Dylan Patel: prediction

Transcript context

…Okay, so they will also increase because they’ll be using some quantity, but there’s not as much of a need as there is for HBM. Makes sense. One thing I didn’t appreciate until I was reading some of your newsletters is that the same constraints preventing logic scaling over the next few years are quite similar to what’s preventing us from producing more memory wafers. In fact, literally the same exact machine, this EUV tool, is needed for memory. So I guess the question someone could ask right now is, why can’t we just make more memory? The constraints, as I was mentioning earlier, are not necessarily EUV tools today or next year. They become that as we get to the latter part of the decade. Currently, the constraints are more that they physically just haven’t built fabs. Over the last three to four years, these vendors have not built new fabs because memory prices were really low. Their margins were low, and in fact, they were losing money in 2023 on memory. So they decided they weren’t building new fabs. The market slowly recovered over time but never really got amazing until last year. In 2024, we were banging on the drums that reasoning means long context, which means a large KV cache, which means you need a lot of memory demand. We’ve been talking about that for a year and a half, two years. People who understand AI went really long on memory then. So you’ve seen that dynamic, but now it has finally played out in pricing. It took so long for what was obvious: long context means the KV cache gets bigger, you need more memory. Half the cost of accelerators is memory. Of course they’re going to start going crazy on it. It took a year for that to actually reflect in memory prices. Once memory prices reflected that, it took another three to six months for the memory vendors to start building fabs. Those fabs take two years to build. So we won’t have really meaningful fabs to even put these tools in until late 2027 or 2028. Instead, you’ve seen some really crazy stuff to get capacity. Micron bought a fab from a company in Taiwan that makes lagging-edge chips. Hynix and Samsung are doing some pretty crazy things to try and expand capacity at their existing fabs, which also have large knock-on effects in the economy. So why can’t we build more capacity? There’s nowhere to put the tools. It’s not just EUV; there are other tools involved in DRAM and logic. In logic, for N3, about 28% of the cost of the final wafer is EUV. When you look at DRAM, it’s in the teens. It’s going up, but it’s a much smaller percentage of the cost. These other tools are also bottlenecks, although their supply chains are not as complex as ASML’s. You see Applied Materials, Lam Research, and all these other companies expanding capacity a lot as well. But you don’t have anywhere to put the tool, because the most complex buildings people make are fabs, and fabs take two years to build. I interviewed Elon recently, and his whole plan is that they’re going to build this TeraFab and they’re going to build the clean rooms. I won’t even ask you about the dirty rooms thing, but let’s say they build the clean rooms. I have a couple of questions. One, do you think this is the kind of thing that Elon Co. could build much faster than people conventionally build it? This is not about building the end tools. This is just about building the facility itself. How complicated is it to just build the clean room extremely fast? Is this something that Elon, with his “move fast” approach, could do much faster if that’s what we’re bottlenecked on this year or next year? Two, does that even matter if, in two years, your view is that we’re not bottlenecked on clean room space, but on the tooling?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.