by liberal japonicus
Another day, another post about AI and another thing with Ezra Klein. Huang is on a tour, I see him talking with all sorts of people, hitting on the same points, which is there is nothing to worry about and nothing to see here.
Huang is a pretty interesting character, his lifestory is detailed in his wikipedia entry. He’s Taiwanese, so I find it interesting how he’s viewed in China vs Taiwan, and how Nvidia is a fabless company, so it just designs the chips, which are produced in Taiwan and South Korea. Huang’s tied to TSMC, which is the Taiwanese powerhouse, and I’m sure he’s well aware that if something happens to TSMC, Nvidia is really just some designs. I feel like this situation contributes to Nvidia’s loans to AI companies and it is easier to think of things like this if you don’t think that you are tied to a place and a society.
Anyway, another thread to talk about the same thing. Have at it.
Interesting that, in addition to making their chips in Taiwan (like nearly everybody outside the PRC), Nvidia is also making chips in South Korea. That suggests to me that the Koreans are positioning themselves to pick up the chip-making baton if/when China decides to retake the island. Well, somebody ought to be.
Japan is also capable of building/restarting fabs as profits warrant, some chip manufacturing moved offshore to SK and TW simply because it was cheaper. I suspect the next big fab location beyond PRC, TW, SK, will be Vietnam.
@Marc, did you leave the US out on purpose? TSMC has committed more then $100B to expanding in Arizona, and has promised to be using their improved 3nm process there in 2027, and their newest 2nm process in 2028. Intel has an 18A fab in Arizona and is building another one in Ohio. Micron is building a new leading edge memory fab in Idaho. SK Hynix is building in Indiana. Samsung’s Tyler, TX fab is sampling 2nm parts and will be in full production next year. By next year, the US will be the only country where all the companies with 3nm or better processes will be producing.
Work is about to begin on Musk’s 100-million-square-foot Terafab. Completion is likely years away, even at Musk’s breakneck speeds.
I follow the podcast Asianometry, and a point he often makes is how tetchy the supply chains are for the cutting edge fabs. TSMC (which now has two fabs in my hometown) has found that just changing just one single element, like changing the supplier of sulfuric acid or neon gas, can lead to entire batches being scrapped. This is why TSMC adopts a Copy-EXACT Strategy, where they _clone_ everything, from equipment layouts to exact same suppliers of things like ultra pure chemicals and raw materials. This has been a big problem at the US TSMC fab, where the engineers are trained to solve problems creatively and Copy-EXACTLY! forces them to blindly adhere to strict recipes without altering steps. Even the Japanese have difficulties matching the TSMC commitment to exact imitation, so the TSMC plants here are all imported Taiwanese engineers. That’s why I’m a little doubtful with the startup times for these cutting edge fabs.
Terafab may be owned by SpaceX/Tesla, but the fab proper is pure Intel. Until it’s built, both SpaceX and Tesla are using TSMC and Samsung.
…how tetchy the supply chains are for the cutting edge fabs.
Yeah, TSMC goes back and forth on their AZ commitments. On the one hand they’ve said AZ will always be a couple of generations (or more) behind Taiwan because their suppliers won’t set up in the US. On the other hand, the Biden administration and some of the huge US customers have said that’s not good enough. My speculation, from rumors, is that TSMC will owe Apple a ton of money if they don’t have full 2nm production running in AZ in 2028.
Samsung’s Tyler, TX fab…
Nitpick: There’s a Tyler, TX, but the fab is in Taylor, TX.
As a US-trained semi-engineer, the dependence on Copy-EXACT feels like we don’t actually understand what’s going on here, and are dependent on happy accident that things work. It gets results, but people should have been aware of supply-chain fragility even long before COVID, and at some point we ought to push back on financialization enough to get more reliability.
The whole “Just in time” management plan should have resulted in everyone at the C-level and above being summarily fired.
Damn the supply-chain lack of kerosene!