Cheapest GPU to train a LoRA (live prices)

Live list prices for single-GPU rentals, fetched 2026-10-20 08:06 UTC from RunPod (public price API) and Vast.ai (cheapest live marketplace offer). TensorDock: unknown (no public price API).

Cheapest 24GB+ card right now: RTX 3090 at $0.12/hr.

GPU	VRAM	RunPod community	RunPod secure	Vast.ai cheapest
RTX 3090	24 GB	$0.22	$0.50	$0.12
RTX 4090	24 GB	$0.34	$0.89	$0.37
RTX 5090	32 GB	$0.69	$1.19	$0.41
L40S	48 GB	$0.79	$1.09	$0.60
A40	48 GB	$0.35	$0.59	—
A100 80GB	80 GB	$1.39	$1.79	$0.66
H100 80GB	80 GB	$2.69	$3.99	—

Which card for which model?

FLUX.2 klein: 12-16 GB is enough (guide).
FLUX.2 dev: 80 GB recommended, a 24 GB config exists in ai-toolkit (guide).
MOSS Image: ~9 GB with ConvRot INT8 or ~13 GB in bf16 per kohya sd-scripts' measurements, 12 GB works with block swap (guide).
Fibo 2.0: ~8.7 GB at 512px with an NF4 build in AcademiaSD LoRAlab; the bf16 transformer alone is 16.5 GB (guide).
FLUX.3 Klein: SimpleTuner ships 24 GB recipes (guide).
Krea 2 / Z-Image: no exact numbers yet; musubi-tuner recommends 12 GB+ for image training (Krea 2, Z-Image).

Vast.ai is a marketplace, so host reliability varies between hosts.

Weekly newsletter: new models, trainer releases, cheapest GPUs. Subscribe

Data updated 2026-10-20. RunPod and Vast.ai links are referral links. unknown = no data yet. · Raw data (JSON) · llms.txt