Real ML is data center based, then, until you have enough electricty production and performant cooling, you could have 28nm ML chips, it should not matter in the end. But while doing that, China should not forget to keep doing its R&D for 2N/18A tooling (include EUV).
Namely, you should be able to dodge the usage of "nearly whole wafer" ML chips manufactured with a 18A/2N process. At a price of a more ML time with more energy consumption.
MonkeyClub•4mo ago