There’s one company trying to etch models into the silicon, they got a working prototype in a 8b model and it does 17k t/s. They suggest at data center scales with bigger models it’d pay for itself in a year in saved costs in electricity, cooling, space required etc.

You can apply lora’s to it, but it will fall behind as it ages, im not quite sure how long a static model would last to make up the money after that 1st year, but for something like creative writing where it doesnt need current data (e.g programming but the languages are always changing and evolving is a problem) it might work?