> ran the same workflow over and over on some api > watched the token costs pile up > then i found out about local ai appliances > bought a 5090 > running mistral 7b at 200 tokens per second > after a year it's basically free https://t.co/tYohyPj1iA