed | One L40S 48GB? |
|---|---|---|
| Llama 3.1 8B inference, BF16 | ≈ 19.2 GB | ✓ |
| Llama 3.1 8B LoRA fine-tune | ≈ 19.2 GB | ✓ |
| Llama 3.1 70B inference, 4-bit | ≈ 42 GB | ✓ |
| Qwen2.5 72B inference, 4-bit | ≈ 43.6 GB | ✓ |
| Llama 3.1 70B QLoRA fine-tune | ≈ 52.5 GB | Borderline: the QLoRA authors fine-tuned a 65B model on one 48 GB GPU |
| FLUX.1 [dev] transformer, BF16 weights | 24 GB, plus text encoders and VAE | ✓ |
| FLUX.1 [dev] LoRA training | ≈ 28.8 GB | ✓ |
| SDXL 1.0 base, FP16 weights | 7 GB | ✓ |
| Llama 3.1 70B inference, FP8 | ≈ 84 GB | No: 2× L40S |
| Llama 3.1 8B full fine-tune | ≈ 128 to 144 GB before activations | ✗ |
What does an L40S cost on GPUwithCrypto?
$0.82/hour, at least 15% under its named on-demand reference: Massed Compute at $0.97, read on 29 October 2026. We do not claim to be the cheapest for every GPU; spot and marketplace offers (Vast.ai) can be lower. Instances are billed per minute at the hourly rate while they run, from the GPU-hours you bought for that GPU. Stopping an instance stops the charge at once. GPU-hours are sold in blocks of 10, 24, 168 or 720 hours for one GPU model, at the same hourly rate; instances of that GPU draw from them per minute, and unused hours never expire.
Popular
Day
24 GPU-hours
$19.68
$0.82/hour
- One day of continuous running
- A fine-tuning run or a batch job
- Unused hours never expire
Week
168 GPU-hours
$137.76
$0.82/hour
- One week of continuous running
- Training runs and evaluations
- Unused hours never expire
Month
720 GPU-hours
$590.40
$0.82/hour
- 30 days of continuous running
- An inference server running all month
- Unused hours never expire
In every plan: Whole GPUs, never shared · Root SSH, CUDA and PyTorch ready · Billed per minute · Unused hours never expire
L40S price questions
What is the cheapest L40S cloud?
On 29 October 2026 the lowest self-serve on-demand L40S price we found was $0.97 at Massed Compute, and GPUwithCrypto charges $0.82. Vast.ai's public price feed showed marketplace offers from $0.33, set by independent hosts.
How much does an L40S cost per month?
720 GPU-hours cost $590.40 on GPUwithCrypto. At the median on-demand price of 29 October 2026 the same month cost $1,203.12, and $698.40 at the lowest.
Can an L40S run a 70B model?
Yes in 4-bit: Llama 3.1 70B needs about 42 GB and fits in 48 GB. In FP8 it needs about 84 GB, so two L40S. The L40S has no NVLink, so multi-GPU runs go over PCIe.
Is the L40S good for ComfyUI and Flux?
It has the memory for it. FLUX.1 [dev] weights take 24 GB in BF16, and LoRA training needs about 28.8 GB, both inside 48 GB. See the best GPU for ComfyUI and Flux.
Can I rent an L40S with crypto?
Yes. GPUwithCrypto takes BTC, USDT, USDC, ETH, SOL, LTC with no ID and no card. Of the 11 self-serve L40S providers we checked, 5 accept crypto: Massed Compute, DigitalOcean, Vultr, Spheron and RunPod. RunPod asks for KYC before the first crypto payment.
Sources (17)
- Massed Compute — VM pricing vm.massedcompute.com/pricing — Massed Compute (read 2026-10-29)
- Runpod — Pricing runpod.io/pricing — RunPod (read 2026-10-29)
- Crusoe — Cloud pricing crusoe.ai/cloud/pricing — Crusoe (read 2026-10-29)
- Nebius — Prices nebius.com/prices — Nebius (read 2026-10-29)
- DigitalOcean — GPU Droplets pricing digitalocean.com/pricing/gpu-droplets — DigitalOcean (read 2026-10-29)
- Vultr — Public plans API api.vultr.com/v2/plans?type=all&per_page=500 — Vultr (read 2026-10-29)
- OVHcloud — Public Cloud prices ovhcloud.com/en/public-cloud/prices — OVHcloud (read 2026-10-29)
- AWS — EC2 on-demand price list, US East (N. Virginia) — Amazon Web Services (read 2026-10-29)
- Verda — Pricing verda.com/pricing — Verda (read 2026-10-29)
- CoreWeave — Pricing coreweave.com/pricing — CoreWeave (read 2026-10-29)
- Spheron — Pricing spheron.network/pricing — Spheron (read 2026-10-29)
- Vast.ai public GPU pricing feed — Vast.ai (read 2026-10-29)
- Shadeform — Instance types API api.shadeform.ai/v1/instances/types — Shadeform (read 2026-10-29)
- Runpod — Billing information docs.runpod.io/references/billing-information — RunPod (read 2026-10-29)
- NVIDIA L40S GPU — NVIDIA (read 2026-10-29)
- QLoRA: Efficient Finetuning of Quantized LLMs — Dettmers et al. (read 2026-10-29)
- Transformer Math 101 — EleutherAI (read 2026-10-29)