NVIDIA L40S · 48 GB · updated 24 October 2026, 18:09 UTC
L40S rental price:$0.80 an hour today.
A 48 GB data-center card from NVIDIA's Ada generation (2023), common for serving models and making images. What it costs to rent one on other platforms right now, and how the price has moved.
Elsewhere today
$0.80 an hour on Vast.ai.
The typical price of one free L40S on Vast.ai, a GPU marketplace, as of 24 October 2026, 18:09 UTC. We read it every 6 hours, counting every verified, on-demand offer. All cards
- Typical price on Vast.ai
- $0.80/hr
- Middle half of prices
- $0.80–$0.80
- Cheapest available now
- $0.47/hr
- Free on Vast.ai now
- 32
- RunPod, cheapest in stock
- $0.79/hr
Price by day
The typical L40S price, last 90 days.
The line is the typical price each day; the shaded band is the middle half of prices. Hover over a day for its numbers. We started reading on 13 October 2026.
Models it runs
Open models that fit in 48 GB.
Each at the best quality that fits, with room for a short conversation. All models
- Llama 3.1 8B (16-bit): about 18 GB
- Llama 3.3 70B (4-bit): about 44 GB
- Qwen3 8B (16-bit): about 18 GB
- Qwen3 14B (16-bit): about 31 GB
- Qwen3 32B (8-bit): about 36 GB
- Qwen3 30B-A3B (8-bit): about 34 GB
- Qwen3.8 27B (8-bit): about 31 GB
- gpt-oss-20b (MXFP4, as published): about 14 GB
- Gemma 3 27B (8-bit): about 30 GB
- Mistral Small 3.2 24B (8-bit): about 27 GB
- Phi-4 14B (16-bit): about 31 GB
- FLUX.1 [dev] (16-bit, full precision): about 24 GB
- Stable Diffusion XL (Comfortable, with LoRAs and upscaling): about 12 GB
- Z-Image-Turbo (As released, the maker's figure): about 16 GB
- Qwen-Image-2.1 (8-bit, ComfyUI's standard files): about 16 GB
- Wan 2.2 5B (720p, the maker's own script): about 24 GB
Common questions
Questions, answered plainly.
How much does it cost to rent an L40S per hour?
On Vast.ai, as of 24 October 2026, 18:09 UTC, an available L40S typically rented for $0.80 an hour, and half of available cards cost $0.80–$0.80. RunPod's cheapest in stock was $0.79 an hour. Storage and data can cost extra.
How much memory does the L40S have?
48 GB per card. That's enough for a chat model of up to about 78B parameters at 4-bit, with room for a short conversation; at 16-bit, about 23B.
Is an L40S cheaper on Vast.ai or RunPod?
As of 24 October 2026, 18:09 UTC: a typical $0.80 an hour on Vast.ai, and $0.79 as RunPod's cheapest in stock. Each charges separately for storage and data, so compare the total for your job.
Other cards
Compare with other big GPUs.
Preparing an invite-only beta
Get an email when access opens.
No launch date is set. We'll email you once when KernelMachine opens. Nothing else.