The HKMG barrier you mentioned is interesting because it's more about process mastery than hard equipment limits. That probably means CXMT could catch up faster there than with the EUV stuff, but it's still gonna slow them down in the near term. It's kinda like how they've worked around DUV limits before but this time the workaround needs way more refinement to get yields up.
Great note. We have a lot of tourists in semiconductor investing today and when the cycle turns a lot of folks are going to get hurt. If anyone reading this has capital at work in semiconductor and really wants to learn how the industry works, this is a superb place to start learning. Kudos. Alex.
My understanding is that right now, hybrid bonding is too expensive and overkill for HBM. I think D2W bonding accuracy still isn’t good enough for HBM.
First, great overview! Second, I'd like to get your take on the capacity of CXMT for HBM2 and HBM3. My understanding is that access to any kind of HBM4 is bottlenecking the AI accelerators that are Made in China. Especially with the US embargos in place, while going for HBM3e and 4 makes sense for CXMT, if CXMT could produce HBM3 in large-enough quantities, it would already increase the capabilities of China's domestic AI accelerators considerably.
Lastly, I doubt that NAMICS would break their exclusive arrangement with SK Hynix, while an overt breaking of IP laws and international agreements by making a knock-off would severely affect China's export-oriented industry.
Thanks! Producing large amounts of HBM3 would definitely be an increase in capabilities for China, but I don't think it's as relevant if it's not keeping pace with Western capabilities. I think CXMT is reasonably trying to leapfrog where they can and catch up with the West rather than being permanently a few generations behind. That is definitely going to be necessary if they have any dream of exporting outside of China too.
I also agree that NAMICS breaking exclusivity is doubtful!
The HKMG barrier you mentioned is interesting because it's more about process mastery than hard equipment limits. That probably means CXMT could catch up faster there than with the EUV stuff, but it's still gonna slow them down in the near term. It's kinda like how they've worked around DUV limits before but this time the workaround needs way more refinement to get yields up.
Great note. We have a lot of tourists in semiconductor investing today and when the cycle turns a lot of folks are going to get hurt. If anyone reading this has capital at work in semiconductor and really wants to learn how the industry works, this is a superb place to start learning. Kudos. Alex.
Excellent read! Is there any chance that CXMT can leapfrog into using hybrid bonding directly and not deal with MRMUF?
My understanding is that right now, hybrid bonding is too expensive and overkill for HBM. I think D2W bonding accuracy still isn’t good enough for HBM.
First, great overview! Second, I'd like to get your take on the capacity of CXMT for HBM2 and HBM3. My understanding is that access to any kind of HBM4 is bottlenecking the AI accelerators that are Made in China. Especially with the US embargos in place, while going for HBM3e and 4 makes sense for CXMT, if CXMT could produce HBM3 in large-enough quantities, it would already increase the capabilities of China's domestic AI accelerators considerably.
Lastly, I doubt that NAMICS would break their exclusive arrangement with SK Hynix, while an overt breaking of IP laws and international agreements by making a knock-off would severely affect China's export-oriented industry.
Thanks! Producing large amounts of HBM3 would definitely be an increase in capabilities for China, but I don't think it's as relevant if it's not keeping pace with Western capabilities. I think CXMT is reasonably trying to leapfrog where they can and catch up with the West rather than being permanently a few generations behind. That is definitely going to be necessary if they have any dream of exporting outside of China too.
I also agree that NAMICS breaking exclusivity is doubtful!
To clarify: the second sentence should have read "My understanding is that access to any kind of HBM is bottlenecking...".
Thanks Substack for not allowing edits, and Gboard for unhelpful autocomplete 😜