Dylan Patel: belief

Transcript context

…A year. A year. The problem with any of these numbers, or the challenge to these numbers, is actually not the power or the data center. We can dive into that, but it’s manufacturing the chips. Take a gigawatt of Nvidia’s Rubin chips. Rubin is announced at GTC, I believe the week this podcast goes live. To make a gigawatt worth of data center capacity of Nvidia’s latest chip that they’re releasing towards the end of this year, you need a few different wafer technologies. You need about 55,000 wafers of 3 nm. You need about 6,000 wafers of 5 nm, and then you need about 170,000 wafers of DRAM memory. Across these three different buckets, each requires different amounts of EUV. When you manufacture a wafer, there are thousands and thousands of process steps where you’re depositing material and removing them. But the key critical step—which at least in advanced logic is 30% of the cost of the chip—is something that doesn’t actually put anything on the wafer. You take the wafer, you deposit photoresist, which is a chemical that chemically changes when you expose it to light. Then you stick it into the EUV tool, which shines light at it in a certain way. It patterns it. There’s what’s called a mask, which is effectively a stencil for the design. When you look at a leading-edge 3 nm wafer, it has 70 or so masks, 70 or so layers of lithography, but 20 of them are the most advanced EUV. If you need 55,000 wafers for a gigawatt, and you do 20 EUV passes per wafer, you can do the math. That’s 1.1 million passes of EUV for a single gigawatt. It’s pretty simple. Once you add the rest of the stuff, it ends up being 2 million, across 5 nm and all the memory. You’re at roughly 2 million EUV passes for a single gigawatt. These tools are very complicated. When you think about what it’s doing across a wafer, it’s taking the wafer and scanning and stepping across. It does this dozens of times across the whole wafer. When you’re talking about how many EUV passes, that’s the entire wafer being exposed at a certain rate. An EUV tool can do roughly 75 wafers per hour, and the tool is up roughly 90% of the time. In the end, you need about three and a half EUV tools to do the 2 million EUV wafer passes for the gigawatt. So three and a half EUV tools satisfies a gigawatt. It’s funny to think about the numbers. What does a gigawatt cost? It costs roughly $50 billion. Whereas what do three and a half EUV tools cost? That’s $1.2 billion. It’s actually quite a lower number, which is interesting to think about. Fifty gigawatts of economic CapEx in the data center, and what gets built on top of that in terms of tokens is even larger. It might be $100 billion worth of AI value into the supply chain, held up by this $1.2 billion worth of tooling that simply cannot expand its supply chain quickly. You wrote an article recently saying over the last three years, TSMC has done $100 billion of CapEx. So it’s $30/$30/$40 billion. A small fraction of that is being used by Nvidia for the 3 nm, or previously 4 nm, that it’s using for its chips. What were its earnings last quarter? It was $40 billion. So $40 billion times four is $160 billion. Nvidia alone is turning some small fraction of $100 billion in CapEx, which is going to be depreciated over many years and not just this one year, into $160 billion in a single year. That gets even more intense when you go down the supply chain to ASML, which is taking a billion dollars’ worth of machines to produce a gigawatt. Of course, those machines last for more than a year so it’s doing more than that. Now I want to understand, how many such machines will there be by 2030, if you include not just the ones that are sold that year, but have been compiling over the previous years? What does that imply? Sam Altman says he wants to do a gigawatt a week in 2030. When you add up those numbers, is it compatible with that?…

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