USED A100 40GB

Indicative pricing for used A100 40GB SXM4 is $4,000 to $6,000 per unit as of Q3 2026, around half of the $11,000 original MSRP. The 40GB variant trades at a steeper discount than the 80GB variant because memory is the binding constraint for many modern inference and fine-tuning workloads.

Inference for models above roughly 30B parameters typically overflows 40GB once activation memory and KV-cache are accounted for. Fine-tuning with full optimizer state for 13B-plus models also requires more than 40GB without aggressive sharding. For workloads that fit comfortably under those thresholds, A100 40GB delivers identical compute throughput to the 80GB variant.

Inference for 7B to 30B parameter models, parameter-efficient fine-tuning with LoRA or QLoRA, classical computer vision and tabular ML training, dev and test environments mirroring Ampere production, and academic research workloads. Stay under the 40GB ceiling and the price-per-throughput economics are unmatched.

Used L40S at $5,500 to $7,500 delivers 733 FP16 sparse TFLOPS plus FP8 acceleration with 48GB GDDR6X. Used A100 40GB at $4,000 to $6,000 delivers 312 FP16 sparse TFLOPS without FP8. For FP8-quantized LLM serving, L40S wins on tokens-per-second-per-dollar. For Ampere-style production workloads still on FP16/BF16, A100 40GB is the cheaper option.