Paul van Gerven

19 February

The plan of OpenAI CEO Sam Altman to raise 7 trillion dollars is a thinly veiled publicity stunt.

OpenAI frontman Sam Altman wants a lot of AI chips and he wants them to be cheap. He’s talking to financiers in the Middle East to secure 5-7 trillion dollars to create a semiconductor manufacturing industry just for AI, the Wall Street Journal reported last week. Altman intends to hand over the money he raises to TSMC to construct and operate chip plants.

Seven trillion dollars, that’s a lot of money. It’s several times bigger than the world’s biggest investment fund and 13 times last year’s semiconductor market. It’s roughly the GDP of the Netherlands, France and the UK combined. You could buy a couple of Big Seven companies or make a few manned trips to Mars.

If we assume a next-gen fab costs 40 billion dollars, seven trillion will build 175 of them. At the reported output of a still-to-be-completed TSMC fab in Arizona, these would spit out 52.5 million 300mm wafers per year. At a die size of 800 square millimeters, roughly 67 dies fit in a wafer, making for 3.5 billion AI chips per year. It’s going to be a lot less if you assume the 7 trillion would also pay for packaging, memory chips, real estate, operating costs of fabs and so on, but it will still be well over an order of magnitude more chips than the couple of million Nvidia sold last year.

It’s of course a ridiculous plan. Even if all that much computing power could be put to good use, who would build the manufacturing equipment or supply the specialty materials? How many times would ASML need to double? The amount of natural resources that would be required is mind-blowing, including the amount of energy to manufacture and run the chips. What about skilled workers? What about market cyclicality? What about technology upgrades – just raise 7 trillion again?

Bits&Chips event sponsor registration-early bird

It’s a ridiculous plan. You can’t assume that a graph that’s going up will continue to go up. But if we’re extrapolating trends, let’s extrapolate the performance improvement of GPUs. Nvidia claims this has been a million-fold over the past ten years. So, in ten years, a couple of million Nvidia chips will be able to do the work of the couple of billion chips that Altman says he needs. “I can do it for less than $1T,” microprocessor engineer Jim Keller tweeted.

It’s a ridiculous plan and Altman knows it. This isn’t a case of a software guy not used to being constrained by the laws of physics. Not so long ago, he subscribed to the concept of controlled AI growth to make sure society can safely absorb its expanding possibilities. The ‘natural’ increase of chip performance and semiconductor manufacturing capacity is a great way to curb growth, but now Altman is calling for a giant leap all of a sudden.

Apparently, Altman’s concern about AI safety last year amounted to nothing more than a publicity stunt. Now that the buzz is fading, he moves on to some other shenanigans to get the hype train going.