The memory on a GPU, separate from system RAM. AI models must fit in VRAM to run on a GPU. A 7B parameter model in 16-bit precision needs ~14GB of VRAM. Consumer GPUs have 8-24GB; datacenter GPUs (A100, H100) have 40-80GB. VRAM is almost always the bottleneck for local AI.
Why it matters
VRAM determines which models you can run. It's why quantization exists (to shrink models to fit), why MoE models are tricky (all experts must fit in VRAM), and why GPU prices scale so steeply with memory. "Will it fit in VRAM?" is the first question of self-hosting AI.