Arvind Krishna on AI infrastructure spending:
“The infrastructure build-out is a bit ahead.
By my math, about 1 gigawatt of power costs roughly $60 billion to $80 billion in semiconductors to populate. So if people have committed to more than 100 gigawatts of AI data center build-out, that points to $6 trillion to $8 trillion in total investment.
If that requires a five to seven year payback, the industry would need an extra $1 trillion to $2 trillion a year in revenue. Even at high margins, that margin might be 20% to 30%. I don’t believe that much incremental revenue is there.
That’s why I think the build-out is a bit ahead.
I also believe many of the largest AI models will become commodities. Commodities can have a lot of value, but they usually have low switching costs. And if switching costs are low, margins may exist, but they won’t come with a massive moat.
So perhaps there won’t be half a dozen to a dozen companies that can build the largest models and survive. Maybe it’s two or three.
That raises the question of how much capital expenditure can realistically go into data centers. If the build-out were half of what we’re seeing today, I’d say it makes complete sense. But when it’s double that, some players may not be able to generate a strong return.”