A Tesla V100 is fairly affordable but older and won’t run newer quants like FP8 or attention kernel routines. But it will run a pile of models, even most new models are compatible or usable.