You've confused engineering compromise for malice, and reversed the purpose. For the model capabilities and inference power draw, what alternative do you see to a (at least mostly) fixed hardware model?