NVIDIA L40S

The L40S starts at $0.47 per GPU-hour, with a median of $1.27 across 43 available configurations. Actual pricing depends on the provider, region, and whether you rent on-demand, spot or reserved.

12 providers currently list the L40S: Cyfuture AI, Sesterce Cloud, AceCloud, Omega Gradient, Scaleway, Runcrate and 6 more. Compare their hourly pricing, regions and machine specs in the table above.

The L40S comes with 48 GB of VRAM, which determines the largest models and batch sizes it can hold in memory for training and inference.

With 48 GB of VRAM, a single L40S can serve models such as Qwen3.6 35B-A3B, Qwen3 32B, Gemma 4 31B, GLM-4.7 Flash. The table above lists the estimated VRAM for each model at 16-bit, 8-bit and 4-bit precision, including the multi-GPU configurations needed for larger models.

The largest models we track that fit across several L40S cards are DeepSeek V4 Flash on 4 GPUs at 4-bit, Solar Open2 250B on 8 GPUs at 8-bit, Qwen3 235B-A22B on 4 GPUs at 4-bit. A multi-GPU configuration means you rent and pay for every card for the whole run, and splitting a model across cards costs throughput as well: the GPUs exchange activations on every token, so performance depends on how fast they are wired together. We do not track interconnect, so check the provider's node specification. The table above gives the smallest configuration for each model at 16-bit, 8-bit and 4-bit precision.

The lowest on-demand price we currently track for the L40S is $0.47 per GPU-hour at Vast.ai. Because providers update pricing and availability frequently, check the live table above before you launch.