12 providers · per GPU-hour
NVIDIA L40S cloud pricing
13 offers from 12 providers, 13 with a published on-demand price. Per GPU-hour, US$. Prices checked 28 Sept 2026.
- Lowest on-demand
- $1.20/GPU-hr
- Lowest spot
- $0.99/GPU-hr
- Memory
- 48 GB GDDR6 ECC
Spec sheet: L40S ↗
| # | Provider | Offer | On-demand | Spot | Committed | As published (original currency) · source | Actions |
|---|---|---|---|---|---|---|---|
| 1 | E2E Networks | L40S x1 1× L40S · 60 vCPU · 220 GB RAM | $1.20/GPU-hr | — | — | ||
| 2 | UpCloud | L40S x1 1× L40S · 8 vCPU · 64 GB RAM | $1.24/GPU-hr | — | — | ||
| 3 | Cyfuture | L40S x1 1× L40S · 16 vCPU · 256 GB RAM | $1.31/GPU-hr | — | — | ||
| 4 | OVHcloud | L40S x1 1× L40S · 15 vCPU · 90 GB RAM · 400 GB disk | $1.35/GPU-hr | — | — | ||
| 5 | Verda (formerly DataCrunch) | L40S x1 1× L40S | $1.41/GPU-hr | — | — | ||
| 6 | Yotta Infrastructure / Shakti Cloud | VM – 1 x L40S On Demand 1× L40S · 28 vCPU · 240 GB RAM · 100 GB disk | $1.44/GPU-hr | — | — | INR 137/hr | |
| 7 | Yotta Infrastructure / Shakti Cloud | VM – 2 x L40S On Demand 2× L40S · 56 vCPU · 480 GB RAM · 100 GB disk | $1.45/GPU-hr | — | — | INR 275/hr | |
| 8 | AceCloud | L40S x1 1× L40S · 16 vCPU · 64 GB RAM | $1.50/GPU-hr | — | — | ||
| 9 | Crusoe | L40S x1 1× L40S | $1.50/GPU-hr | — | — | ||
| 10 | Nebius | L40S x1 1× L40S | $1.55/GPU-hr | — | — | ||
| 11 | Vultr | L40S x1 1× L40S | $1.67/GPU-hr | — | — | ||
| 12 | CoreWeave | NVIDIA L40S (8x) 8× L40S · 128 vCPU · 1024 GB RAM | $2.25/GPU-hr | $0.99/GPU-hr | — | ||
| 13 | Oracle Cloud Infrastructure | L40S x1 1× L40S | $3.50/GPU-hr | — | — |
What is the L40S used for?
The L40S is a data-centre GPU positioned by its maker for both training and inference, built on the Ada Lovelace architecture and released in 2023. Multi-GPU servers link cards over No NVLink, which matters when a job is split across several GPUs.
How large a model fits on one L40S?
With 48 GB of memory, one card holds the weights of a model up to about 24 billion parameters at 16-bit precision, or about 48 billion at 8-bit. That is weights only: inference also needs room for the KV cache, and training needs several times more for gradients and optimiser state, so larger jobs are split across multiple GPUs.
| Spec (as published) | L40S |
|---|---|
| Architecture | Ada Lovelace |
| Form factor | PCIe dual-slot |
| Memory | 48 GB GDDR6 ECC |
| Memory bandwidth | 864 GB/s |
| Interconnect | No NVLink |
| Max power (TDP) | 350 W |
| Precisions listed | FP32, FP8 |
| Partitioning | No MIG; vGPU supported |
| Released | 2023 |
| Source | maker spec sheet ↗ |
Need L40S capacity for a longer run?
Committed and cluster pricing is often negotiated. Tell us the GPU count, duration and region.