The index number understated the damage. The VanEck Semiconductor ETF closed down more than 2%, with intraday prints closer to 4%, but the dispersion underneath was violent: SanDisk fell over 14%, AMD between 5% and 8% depending on the print, Teradyne 4% to 5%, Nvidia nearly 5% intraday, ASML’s US listing almost 6%, and BE Semiconductor roughly 8.5% in Amsterdam. Micron shed about 2%. Three separate headlines landed inside a single session and the tape treated them as one thesis. They are not one thesis, and only one of them changes the arithmetic of the sector over the next two years.
The lithography story is the weakest of the three and it inflicted the most damage. A Shanghai-based state-backed manufacturer has begun producing domestically developed 193nm immersion deep ultraviolet systems, with first deliveries expected this year to SMIC, Hua Hong and ChangXin Memory Technologies. The volumes are the entire story and they are small: roughly five systems in 2026, rising to about twenty in 2027. ASML expects to ship approximately 130 immersion units in 2026, flat against 2025, and has told investors it intends to raise immersion capacity by 30% in 2027 with another 30% under investigation for 2028. The domestic tools trail ASML on performance, reliability and build quality. Qualifying a new scanner onto a live production line is a multi-quarter exercise before a single production wafer moves through it. Independent work published in June placed commercial-scale Chinese immersion DUV in the mid-2030s, against an installed ASML immersion share near 98.7%.
The China de-rating has also largely already occurred. China fell from roughly a third of ASML net sales in 2025 to around 20% this year, and from 19% of net system sales in the first quarter to 14% in the second. The revenue that a domestic alternative would displace is revenue the market has been marking down for eighteen months. What is genuinely new in the reporting is not the machine. It is the line about Washington weighing tighter restrictions on foreign lithography tool exports and on servicing existing installed base inside China, with the MATCH Act capable of forcing Dutch and Japanese alignment with US rules inside a 150-day window. Servicing revenue is the high-margin annuity underneath every equipment franchise. That is a policy risk with a real number attached, and it is the opposite of the risk that was traded.
The spillover into the rest of European semicap was indiscriminate. Hybrid bonding is a back-end interconnect process. It has no functional relationship to a front-end immersion scanner, and a Chinese 193nm tool does not compress the die-to-wafer bonder market or alter HBM4 stacking economics. Soitec sells engineered substrates. Infineon sells power and automotive silicon. None of these are lithography franchises. The move was a sell of the category label rather than the exposure, which is the signature of positioning unwinding rather than a fundamental repricing, and it is also what a crowded long looks like when the marginal buyer steps back.
The Nvidia number is being described incorrectly in most of the coverage, including by the more prominent bears. The roughly $250 billion under discussion is a financing guarantee covering the lease and construction debt on a 10-gigawatt campus in Pike County, Ohio, developed by SoftBank’s SB Energy on a decommissioned uranium enrichment site. Chips are explicitly outside that figure. A separate and parallel negotiation covers financing of the chip purchases themselves, reported at up to $350 billion. Collapsing the two into a single circularity complaint obscures what is actually happening. On the $250 billion, Nvidia is not booking revenue against its own capital. It is lending its balance sheet and credit rating to a counterparty that cannot obtain an investment-grade rating on its own, in exchange for locking in demand. That is a contingent liability, not a revenue loop, and it is the harder criticism of the two. The circularity objection applies cleanly to the $350 billion chip leg and to the separate $5 billion equity commitment into Safe Superintelligence announced the same morning, which pairs capital with access to the Vera Rubin platform and routes the recipient’s research back into Nvidia’s own compute roadmap.
The CXMT listing is being read backwards. ChangXin raised between $8.6 billion and $9.8 billion on the Shanghai STAR Market at 8.66 yuan, rose 466% on debut and closed as the most valuable company listed on a mainland exchange at roughly 3.3 trillion yuan, or about $488 billion. It is the second-largest IPO in mainland history behind Agricultural Bank of China in 2010 and the largest mainland semiconductor offering on record. A 466% first-day move on a state-championed national asset is a statement about the Chinese retail bid, not about global memory supply. The proceeds fund conventional DRAM production and research. There is no standalone HBM project in the use of proceeds, and commercial high-bandwidth memory requires advances in stacking, packaging, test and customer validation that CXMT has not demonstrated. US restrictions continue to bar China from importing HBM.
That distinction matters more than the headline size of the raise. Chinese commodity DDR5 capacity absorbing the low end of the market is precisely what allows Samsung, SK Hynix and Micron to keep converting wafer starts toward HBM without collapsing conventional pricing. Displacement at the bottom supports conversion at the top. The supply overhang worth modelling is a 2028 and beyond question in conventional DRAM, not a 2027 question in the high-bandwidth tier where capacity remains committed well into next year. The thread connecting the first story to the third is that CXMT is named as an early customer for the domestic DUV tool, which is the trade nobody put on: a memory expansion whose equipment dependency has just been partially domesticated.
The structural signal from the session is the one that received the least attention. The 60-day rolling correlation between the software and semiconductor ETFs has fallen to approximately zero from a long-run average near 0.75, with the two moving in opposite directions in 32 of the last 60 sessions, the highest count since their 2001 inception. Both are up roughly 7% or better over three months. The AI trade has stopped being a single trade, with leadership rotating into cybersecurity and data protection names while the silicon complex absorbs a one-month drawdown of roughly 9% on the semiconductor ETF.
The bear case deserves its due. Analysts calling every leg of this drawdown mechanical rather than fundamental have been wrong for a month, and forced deleveraging can run considerably further than any valuation framework justifies. But the number that decides this is not the DUV unit count or the size of Nvidia’s guarantee. It is ASML’s immersion shipment guidance for 2027. If that 30% capacity expansion holds through the next two earnings calls, the market spent Monday selling a tool that does not exist at scale against a franchise that is being told to build more.
Leave a Reply