NVIDIA L40S rental from €1.06/h
Renting an NVIDIA L40S on VRAMfarm costs from €1.06 an hour. It is a data centre card on the Ada Lovelace architecture (2023) with 48 GB of GDDR6 memory: data-centre Ada with 48 GB for inference and image generation. You pay by the minute from your balance, with no subscription.
To launch, your balance must cover one hour of the machine. You are billed by the minute for actual use.
What the L40S is good for
LLM inference
48 GB holds a model of up to about 21B parameters in FP16 and up to 72B in 4-bit, for example Llama 3.3 70B and Qwen 2.5 72B. The memory bandwidth of 864 GB/s sets how fast tokens are generated.
Image generation
Good for Flux.1 [dev] at full precision, SDXL with ControlNet and LoRA training without compromises.
Training and fine-tuning
Full training and full fine-tuning: the tensor cores compute in FP16 and BF16.
Rental cost by period
| Period | Cost |
|---|---|
| 1 hour | €1.06 |
| A day (24 hours) | €25 |
| A week | €178 |
| A month (720 hours) | €763 |
Prices updated 12 November 2026. While the page is open we show the price of the cheapest reliable machine available in the catalogue. Prices are in euros. How VAT applies to your account will be shown at checkout.
Similar GPUs
How to rent the L40S
- Sign up and top up your balance from €10. To launch, your balance must cover at least one hour of the machine.
- In the dashboard, click Rent a GPU, choose what to run (a ready-made app, PyTorch with Jupyter, vLLM or your own Docker image), then the L40S and a machine.
- The server is usually ready in 2–5 minutes. When you’re done, click Stop and billing stops. If you will need your models and results again, click Save and stop: /workspace is kept for €0.03 per GB per month and comes back on your next server.
Rental questions
How much does L40S rental cost?
From €1.06 an hour. Billing is per minute, with a minimum charge of 15 minutes per launch. To launch, your balance must cover one hour of the machine; the rest stays on your balance. Prices updated 12 November 2026.
Which AI models can I run on this card?
48 GB holds an LLM of up to about 21B parameters in FP16 or up to 72B in 4-bit, for example Llama 3.3 70B and Qwen 2.5 72B.
How do I access the rented GPU?
The server is usually ready 2–5 minutes after launch. Ready-made apps for images, video and chat open with a button in the dashboard, and for your own code you get SSH and Jupyter. Developers can choose PyTorch with Jupyter, vLLM, an empty ComfyUI or their own Docker image.
How do I pay?
Top up your balance in euros by card or SEPA transfer, from €10. Payments go through a payment provider that we will announce before launch. VAT, where it applies, is shown at checkout.
How does the L40S compare with similar cards?
A100: 38% more expensive, 32 GB more memory, FP32 ≈ 19.5 vs 91.6 TFLOPS. RTX 5090: 28% cheaper, 16 GB less memory, FP32 ≈ 104.8 vs 91.6 TFLOPS. L40: 57% cheaper, same memory, FP32 ≈ 90.5 vs 91.6 TFLOPS.
What happens to my data after I stop?
The server disk is wiped when you stop. If you need your models and results, click Save and stop: we keep /workspace for €0.03 per GB per month, and your next server gets it automatically. Saving does not work yet for custom Docker images.
Launch the L40S
Available machines with an hourly price, billed by the minute from your balance.