Five machines.
That is the number the market did not price when it wiped more than €8bn off ASML in a single session on July 27, after The Information reported that a state-backed firm had reportedly begun mass-producing homegrown immersion DUV scanners. Applied Materials, Lam Research, and KLA fell alongside it. Traders repriced the long-term addressable market in China on a headline about "mass production."
Now set the headline beside the arithmetic. China's domestic immersion output target is roughly five machines in 2026, and around twenty in 2027. ASML shipped about 130 immersion systems in 2025 and expects about 130 again this year before increasing capacity by 30% in 2027. At its 2027 target, the Chinese line produces roughly one scanner for every six or seven ASML ships in a year that has already passed.
The tool itself reportedly resembles ASML's Twinscan NXT:1950i — a system ASML brought to market around 2008. Its optics still come from Japan. It is designed for 28nm-class work, the silicon inside cars and appliances, not flagship logic.
This week the story moves one layer down, into the tools that print the wafers and the memory that feeds the chips — and the picture that emerges is not a country closing the frontier gap. It is a country deliberately choosing not to try.
That is the subject: the distance between the market's reflex and the actual bet Beijing is placing.
Start with the tooling. Independent analysis from the AI Futures Project reportedly put commercially viable Chinese immersion DUV in the mid-2030s, with ASML currently holding 98.7% of the immersion market. Qualifying a new scanner into a production line can take many months. The five machines arriving this year reportedly go to SMIC, Hua Hong, and CXMT — and all three sit on the list of firms a bill in Congress would cut off from ASML by statute.
That last fact is the point. The MATCH Act, reportedly reported out of the House Foreign Affairs Committee on April 22 with a Senate companion filed as S. 4281, does not just block new exports. Its immersion DUV provisions cover servicing and technical assistance — which reaches the installed fleet already humming inside Chinese fabs, the fleet those fabs have spent two years stretching through secondary-channel upgrades.
So a domestic scanner that trails ASML by most of two decades is not built to win a performance contest. It is built to make a servicing embargo survivable. Five inferior machines that no American statute can switch off are worth more, strategically, than a hundred superior ones that Congress can strand.
Five inferior machines that no American statute can switch off are worth more, strategically, than a hundred superior ones that Congress can strand.
The harder half of that problem is the half that is not yet Chinese. ASML's machines draw on more than 5,100 specialized suppliers and over 100,000 components per system, anchored by Carl Zeiss optics and Trumpf light sources refined over decades. Yuliangsheng's tool is mostly domestic — but the precision optics still come from Japan, and replacing that remaining slice is widely regarded as the hard part, precisely because the easy part is already done.
Now the memory layer, where the same logic repeats in cash terms.
ChangXin Memory Technologies closed its first day on Shanghai's STAR Market up roughly 466% from its 8.66 yuan offer, at 49 yuan — a market capitalization near 3.3 trillion yuan, about $487bn, vaulting it past ICBC to the top of the mainland market. Asia's largest IPO of 2026 raised 57.92 billion yuan.
Read the prospectus and the ambition sharpens into something narrower than the valuation implies. It splits 29.5 billion yuan of named project spending across three lines: 13 billion for DRAM technology upgrades, 9 billion for next-generation DRAM research, and 7.5 billion for wafer manufacturing line upgrades.
Nearly 70% of the named spending, then, flows into DRAM CXMT already makes. There is no dedicated HBM project. Not a yuan of the itemized spend is earmarked for the high-bandwidth memory that sits beside every frontier AI accelerator.
That is not an oversight. Conventional DRAM yields more than three times the bits per wafer that HBM does, and SemiAnalysis models CXMT's 8-high HBM3 yield at around 25%. HBM is the frontier's hardest, lowest-yielding memory — and CXMT, lacking EUV, is routing around it exactly as the scanner routes around servicing risk.
The tell is in the retail aisle. A 64GB DDR5-5600 RDIMM built on CXMT dies reportedly sells for 18,999 CNY on JD.com. The same module on Samsung or SK hynix silicon sells for 18,595 CNY.
The Chinese part is 2.2% more expensive — not cheaper.
Everyone expected CXMT to be the budget savior that broke big-three pricing. It isn't. Its dies are made on older processes, draw more power, and overclock poorly, yet they track — or slightly exceed — big-three prices, because in a supply-constrained market there is no incentive to undercut and every incentive to milk the surge.
That is the whole strategy in one price tag. This is not memory that competes on the frontier or even on cost. It is un-sanctionable bulk, priced at what a shortage will bear.
And there is bulk. CXMT will add roughly 85,000 wafer starts per month of DRAM this year — against 60,000 at SK hynix, 30,000 at Micron, and 15,000 at Samsung, per SemiAnalysis — putting it on course for around 350,000 by the end of 2026, within 25,000 of Micron's total per Citrini Research's model. Output is reportedly booked through 2027, with Dell, HP, Lenovo, and Apple ahead of smaller buyers, plus a $7bn ByteDance server-DRAM deal and a $3bn Tencent deal.
That volume is the geopolitical asset. It is enough conventional DRAM to matter, all of it made on tools no export rule can reach.
The wall is visible in the analyst spread. Nomura opened with a buy and a 116 yuan target on rising global share; Morningstar puts fair value at 14.90 yuan — under a third of the close — citing the lack of EUV as the ceiling on conventional scaling. Nomura's own downside case, built on equipment and materials embargoes, cuts 2027–2028 profit by up to a third.
So the €8bn selloff read the story backwards. The threat to ASML is not that China is about to match it — the mid-2030s forecast and the 2008-era design say otherwise.
The threat is subtler: that China is engineering enough domestic capacity in the un-sanctionable half of the market — mature-node logic tools and commodity DRAM — that export controls lose their bite where they can still theoretically apply. EUV, HBM, and the Japanese optics inside the DUV scanner mark the wall.
Beijing is not trying to climb it. It is building a very large, very cheap floor beneath it — and buying statutory immunity in the process.