Chip stocks fall as CXMT advances and Broadcom warns on AI bets

A broad retreat across Asia

South Korea’s Kospi fell 6.25%. Japan’s Nikkei dropped 1.8%, and Taiwan’s market lost almost 2%, with TSMC shares down by a similar amount.

US stock futures pointed to a mixed open on Wall Street. The Nasdaq Composite fell 0.4% on Monday. Micron Technology shares lost 3.9%, while Nvidia, Advanced Micro Devices and Broadcom also declined. Oil prices fell as tensions in the Middle East escalated.

AI-related stocks have swung sharply in recent months as investors question the huge sums companies are spending on AI infrastructure, much of it financed with debt.

CXMT adds pressure on memory makers

The latest wave of selling followed a Monday report by the South China Morning Post that Chinese company CXMT had made significant progress on a new DRAM architecture designed to increase memory density. Investors fear it could challenge Samsung, SK Hynix and Micron, the world’s largest memory-chip makers.

CXMT, which makes DRAM used to store data for other AI chips to process, successfully listed on the Shanghai Stock Exchange in late July.

Revenue is not the same as profit

At a conference in Singapore, Broadcom CEO Hock Tan warned that the AI boom had drawn in investors who do not understand what they are betting on. Many, he said, build too much into their expectations and end up like players at a casino table, Bloomberg reported.

Tan said OpenAI’s annualized revenue, based on its performance at the end of September, had reached $50 billion (£38 billion). He called that evidence of a viable business model, but said it revealed little about long-term profitability. OpenAI and other AI companies remain unprofitable; Tan said OpenAI is still losing money because it spends heavily on training models. The current figures, in his view, show the potential for substantial revenue from providing advanced generative AI models.

He also said a notable share of AI labs’ corporate revenue comes from pilot projects, comparing them to slot machines: a child gets $10 in tokens from their parents, spends them in 10 minutes and learns nothing.

I think that distinction matters more than the revenue headline. A pilot can show that a company will pay to experiment with AI; it does not establish that the use will last or justify the infrastructure bill. The market is being asked to price both a possible new memory competitor and a business model whose impressive sales figures, by Tan’s own account, have yet to prove lasting profitability.