GPU class · 48 GB
A40
What does A40 cost and what fits on it?
A40 has 48 GB and costs $0.42 per hour to train on, $0.48 to serve. 36 models in the catalogue fit half-precision LoRA on it and 5 more fit in four-bit.
Forty-eight gigabytes at close to 4090 money. That extra headroom is what makes preference tuning practical, because DPO holds a frozen reference model in memory alongside the one being trained and a 24GB card runs out well before a 14B pair fits.
When to pick this class
Forty-eight gigabytes at close to 4090 money is the whole argument, and the specific thing it unlocks is preference tuning. DPO holds a frozen reference model in memory alongside the one being trained, and a 24GB card runs out well before a 14B pair fits — so this is the cheapest class where preference tuning at mid-large scale is practical.
Against the RTX 6000 Ada at the same capacity it is slower and considerably cheaper. For a long run the faster card can win on total cost; for a short one this is the value pick and the difference is not close.
Rates
| VRAM | 48 GB |
|---|---|
| Training | $0.42 / GPU-hourMetered per GPU-second, one-minute floor |
| Serving | $0.48 / GPU-hourWhile a replica is resident |
| Warm for a day | $11.5224 hours resident, regardless of traffic |
| Warm for 30 days | $345.60Set 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 A40 on capacity and rate.
| Class | VRAM | Training | Serving | Versus this one |
|---|---|---|---|---|
| RTX 4090 | 24 GB | $0.38 | $0.43 | 24 GB less, $0.04/hr cheaper |
| A100 40 GB | 40 GB | $1.17 | $1.35 | 8 GB less, $0.75/hr dearer |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | same capacity, $0.19/hr dearer |
| A6000 | 48 GB | $0.65 | $0.75 | same capacity, $0.23/hr dearer |
Frequently asked questions
How much does A40 cost per hour?
$0.42 per GPU-hour for training and $0.48 for serving, metered per GPU-second above a one-minute floor. A continuously warm endpoint on this class is about $11.52 a day, or $345.60 over thirty days.
What models fit on A40?
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 A40 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.