What an NVIDIA L40S costs per hour in November 2026, read on eight providers' own pages
I run VoltageGPU, which rents L40S GPUs by the hour , so this is a comparison written by one of the vendors in it. Our row comes from our public pricing feed ( /api/pricing/snapshot , built at 07:45 UTC on 20 November 2026, 9 L40S on demand at that moment). Every other row comes from the provider's own pricing page, read on 19 November 2026 between 20:56 and 21:06 UTC; I read the prices again on…
In November 2026, the NVIDIA L40S GPU is available for rent from eight different providers, with prices ranging from $0.79 to $1.95 per GPU-hour. The NVIDIA L40S is based on the Ada Lovelace architecture, featuring 48 GB of GDDR6 memory with ECC, a bandwidth of 864 GB/s, and a power consumption of 350 W. The card supports FP8 Tensor Cores, which offer a denser format for FP16, enabling higher performance with more efficient memory usage.
When considering the price per hour, the cost varies significantly depending on the provider and the specific configuration of the GPU. RunPod, for instance, charges $0.79 per GPU-hour on-demand, while Nebius, with an AMD CPU, offers a spot price of $0.74 per GPU-hour, making it the cheapest option. On the other end of the spectrum, Modal charges $1.95 per GPU-hour for serverless functions, with the price being CPU-based.
The amount of RAM available per GPU varies significantly as well, ranging from 64 GB to 180 GB. The amount of RAM can be a critical factor in the cost of the GPU, as it determines the size of the model that can be run on the card. For example, an 8B or 14B model can fit into the 48 GB of memory with room to spare when using BF16 (2 bytes per parameter) or 4-bit precision (0.5 byte per parameter). However, a 32B model would require the use of FP8 or 4-bit precision to fit within the memory constraints.
The L40S GPU does not support NVLink, which means that if a model needs to be split across multiple cards, data must be exchanged via PCIe. Additionally, there is no Multi-Instance GPU (MIG) support, which means that only one workload can be run on each card. This can impact the overall performance and efficiency of the GPU when running multiple models or workloads simultaneously.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.