GPU class · 48 GB
L40S
What does L40S cost and what fits on it?
L40S has 48 GB and costs $0.78 per hour to train on, $0.90 to serve. 36 models in the catalogue fit half-precision LoRA on it and 5 more fit in four-bit.
A newer datacentre 48GB part, and the class we reach for when an endpoint needs to hold many adapters against one base model at once. It is priced above the A40 and the RTX 6000 Ada, so for a single training run it is rarely the value pick; for a busy multi-tenant-style serving deployment inside one account, it usually is.
When to pick this class
This is the class we reach for when an endpoint needs to hold many adapters against one base model at once. For a single training run it is rarely the value pick — the A40 and the RTX 6000 Ada are both cheaper at the same capacity — but for a busy multi-variant serving deployment the throughput and memory bandwidth earn the difference.
If you are choosing it for training, check the cost calculator against the two cheaper 48GB classes first. If you are choosing it for serving several adapters under real traffic, it is usually correct and the alternatives are false economy.
Rates
| VRAM | 48 GB |
|---|---|
| Training | $0.78 / GPU-hourMetered per GPU-second, one-minute floor |
| Serving | $0.90 / GPU-hourWhile a replica is resident |
| Warm for a day | $21.6024 hours resident, regardless of traffic |
| Warm for 30 days | $648Set the endpoint to scale to zero if nobody is waiting |
| Models — LoRA | 36 |
| Models — QLoRA only | 5 |
Largest models that fit for LoRA
Half precision, frozen base, within 48 GB.
Mistral Small 3 24B
24BNeeds 48 GB · Apache 2.0
LFM2 24B-A2B
24BNeeds 48 GB · LFM Open
gpt-oss 20B
20.9BNeeds 40 GB · Apache 2.0
Qwen3 14B
14BNeeds 28 GB · Apache 2.0
Phi-4 14B
14BNeeds 28 GB · MIT
Mistral Nemo 12B
12BNeeds 24 GB · Apache 2.0
Gemma 3 12B
12BNeeds 24 GB · Gemma Terms of Use
Falcon 3 10B
10BNeeds 22 GB · TII Falcon LLM
Models that need four-bit training here
These exceed 48 GB in half precision but fit quantised. Quantised training is slower per step, so compare the total run cost against the next class up rather than the hourly rate alone.
Compared with its neighbours
The classes immediately either side of L40S on capacity and rate.
| Class | VRAM | Training | Serving | Versus this one |
|---|---|---|---|---|
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | same capacity, $0.17/hr cheaper |
| A6000 | 48 GB | $0.65 | $0.75 | same capacity, $0.13/hr cheaper |
| A100 80 GB | 80 GB | $1.68 | $1.94 | 32 GB more, $0.90/hr dearer |
| H100 80 GB | 80 GB | $2.59 | $2.99 | 32 GB more, $1.81/hr dearer |
Frequently asked questions
How much does L40S cost per hour?
$0.78 per GPU-hour for training and $0.90 for serving, metered per GPU-second above a one-minute floor. A continuously warm endpoint on this class is about $21.60 a day, or $648 over thirty days.
What models fit on L40S?
36 models in the catalogue fit half-precision LoRA training within 48 GB, and 5 more fit in four-bit. 37 can be served from this class before accounting for the attention cache.
Is L40S the cheapest option?
Cheapest per hour is not the same as cheapest per run. Total cost is the rate multiplied by wall time, so a faster class that finishes sooner often costs less — particularly when the cheaper alternative would require quantised training, which is slower per step.
Last verified 6 August 2026.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.