Microsoft's Maia 300 swaps raw compute for a 2.5x memory bandwidth gain as it moves to make in-house silicon its main AI engine while claiming a 40% efficiency bump — could going the Google route be the pin that pops the Nvidia AI bubble?