Is the AI Boom Changing the Semiconductor Cycle?

During their recent episode, Taylor, Carlisle, and David Samra discussed Is the AI Boom Changing the Semiconductor Cycle? Here’s an excerpt from the episode:

[David] First of all, again, not to be pithy, but we’re clearly not bumping along the bottom of the cycle. So let’s just start there. 85% gross margins.

It’s a pretty good earning stream coming out of these businesses. Now, how far can it go? The speculation is it’s going to go again next year, because the hyperscalers are ramping up their capital spending yet again.

Now, you listened to Buffett talk to CNBC, and he talked about the fact that, and this is a simple observation, that there are participants in the AI game that are spending money because they have to. They have to compete. So you’ve got five, six of these LLMs, all who need capacity in order to try and win this game where they’re not really making any money.

And some have competitive advantages and some don’t. And the persistence of that capital spending is what is driving today’s level of demand for DRAM. And just like any other commodity-based cycle, there are two things that happen when the price goes above the marginal cost of production.

The users of that commodity are doing everything that they can to reduce that usage. And the manufacturers of that commodity are doing everything they can to increase the production, because, well, it’s 85% gross margins. Why wouldn’t you want to sell more of that?

And so the supply-demand balance is not going to persist. It’s just not. And you don’t bring capacity on in small increments.

You bring it on in huge increments, and you have Samsung, Hynix, and Micron all spending capital in ways that we have never seen before. Some of it’s subsidised. A fair amount of it’s subsidised.

And then you’ve got China, who wants to be self-reliant, spending an enormous amount of money building out DRAM capacity. It’s not at the edge, but it’s still offsetting capacity that others would have been able to sell. So the risk-reward is definitely not where it was when you were paying book value for these things.

I could try to speculate. I’ve looked at supply-demand models. They’re all over the map.

We tried to do it ourselves. It’s just hard to tell. It’s hard to tell.

And that says a lot. If you say it’s hard to tell, then are you really getting a good risk-reward? Isn’t that enough to make you wonder?