You want to look into frigate, and you don't need much of a graphics card. You could use a old gaming PC with a gtx 1060 from 10 years ago.

There's cameras with tensor cores in them too. I only played around with the pi ai camera but frigate uses the same yolo models. so you can have your camera process the image detection or the server's GPU.

It doesn't take as much as you think. Its when you want to generate content that its important you need lots of vram or cram the entire model into context.

I am the biggest ai hater I know but local models, and old school ai technologies before LLMs are pretty cool.