L40S from $0.15 per GPU-hour
Ada Lovelace in a data-center card: 48 GB, strong FP8 throughput and no NVLink. It is the value pick for inference servers and for generative image work, where memory bandwidth per dollar matters more than multi-GPU scaling.
Billed per started minute. 384 GB of GPU memory on a 8× node. Persistent storage $0.08 per GB-month. No egress fees.
- GPU memory
- 48GBGDDR6
- Bandwidth
- 864GB/sSets token throughput on inference
- Node sizes
- 1× · 2× · 4× · 8×Same price per GPU at every size
- Reclaim rate
- <5%Spot instances reclaimed over 30 days
What you get with each L40S node.
| Node | GPU memory | vCPU | RAM | Local NVMe | Spot | On-demand | Action |
|---|---|---|---|---|---|---|---|
| 1× L40S | 48 GB | 12 | 96 GB | 1 TB | $0.15/h | $0.25/h | Configure |
| 2× L40S | 96 GB | 24 | 192 GB | 2 TB | $0.30/h | $0.50/h | Configure |
| 4× L40S | 192 GB | 48 | 384 GB | 4 TB | $0.60/h | $1.00/h | Configure |
| 8× L40S | 384 GB | 96 | 768 GB | 8 TB | $1.20/h | $2.00/h | Configure |
GPU-to-GPU link: PCIe Gen4 x16 64 GB/s. Local NVMe is scratch space wiped when the instance ends; keep anything you need on a persistent disk.
96 L40S GPUs free right now.
Straight from the capacity pool the console books against. Prices are identical in every region — pick the one closest to your data.
A region with zero free GPUs still accepts reserved capacity requests — we hold hardware for a term rather than sell what is already taken.
What fits in 48 GB — and what it costs per hour.
Weights need about 2 GB per billion parameters at FP16 / BF16, 1 GB per billion parameters at FP8, 0.5 GB per billion parameters at INT4. We keep 20% of the memory free for activations, the KV cache and the CUDA context, then take the smallest node that still fits.
| Open model | Parameters | FP16 / BF16node · spot price | FP8node · spot price | INT4node · spot price |
|---|---|---|---|---|
| Mistral 7BAssistants, RAG, classification | 7.2B | 1× $0.15/h | 1× $0.15/h | 1× $0.15/h |
| Llama 3.1 8BAssistants, agents, fine-tuning | 8B | 1× $0.15/h | 1× $0.15/h | 1× $0.15/h |
| Gemma 2 27BHigher-quality assistants | 27B | 2× $0.30/h | 1× $0.15/h | 1× $0.15/h |
| Qwen2.5 32BCode and reasoning | 32B | 2× $0.30/h | 1× $0.15/h | 1× $0.15/h |
| Mixtral 8x7BMixture of experts, high throughput | 46.7B | 4× $0.60/h | 2× $0.30/h | 1× $0.15/h |
| Llama 3.3 70BThe common production baseline | 70B | 4× $0.60/h | 2× $0.30/h | 1× $0.15/h |
| Qwen2.5 72BMultilingual, long context | 72B | 4× $0.60/h | 2× $0.30/h | 1× $0.15/h |
| Mixtral 8x22BMixture of experts, large capacity | 141B | 8× $1.20/h | 4× $0.60/h | 2× $0.30/h |
| Llama 3.1 405BLargest widely used open model | 405B | over 8× | over 8× | 8× $1.20/h |
Real usage depends on context length, batch size and the serving engine: a long context can add tens of gigabytes of KV cache. Treat the table as the floor, not the ceiling — and if a job is close to the limit, take the next node size or quantise one step further.
What a dollar buys on an L40S.
Ranked 11 of 21 data-center models in this catalogue on cost per gigabyte of GPU memory.
GPU memory you get for $1.00 of spot time, per hour.
Bandwidth decides token throughput far more often than raw FLOPS.
Spot price divided by published FP16 tensor throughput.
One GPU on spot, billed per minute, storage excluded.
Against the rest of the market
64 providers · September 2026- Median on-demand elsewhere
- $1.25Our on-demand is 80% below it
- Cheapest on-demand seen
- $0.35Our on-demand $0.25 is 29% below it
- Cheapest spot seen
- $0.28Our spot $0.15 is 46% below it
- 24 hours on one GPU
- $3.60 on spot$6.00 on-demand · $8.40 at the cheapest rate found elsewhere
Index built from published prices for the same GPU across 137 providers (public price index (getdeploying.com)), September 2026. How the guarantee is enforced.
What people run on an L40S.
Fine-tuning with LoRA and QLoRA
Comfortable for parameter-efficient training on 13B–34B models without sharding.
Inference servers
Holds 30B-class models quantised, or several smaller models side by side.
Anything that can checkpoint
On the spot tier this model is 40% below its own on-demand price, with a 2-minute notice before a reclaim.
NVIDIA L40S, as published by NVIDIA.
Taken from the vendor datasheet. Where a figure is not published, the row is absent rather than estimated.
- Vendor
- NVIDIA
- Architecture
- Ada Lovelace (2023)
- Segment
- Datacenter
- Memory
- 48 GB GDDR6
- Memory bandwidth
- 864 GB/s
- FP16 / BF16 tensor
- 362 TFLOPS
- FP8 tensor
- 733 TFLOPS
- FP32
- 92 TFLOPS
- CUDA cores
- 18,176
- Board power
- 350 W
- Form factor
- PCIe dual-slot
- Interconnect
- PCIe Gen4 x16 64 GB/s
- Multi-instance (MIG)
- Not available
- 8-bit float (FP8)
- Supported by the architecture
Source: www.nvidia.com · verified against the vendor document
Software that runs on it
8 templatesPre-built environments, pulled on the node before you land on it. Or bring any OCI image from a public or private registry.
On the spot tier
<5% reclaimed · 30 d- 2 minutes of notice on the metadata endpoint, a webhook and the console
- The instance is stopped, not deleted — disk and IP stay attached
- Auto-relaunch on the next free L40S, or switch the same disk to on-demand
- You save $2.40 a day per GPU against on-demand
If the L40S is not the right fit.
Same memory or more, 40% cheaper on spot
A100 PCIe 80GB 80 GB HBM2e · Ampere $0.09/h spotNext step up in memory: 80 GB instead of 48 GB
L40 48 GB GDDR6 · Ada Lovelace $0.09/h spotSame Ada Lovelace generation
L4 24 GB GDDR6 · Ada Lovelace $0.05/h spotSame Ada Lovelace generation
Or put two of them side by side: L40S vs RTX A6000 · L40S vs A100 PCIe · L40S vs L40
L40S — the questions that come up
How much does an L40S cost per hour?
Spot is $0.15 per GPU-hour and on-demand is $0.25, both billed per minute — every started minute costs the hourly price divided by 60, so $0.0025 on spot. Reserved capacity is $0.15 per GPU-hour on a monthly commitment. Storage and egress are not included in that rate: persistent disks are $0.08 per GB-month and there are no egress fees.
How many GPUs can I put in one instance?
Node sizes are 1×, 2×, 4×, 8× — up to 8 L40S GPUs in a single virtual machine, with CPU, RAM and local NVMe scaled with the GPU count. The price per GPU is identical at every node size.
Can it run a 70B model?
Yes. At FP8, a 70B model needs about 70 GB for the weights, so it fits on 2× L40S ($0.30 per hour on spot) with roughly 20% of the memory left for activations and the KV cache.
Which regions have it?
Available now in US East (Virginia), US West (Oregon), EU Central (Frankfurt). Prices are identical in every region; pick the one closest to your data.
How often is a spot L40S reclaimed?
Over the trailing 30 days, <5% of spot instances on this model were reclaimed. You get a 2-minute notice on the metadata endpoint, through a webhook and in the console; the instance is then stopped, never deleted, and the persistent disk stays attached.
Anything else about this model? Ask an engineer — the same people run the nodes.
An L40S for $0.15 an hour, running in under 60 seconds.
Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription. 96 L40S GPUs available right now.