What is the cheapest 24GB GPU to rent?

Twenty-four gigabytes is enough for a lot of real work: single-user LLM inference on small and mid-size models, Stable Diffusion and SDXL generation, LoRA training, and light fine-tuning. It is not enough to hold a 70B model at usable precision or to fully fine-tune large models, which is where 48GB and 80GB cards take over.

The cards in this tier differ in ways beyond price. The RTX 4090 is the fastest for most single-GPU tasks, the RTX 3090 trades a little speed for an even lower rate, and the L4 is a low-power datacenter card that some providers offer on more stable tiers. Pick on availability and reliability once the price is close.

The table below lists the 24GB-and-up cards we track, sorted by price, so you can see where each currently sits across providers. It is rebuilt from the daily snapshot rather than a fixed list price.