Rent NVIDIA L40S GPUs by the minute.
Inference, fine-tuning, graphics. 48 GB per GPU, billed by the minute.
Ada LovelaceDatacenter
L40S
Inference, fine-tuning, graphics
- VRAM
- 48 GB
- From
- $2.01/h
Finland
GPU Cloud · Secure
L40S rates per GPU-hour.
| GPU | VRAM | Countries | On demand | SpotInterruptible |
|---|---|---|---|---|
| L40S | 48 GB |
| from $2.01/h | Not in stock |
Prices on Nov 14, 2026, 20:35 UTC. US dollars, per GPU, before tax.
Community machines. Priced by their owners, per hour. Each rate in the console.Explore Community
The L40S, in numbers.
- Architecture
- NVIDIA Ada Lovelace
- GPU memory
- 48 GB per GPUGDDR6
- Memory bandwidth
- 864 GB/s
- Interconnect
- PCIe Gen4
- Form factor
- L40S
- Image
- Ubuntu 24.04 LTS, NVIDIA driver 570, CUDA 12.8
- Billing
- By the minute
What fits in 48 GB.
Open models on one L40S, or split across up to 8.
| Model | 16-bit | 4-bit |
|---|---|---|
| Llama 3.1 8B Instruct | 1 GPUNeeds 19.5 GiB | 1 GPUNeeds 6.1 GiB |
| Mistral 7B Instruct v0.3 | 1 GPUNeeds 17.8 GiB | 1 GPUNeeds 5.7 GiB |
| Gemma 2 27B IT | 2 GPUsNeeds 63.5 GiB | 1 GPUNeeds 17.9 GiB |
| Qwen2.5 32B Instruct | 2 GPUsNeeds 74.8 GiB | 1 GPUNeeds 20.4 GiB |
| Mixtral 8x7B Instruct | 4 GPUsNeeds 106.0 GiB | 1 GPUNeeds 27.7 GiB |
| Llama 3.3 70B Instruct | 4 GPUsNeeds 160.3 GiB | 1 GPUNeeds 42.0 GiB |
| Qwen2.5 72B Instruct | 4 GPUsNeeds 165.0 GiB | 1 GPUNeeds 43.1 GiB |
| FLUX.1 [dev] | 1 GPUNeeds 38.9 GiB | 1 GPUNeeds 10.5 GiB |
| Whisper large-v3 | 1 GPUNeeds 4.5 GiB | 1 GPUNeeds 1.9 GiB |
Estimated by the GPU calculator for a 4,096-token context and one request: weights, KV cache and runtime overhead. Launch by workload
What runs well on the L40S.
- LLM inferenceAn OpenAI-compatible endpoint with vLLM.
- Fine-tuningTrain on your data. Keep the weights.
- 3D renderingBlender and Cycles, headless.
- RecipesvLLMServe an open-weight LLM with vLLM on a Node Lodge machine. Install with uv, start the OpenAI-compatible server, query it through an SSH tunnel.
- RecipesComfyUIRun ComfyUI on a Node Lodge GPU. Install with uv, download a checkpoint, open the web UI on your computer through an SSH tunnel.
Next to the L40S.
| GPU | VRAM | Memory | Architecture | From |
|---|---|---|---|---|
| RTX PRO 6000 | 96 GB | GDDR7 | Blackwell | $2.16/h |
| L40SThis page | 48 GB | GDDR6 | Ada Lovelace | $2.01/h |
| L40 | 48 GB | GDDR6 | Ada Lovelace | $1.11/h |
About the L40S.
- How much does an NVIDIA L40S cost per hour?
- From $2.01 per GPU-hour on demand on Nov 14, 2026, 20:35 UTC, in US dollars before tax. Billed per minute from prepaid credit, with no minimum.
- How much memory does the L40S have?
- 48 GB of GDDR6 per GPU. Memory bandwidth: 864 GB/s.
- Where are L40S GPUs located?
- Countries: Finland. In partner datacenters, under Node Lodge Secure. The console shows the city of each machine.
- Which models run on one L40S?
- At 16-bit: Llama 3.1 8B Instruct, Mistral 7B Instruct v0.3, FLUX.1 [dev], and Whisper large-v3. At 4-bit: Llama 3.1 8B Instruct, Mistral 7B Instruct v0.3, Gemma 2 27B IT, Qwen2.5 32B Instruct, Mixtral 8x7B Instruct, Llama 3.3 70B Instruct, Qwen2.5 72B Instruct, FLUX.1 [dev], and Whisper large-v3. Size any other model with the GPU calculator.
Request access.
Tell us what you will run. Your access arrives by email.