CXMT is adding 50,000 monthly wafers of HBM capacity
I have no idea where "2M HBM stacks" comes from. Now CXMT should be around 50-60K WPM for HBM. You get, what, 2TB HBM3E per wafer? 75PB needed per SuperCluster, 37.5 thousand wafers. 3 weeks of output. I may be way way off but it doesn't look like a bottleneck. @zephyr_z9
If HBM is the binding constraint and can all be utilized, this would mean China can ship like 24 GW of compute annualized by mid-late 2027 (using Huawei's specs as the baseline). Likely this is the upper range, but I think "1-2% of the US" Hawks cite is ruled out. Even 10% is.
Also, CXMT isn't the only HBM manufacturer Huawei gets their supply from XMC Chinese HBM capacity will be in the 150k WPM range by mid to late 2027
Chinese CSPs have tendered for over 4.5GW of IT capacity this year
If HBM is the binding constraint and can all be utilized, this would mean China can ship like 24 GW of compute annualized by mid-late 2027 (using Huawei's specs as the baseline). Likely this is the upper range, but I think "1-2% of the US" Hawks cite is ruled out. Even 10% is.
@teortaxesTex @zephyr_z9 此HBM非彼HBM。事实上华为用的是自有协议。HBM所有的技术标准都掌握在韩国人手里，这对于AI芯片公司来讲是不利的。我怀疑几年之后英伟达也会转向自有协议。
@morterlier @zephyr_z9 I don't care what Huawei calls it, I care that they claim 4 TB/s.