5 year old GPUs are still at crazy prices.

Plenty of businesses want in local inference and are priced out.

You can't do quality inference on such a toy. It's a myth.