Deploy NVIDIA L40S and L40 GPUs

The L40S and L40 are two Ada Lovelace data center cards built around the same 48 GB of GDDR6 memory at the same 864 GB/s of bandwidth, so the choice between them comes down to compute, not capacity. Every offer on this page is a VM you can launch now with SSH access, billed by the hour and settled to the second when you stop.

Offers now

12

From

$0.97/GPU-hr

GPUs per VM

1x, 2x, 4x, 8x

Providers

2

Massed Compute, QuantaCloud · checked 2026-10-16 03:18 UTC

L40S and L40 offers you can launch now

The first table lists every L40S VM in stock and the second every L40, cheapest first, refreshed every 15 minutes.

For L40S prices across the whole market, compare all providers on GpuPerHour. For a reference price over time, see the L40S rental index (GPUIndexes).

L40S or L40

Specifications from NVIDIA's L40S page and the NVIDIA L40 datasheet. The memory is identical between the two cards. What differs is compute: NVIDIA rates the L40S at roughly double the L40's FP8 Tensor Core throughput, for an extra 50 W of power draw. For AI training, fine-tuning and LLM inference, that makes the L40S the better default. The L40 remains a solid choice for rendering, video and graphics work where the extra Tensor Core throughput matters less.

What fits in 48 GB

These are weights-only sizes of the published checkpoints, calculated rather than measured. The space left over is what vLLM has for the KV cache after its default 92% memory limit, and that cache grows with context length and the number of concurrent requests.

A 48 GB card comfortably holds a 20B-class model on its own. Past that, most current dense or mid-size MoE checkpoints in BF16 need a second card, which is where the 80 GB tier starts to look simpler for a single, larger model.

How long until it is ready

A VM usually reaches Running in 1 to 3 minutes. App templates take about 6 to 8 minutes, because the platform waits until the app itself answers before it marks the deployment ready. On some providers the machine reports Running a little before SSH accepts connections, so a Connection refused in the first minute usually clears on its own. The Connection refused fix covers what to check if it does not.

Paid time starts when provisioning starts, but only a launch that reaches Running is charged. A launch that fails refunds the reserved hour automatically.

One GPU or eight

L40S and L40 VMs come in fixed sizes of 1, 2, 4 or 8 GPUs, and the GPUs column shows what is in stock right now. Neither card has NVLink, so the Interconnect column reads PCIe at every size. One VM is the largest unit DeployGPU launches: there is no self-serve multi-node cluster. For more than 8 GPUs, or a fixed term at a fixed price, request a cluster quote and a person replies by email. The multi-GPU page goes deeper on how GPUs inside one node talk to each other over PCIe.

Software on the machine

In the dashboard, every L40S or L40 VM boots an Ubuntu image with NVIDIA drivers and Docker installed by the provider, and you connect over SSH as the user shown on the deployment. Three app templates set up a ready app on a private HTTPS link instead: JupyterLab with PyTorch, ComfyUI and Open WebUI with Ollama. For an OpenAI-compatible server on your own card, follow the vLLM guide.

What you pay

The price on each row is the hourly rate you are billed, including any platform fee. Launching reserves the first hour from your credit balance, and when you stop, the unused seconds of that hour go back to the balance. Credits are bought by card with a 5% + $0.30 processing fee on each top-up. The pricing page has the full breakdown.

Stopping an L40S or L40 deletes its disk, so copy results off first. The guide to stopping a GPU without losing work shows how.