Smaller end of the spectrum here: an RTX 4060 with 8 GB, in a gaming PC. I don't run other people's models on it, I train my own small French one (97M parameters). 8 GB is enough for that if you use activation checkpointing and keep the batch small; the 4060 does about 19k tokens/s on a 97M model in bf16, so a 2B-token pretraining run takes a bit over a day. For anything bigger I use free Kaggle sessions (two T4s, 30 h a week), which are roughly 1.4x the 4060 when you run both cards. Inference of the result is the easy part: int8 weights, 98 MB, it runs on a laptop CPU at ~60 ms/token with no GPU at all.