AI Companies Embrace Efficient Models That Run on Fewer Chips