China's STAR 50 hands back roughly 70% of its three-month rally

Compounding the two moves changes the shape of the quarter. A rise of nearly 75% followed by a fall of about 30% leaves the STAR 50 at roughly 1.75 times 0.70, or about 1.23 times its level when the rally began ^[1]. Of the 0.75 it gained, about 0.525 is gone ^[2]. A holder from the start of the rally is still up by a little over a fifth ^[1].

Cryptobriefing, the source for all of these figures, attributes the fall to investors questioning whether the money going into AI infrastructure will earn proportional returns ^[5]. It reports that semiconductor indices in South Korea, Taiwan and the US slid at the same time ^[4]. The article does not say how far those indices fell, and that figure decides the argument. A drop in Seoul, Taipei and New York close in size to the STAR 50's July loss ^[2] would make the doubt global and about AI spending itself. A drop of a few points would leave most of China's fall with a Chinese cause.

The same article offers two Chinese causes. The first is efficiency. It says open-source models such as Moonshot's Kimi K3 have undercut the assumption that AI leadership needs the most expensive hardware stack ^[9]. If competitive systems can run on leaner infrastructure, it argues, the case for premium semiconductor and storage companies gets harder to defend ^[10]. That argument concerns which suppliers get paid out of AI budgets, and it can hold even if the budgets earn their returns.

The second cause is China itself. The broad blue-chip index also fell hard in July ^[6], though the STAR 50 fell about three times as far ^[4]. Sellers moved their money into defensive sectors ^[13].

I think the efficiency case, on top of a fast rally unwinding, fits the price action best. Yuanjie Semiconductor and Hua Hong Grace Semiconductor each had single-day drops of 14% or more during July ^[8]. The STAR Market was set up for high-growth companies that might not meet the profitability rules of older exchanges ^[12], so its prices rest on earnings several years out. A lower assumed hardware bill per AI model cuts those distant earnings, and stocks priced on them move further than a blue-chip index does. The counter-thesis is the source's own. It wins if the Korean, Taiwanese and US chip indices turn out to have fallen by similar amounts.