EUV非中国AI芯片瓶颈，DUV足以支撑竞争

I think the EUV issue is a red herring at this point, tbh. Americans focus on it *because* it's a defensible moat, but the only thing *this moat durably protects* in AI is accelerator export markets; given the shortages, US capex, and aspects of arms race and weaponization of AI, Americans won't be exporting a lot. DUV is sufficient to decide volume production of AI-relevant chips, which will be sufficient to enable China to stay a player in the strategic AGI competition, provided it's not taken out in the process somehow. Ryan does care about exports; but they'll definitely take a backseat to Wunderwaffe considerations in any case. The part with DUV "5 units in 2026" smells like a mistranslation/broken telephone, I am almost certain that >5 machines have been delivered to date. So the real question is the timeline to scaling production of subcomponents for DUV machines around the level of NXT 1980Di (give or take), QC for consumables, and metrology for stuff like mirrors; integration will come along the way. The whole thing is getting tedious. The endless hype-mongering around sci-fi requirements of photolithography is just main character syndrome. It ignores the core fact that everything up to "7nm grade DUV" (say, 2019) had been achieved by a relatively small and poor industry, with key nodes heavily concentrated in three small nations, drawing on a limited technical talent contested by lucrative non-engineering careers. And because the domestic talent was… insufficiently incentivized, even the "espionage" and "talent poaching" angles are misconstrued – China receives not some janitors who looked over the shoulder at genius wypipo blueprints or stole a pen drive, they are getting back some of the *developers* of the current-era stack, which let's say does simplify the absorption of the intel. Trumpf, Zeiss, ASML, even Cymer – what does it all matter? How many engineer-hours was that? And how many hours adjusted for modern AI-boosted productivity and better basic science? Might be <5 years of labor by 2026-grade Huawei alone. Despite the astronomical gap in technical specifics, the conversation is analogous to "the Chynese can't make Quality German ICE cars". Well, they did specialize in EVs, but they absolutely can make any ICE now and lead in several categories like heavy diesel efficiency. So, I think the serious questions are: timeline to Dario's Durable Strategic Advantage (a function of peak capabilities, gap + inference capacity, inherent offense/defense balance and amenability of general R&D to automation), and how well the maturation of components of DUV stack can be parallelized. Other things will get rendered irrelevant conditional on these ones.