Two headlines, one from a chip boss and one from a Chinese rival, were enough to knock the wind out of semiconductor stocks this week. Chipmakers are now going through their roughest stretch in about 16 months, and the pain was sharpest in Asia on Tuesday.
The first spark came from Broadcom CEO Hock Tan. Speaking at an event in Singapore on Monday, he said many AI investors don't really understand what they're putting their money into, and warned that some are chasing extreme expectations in a way that resembles a casino. He compared companies experimenting with AI to kids burning through arcade tokens without learning much, and said access to financing could become the industry's biggest constraint, ahead of even memory shortages. Coming from the head of one of the biggest AI chip suppliers, the comments landed hard.
The second came from China. CXMT, the country's leading DRAM maker, said over the weekend it had made key progress on a design known as 4F², which shrinks the space each memory cell takes up by about a third and could, in theory, pack roughly 50% more cells onto a chip. CXMT plans a server memory product using the design by the end of 2026. Rival Samsung reportedly targets mass production of similar technology by 2028, so on announced dates CXMT would be about two years ahead.
Markets didn't wait for details. In Seoul on Tuesday, SK Hynix fell about 10%, Samsung dropped roughly 7%, and South Korea's Kospi lost more than 5%. Japan's Kioxia slid around 13% and SoftBank Group fell 8.5%, while TSMC eased about 2%. That followed a Monday session in New York where Micron fell 3.88% and the Philadelphia Semiconductor Index dropped 2.67%. Back in the U.S. on Tuesday, the Nasdaq ended lower as chipmakers weighed, and the chip-heavy Nasdaq 100 slipped about 0.3% while the broader S&P 500 finished roughly flat thanks to gains in other sectors.
It's worth separating the noise from the fundamentals. Neither story shows AI chip demand actually falling. Tan's warning was aimed at investor behavior, and Broadcom itself still forecasts its AI chip revenue climbing from about $58 billion this fiscal year to $230 billion by fiscal 2028. CXMT, for its part, has not published specs, yield rates or customer approvals for the new design, and it has a track record of delays. Its DDR5 memory, first targeted for 2022, didn't become widely available until late 2024.
Still, CXMT's momentum is real. According to Counterpoint Research, its share of global DRAM revenue climbed to 10% in the second quarter of 2026, up from 4% a year earlier, and that was before 4F² has shipped. That trend is a growing concern for Samsung, SK Hynix and Micron, which have enjoyed strong pricing power during the AI boom.
MarketWatch notes that a broader sense of AI fatigue is building on Wall Street, with investors growing less patient with heavy spending and looking ahead to Anthropic's upcoming investor day for clues. With ASML and TSMC due to report earnings this week, the chip sector's next moves may depend on whether those results can calm nerves or deepen them.