Rent NVIDIA L40S GPU servers by the month

When you rent NVIDIA L40S capacity from us, you get one whole Ada Lovelace GPU with 48 GB of GDDR6 in a dedicated virtual machine, for $890 per month. It is our entry-level data-center GPU: large enough for 7B–14B language models in BF16, image generation and GPU rendering, and it is in stock now.

What the L40S is, and where it fits in 2026

The L40S is NVIDIA's Ada Lovelace data-center GPU. Per NVIDIA's specifications, it has 48 GB of GDDR6 at 864 GB/s and Tensor Cores with FP8 support at 733 TFLOPS dense (1,466 TFLOPS with sparsity), plus 362 TFLOPS of dense BF16. It is a 350 W PCIe Gen4 card without NVLink. Unlike compute-only GPUs such as the H100, it also carries RT cores and NVENC video encoders, so the same card can serve models, render scenes and transcode video.

In 2026 it sits at the value end of our data-center lineup. The H100 has about 3.9 times its memory bandwidth and NVIDIA lists 2.7 times its dense FP8 compute, which matters for large models and high concurrency. The L40S makes sense when your model fits in 48 GB and the monthly bill matters more than peak throughput. Below it in price, the GeForce RTX 5090 has more bandwidth (1.79 TB/s) but only 32 GB; the L40S's extra 16 GB is what lets a 14B model run in BF16 with a useful KV cache, or an image pipeline keep its text encoders resident. Our L40S vs H100 and RTX PRO 6000 vs L40S comparisons cover both directions.

What fits in 48 GB

Weights take roughly 2 bytes per parameter in BF16, 1 in FP8 and 0.5 in 4-bit, plus KV cache and 10–30% headroom.

• 7B–8B models in BF16: about 16 GB of weights, leaving roughly 25 GB for KV cache after headroom: comfortable for chat, RAG and agent backends.
• 14B models in BF16: about 28 GB, with around 15 GB left for KV cache. In FP8, the same model takes about 14 GB.
• 32B models: about 32 GB in FP8, which works for short contexts and modest concurrency; in 4-bit, at about 16 GB, there is plenty of room.
• 70B models: FP8, at about 70 GB, does not fit. 4-bit, at about 35 GB, does, but leaves too little KV cache for production use. Move to a 96 GB RTX PRO 6000, or to an H200 for more KV cache, instead.
• Image generation: SDXL, with about 3.5B parameters, runs with large batches, and a 12B-parameter model such as FLUX.1, about 24 GB in BF16, fits together with its text encoders.

For training, LoRA on an 8B model in BF16 and QLoRA on models up to about 32B fit comfortably. QLoRA on 70B is possible but tight. Full fine-tuning with Adam needs roughly 16–20 bytes per parameter, so on 48 GB it is limited to small models of around 1–2B parameters. Our article on how much VRAM LLMs need explains the KV cache side of the math.

Rendering, video and graphics workloads

The L40S's RT cores accelerate GPU renderers such as Blender Cycles with OptiX, Redshift and Octane. The scene has to fit in the GPU's 48 GB, or the renderer falls back to slower out-of-core modes where it supports them; 48 GB is enough for many look-dev and mid-size scenes. The NVENC encoders handle video transcoding next to AI work, for example encoding the output of a generation pipeline on the same machine. If your scenes outgrow 48 GB, the RTX PRO 6000 doubles the memory. See rendering and VFX for more on render workloads.

Software on an L40S server

The VM comes with root access over SSH and Ubuntu 22.04 or 24.04 with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit preinstalled; Debian or Rocky Linux are available on request. Ada's Tensor Cores support FP8, and both vLLM and TensorRT-LLM use it on this GPU, so FP8 checkpoints run on the hardware FP8 path. A typical small-model endpoint:

vllm serve <your-8b-model> --quantization fp8 --max-model-len 32768 \
  --host 127.0.0.1 --port 8000

Binding to localhost keeps the endpoint private until you put your own authentication in front of it. Our guide to self-hosting inference with vLLM goes further.

What it costs to rent NVIDIA L40S

The L40S costs $890 per month, about $1.22 per GPU-hour if it runs all 730 hours. For an always-on endpoint, a render queue or a batch pipeline that never empties, that flat price is the whole bill. For occasional jobs it is not: at 100 hours of use a month, the effective cost is $8.90 per used hour, and hourly billing elsewhere will usually cost less. The break-even point is simply $890 divided by the hourly rate you would pay.

One sizing trap to avoid: two L40S plans cost $1,780 per month for 96 GB split across two GPUs, while one RTX PRO 6000 costs $1,390 for 96 GB on a single card. If your model needs more than 48 GB, the larger card is both simpler and less expensive, because nothing has to be split across GPUs.

The price includes 20 TB of outbound traffic per month with no egress fees within that allowance, a public IPv4 address and an IPv6 /64, DDoS mitigation, 24/7 support from engineers and the 99.9% availability SLA. There is no setup fee. The L40S is offered as a single-GPU plan only; if you rent several, private networking between your servers is available on request.

Ordering an L40S

Send a request on the site. We confirm in writing the configuration, monthly price, location and delivery date; you pay the first month by card or in crypto; access details arrive by email. If we miss the confirmed delivery date, you can cancel for a full refund, and you can cancel any later month with no fee, the server running until the end of the paid month. Questions about sizing go through the contact form.