USED L40S
NVIDIA L40S 48GB
Used L40S units are popular for inference and video or image generation. PCIe compatibility, lower power draw, and FP8 acceleration make them straightforward to deploy in existing infrastructure for FP8-quantized LLM serving and multi-modal workloads.
USED L40S — Indicative Range (Q3 2026)
$5,500 – $7,500
Steady demand for inference and visual workloads. Pricing reflects ongoing utility.
Based on aggregated supplier quotes and broker market data. Final pricing depends on quantity, region, condition, and warranty terms. Compute Exchange does not publish active counterparty listings.
TECHNICAL SPECIFICATIONS
USEDVRAM
48 GB GDDR6X
MEMORY BANDWIDTH
864 GB/s
FP 16 TENSOR
733 TFLOPS (sparse)
FP 8 TENSOR
1,466 TFLOPS (sparse)
TDP
350W
FORM FACTOR
PCIe Gen4
INTERCONNECT
PCIe 4.0 x16
ARCHITECTURE
Ada Lovelace
Frequently Asked Questions
USED L40S — KEY QUESTIONS
What is the price range for used L40S?+
Indicative pricing for used L40S is $5,500 to $7,500 per unit as of Q3 2026, around 30 to 45 percent below the $10,000 original MSRP. The wider range reflects condition variance and bundling of original heatsinks and brackets across suppliers.
Why is used L40S pricing more stable than Hopper?+
L40S occupies a unique inference niche — Ada Lovelace FP8 acceleration with 48GB GDDR6X at lower power and PCIe form factor — that does not directly compete with Hopper or Blackwell. Demand for FP8 inference at modest scale keeps used L40S pricing relatively firm even as Hopper-generation pricing softens broadly.
What workloads run well on used L40S?+
FP8-quantized inference of LLMs up to 70B parameters, video and image generation pipelines, multi-modal model serving, hybrid graphics-plus-compute workloads such as digital twins, and small-scale fine-tuning that fits in 48GB. The L40S is not optimized for distributed training or HBM-bandwidth-bound workloads.
Used L40S versus used A100 80GB for inference?+
Used L40S at $5,500 to $7,500 delivers 1,466 FP8 sparse TFLOPS that A100 lacks. Used A100 80GB at $7,000 to $9,500 delivers 312 FP16 sparse TFLOPS and 80GB HBM2e. For FP8-quantized recent LLMs, L40S typically wins on tokens-per-second-per-dollar. For workloads that need more than 48GB of memory, A100 80GB is the only option.
Ready to Source or List?
LIVE QUOTE FOR
USED L40S
Compute Exchange returns indicative pricing within 24 hours, anchored to your specific quantity, region, and condition. We do not publish active counterparty listings.
Request a Quote