Korean Chip Stocks Crashed 13%. That Same Day, a Top Chinese AI Lab Was Lining Up More Nvidia Chips.

On the day Samsung Electronics fell about 13% and SK Hynix about 14%, among their steepest single-day drops on record, a Beijing-based AI lab was reported to be securing additional Nvidia Blackwell GPUs for its next model. That juxtaposition is the story: Korean memory stocks got priced on a "China has achieved semiconductor self-sufficiency" narrative while one of China's leading AI startups was still buying the American chips that narrative claims to replace.

South Korea's KOSPI index, the country's main stock benchmark, opened down more than 5% near 6,400 and broke below 6,000 intraday, according to The Guardian, erasing three months of gains for the country's two largest chip names. Both stocks had entered the session with stretched valuations: SK Hynix had briefly become the world's most valuable chipmaker by market cap, drawing comparisons to Cisco at the peak of the dot-com bubble, when a network-equipment leader's stock had run far ahead of its actual earnings. Leveraged products tracking the two names collapsed 29% in a single session, amplifying the move.

The trigger was a substitution story, not an order-book event. CXMT, China's leading DRAM maker, debuted on Shanghai's STAR Market on July 27 with shares up roughly 466% on the first day, according to The Guardian, and the OZ Talking newsletter pointed to parallel reporting that a Shanghai state-owned enterprise had moved into mass production of immersion DUV lithography, the deep-ultraviolet machines used to etch patterns onto silicon wafers. The narrative that followed, that China had leapfrogged Korean memory and the ASML-led Western lithography stack, moved faster than the procurement data.

On the same trading day, The Information reported that Moonshot AI, the Beijing-based lab behind the open-weight Kimi K3 model (2.8 trillion parameters, released July 16), was working to secure additional Nvidia Blackwell GPUs for its next-generation "Kimi K4." Jensen Huang's response to the same dynamic, that "free AI is good for hardware," captures the demand side of the argument: cheap, open-weight Chinese models do not necessarily shrink the addressable market for AI compute, including high-bandwidth memory (HBM, the stacked memory chips that sit next to AI accelerators). They may expand it by training and serving more models, on more hardware, in more geographies. The Guardian also flagged a parallel signal: AMD and Core Scientific announced a 2.5-gigawatt AI data-center partnership, with the first 500 megawatts scheduled to come online in 2027, a reminder that the demand pipeline beyond Nvidia is still expanding.

CXMT controls roughly 11% of global DRAM wafer capacity as of 2024, with projections near 15% by 2028, and is reportedly converting about 20% of its Shanghai fab to HBM3 production. That is meaningful, but it is not a replacement of the HBM tier that trains frontier models. On the lithography side, the gap is wider still. China is on track to install about 5 domestic immersion DUV units this year; ASML shipped 131 last year. The Chinese tools operate at the 28nm class using ArF immersion, a mid-2000s ASML technology tier, two to three generations behind current extreme-ultraviolet (EUV) and High-NA EUV systems. Per Goldman Sachs analysis cited by OZ Talking, running sub-7nm processes on DUV-only multi-patterning implies a 40–50% cost penalty compared with TSMC. ASML's stock fell 8% in sympathy with the Korean selloff, but The Guardian noted that analysts at StoneX, Morningstar, and Swissquote described the move as overdone.

The U.S. MATCH Act, which is before Congress and not yet law, would cut off not just new exports but maintenance and support for already-installed ASML tools at SMIC, Hua Hong, CXMT, Huawei, and YMTC. If enacted, it would give even generationally-backward domestic tools a new value: as insurance against a service cutoff on the Western-installed base. That dynamic is part of why Chinese fabs and Chinese lithography names have moved in sympathy, and why the Korean tape is reading as a referendum on the whole stack rather than on memory alone.

The legitimate criticism embedded in the selloff is that the market is pricing a narrative, not an order book. The strongest falsifier is also the open question: if open-weight Chinese models genuinely compress Big Tech infrastructure spend and HBM orders, and if CXMT's HBM3 line scales on time and at yield, the substitution story will be vindicated by the next earnings cycle, not the trading tape. Until then, the gap between the Seoul tape and the Moonshot procurement desk is the most concrete read on how AI supply chains are being priced, and on which way the next leg moves when the order book catches up to the narrative.