SMIC is reportedly pushing toward a 5nm-class node, on DUV, which is a sentence that makes process engineers wince. If 7nm on deep-ultraviolet light was brutal, 5nm on the same light is closer to masochism, more stacked exposures still, yields that fall off a cliff, every problem of the 7nm DUV slog made sharply worse and a few new ones stacked on top. SMIC 5nm, if it arrives in any real volume, will be the most expensive leading-edge-ish silicon in the world to manufacture. SMIC is reaching for it anyway, because the thing pulling it forward now is not phones. It is artificial intelligence.
The reason the node suddenly matters more is that the United States has banned the sale of advanced AI chips to China, the top Nvidia accelerators that train large models, the H100 and its successors, with even the cut-down China-only parts under tightening pressure, and China has to build its own or fall behind in the technology that may matter most this decade. Huawei is designing those AI chips, the Ascend line, and SMIC is the only foundry on earth that will build them for China. While TSMC builds the world’s AI silicon on 5nm and below with EUV, SMIC has to build China’s AI silicon on DUV, behind the embargo, at whatever cost it takes. The AI chip ban turned SMIC from a strategically important company into the sole load-bearing company in China’s entire technology strategy.
What has formed here is a partnership unlike anything in the rest of the industry: a designer and a foundry, both sanctioned, both national champions, welded together into the backbone of a country’s self-sufficiency drive. Huawei designs, SMIC builds, and the state funds, and between them, they are trying to reproduce, within a sanctions perimeter, what the entire global supply chain produces outside it. It is staggeringly inefficient, and that is beside the point, because it cannot be switched off from Washington. Every Ascend chip SMIC builds is a chip China did not have to buy from a company, an American export rule could block. Beijing has done the math and decided it would rather pay many times over for its own chips than depend on chips an adversary can deny it.
The honest picture is harsh. Building 5nm-class logic on DUV means yields that may be a fraction of what the leaders achieve, costs that only a state subsidy can absorb, and a tool base that the October controls and their successors are steadily strangling, because the DUV machines themselves are now restricted and SMIC cannot freely buy more or get the best ones serviced. SMIC is trying to advance to a new node while the ground beneath its existing nodes is slowly being pulled away. It is running up a down escalator that Washington keeps speeding up. Whether it can reach 5nm in any volume that matters is deeply uncertain, and the most likely outcome is small quantities at enormous cost.
Everything China wants to do in AI now runs through this chokepoint. The models, the data centers, the ambition to not depend on American silicon, all of it bottlenecks at whether SMIC can build enough advanced chips, at any cost, to feed the buildout. This is an enormous amount of weight on one foundry that cannot buy a new EUV machine and is fighting for every DUV tool. Scale it and China gets a sovereign AI hardware base, costly but real. The other outcome is worse than a missed node: the entire self-sufficiency story, every data center and model roadmap behind it, inherits a hole in the middle that money alone cannot fill. The most important number in Chinese technology is SMIC’s advanced-node yield, and almost nobody outside the company knows what it really is.
SMIC reaching for 5nm on DUV is either the next proof that the embargo only slows China down or the place where the slowing finally becomes stopping. Perso, I think SMIC gets some 5nm-class capability working but never at the volume or cost to truly feed China’s AI ambitions, that this is the node where DUV’s punishing math finally outruns even state patience, and that the gap to TSMC starts widening again rather than closing, because the leaders have EUV and SMIC never will. I have been wrong about SMIC’s ceiling before; everyone who bet against the Mate 60 was. So I hold the prediction loosely. What is not in doubt is that China will spend whatever it takes to find out, because the alternative, depending on chips an adversary can deny, is the one outcome Beijing has decided is unacceptable. The escalator keeps speeding up. SMIC keeps climbing it. Nobody knows yet who tires first.