The big picture is to think about computers not as supercomputers but as AI factories producing tokens/numbers.

You have to use the factory for three fundamental things. The first is to train your next frontier model to build the best AI and get to market first.

You want to train as fast as you can. Comparing Rubin to Blackwell, Rubin is a 4x leap.

That 4x jump means you could do something in four months or do it in one month - a very big difference. You could either do something with four times as many Blackwells or one fourth as many, saving a ton of money.