Cathie Wood, founder of ARK Investment Management, is keeping her distance from the memory chip sector and cautioning that the industry is prone to a "brutal" boom-and-bust cycle, a problem that could only worsen as Chinese companies accelerate capacity expansion.

Wood told investors that once supply outpaces demand, memory chip prices could fall rapidly. While high-bandwidth memory (HBM) remains in tight supply at present, the growing entry of Chinese manufacturers into the market could intensify price volatility across the entire industry. "The trend is up, but the cycle swings are very brutal," she said, adding that "a few more competitors, especially from China, would amplify market volatility."

Wood's view is not without merit. For example, Chinese memory chip maker ChangXin Memory Technologies (CXMT) is ramping up production of advanced memory chips to meet surging artificial intelligence (AI) demand. Although CXMT still lags in the HBM space, it is steadily making progress in the broader DRAM market.

Data compiled by Counterpoint Research shows that in the second quarter of this year, CXMT's share of the DRAM market expanded to 10%, up from just 4% a year earlier. Meanwhile, news of CXMT's progress on next-generation high-density DRAM architectures has also sparked concerns about intensifying competition and dragged shares of Samsung Electronics and SK Hynix sharply lower in South Korea on Tuesday.

Memory chip stocks such as Micron Technology (MU.US) and SanDisk (SNDK.US) also weakened in U.S. premarket trading on Tuesday. It is reported that CXMT is accelerating its next-generation DRAM technology roadmap and plans to launch DDR5 RDIMM server memory products using a 4F虏 cell architecture before the end of this year.

For a long time, major global memory manufacturers including Samsung Electronics, SK Hynix and Micron Technology have been researching next-generation high-density DRAM architectures. However, due to extreme difficulty in design complexity, manufacturing processes and yield control, large-scale commercialization has progressed relatively slowly. Reports indicate that Samsung Electronics' mass production timeline for a similar technology route is still expected to take several years. If that assessment is accurate, CXMT could gain roughly a two-year advantage in commercializing 4F虏 architecture-related products. Being able to manufacture more complex products would also give the company a more favorable competitive position in the coming years as global supply tightness begins to ease.

In addition to the news of CXMT's progress on next-generation high-density DRAM architectures, remarks by executives at two major chip giants also contributed to the selloff in memory chip stocks. Broadcom (AVGO.US) CEO Hock Tan said Monday at an event in Singapore that as the entire industry races to build infrastructure to provide computing power for increasingly complex AI models, financing could become a bigger obstacle to AI expansion, potentially even more significant than shortages of semiconductors, memory chips or electricity.

Intel (INTC.US) CEO Lip-Bu Tan, who appeared on the same stage, also expressed concerns about financing. He pointed to rising capital costs, while the market is also questioning whether companies making massive AI investments will ultimately generate sufficient returns. "If there is any challenge right now, it is capital and the cost of capital. On the other hand, in the long run, people want to see return on investment (ROI)," Lip-Bu Tan said.

Hock Tan also said the AI boom has attracted a group of investors whose investment strategies resemble those of gamblers, and he believes some of them have no understanding of what they are actually betting on. "A lot of investors have no idea what they are investing in. They set extremely outrageous expectations, only to find themselves almost like they are in a casino," Hock Tan said. He added that several AI companies, including OpenAI, are losing billions of dollars annually. Such services require massive investment in high-end chips, data centers and model development, and the industry has yet to find a sustainable business model.

Wood disagrees with the remarks by Hock Tan and Lip-Bu Tan. She dismissed market fears of a bubble and expects AI to significantly boost productivity.

Why is Wood staying away from the memory sector?

It is worth noting that back in August, Wood responded to questions about her "missing the memory rally." Wood pointed out that memory chips are the "most commoditized and most cyclical" link in the semiconductor supply chain. Samsung Electronics, SK Hynix and Micron together control about 90% of global DRAM output, with products that are highly standardized and easily substitutable, and historically prices could multiply several times within two years or be cut in half within two years.

In her view, although HBM currently carries the label of a "core AI component," its commercial nature has not departed from the essence of the memory industry 鈥?it is a link that can be standardized, price-compared and bypassed. Wood also noted that the surge in memory prices is not a positive signal but a negative one. She pointed out that when the price of a supply chain element rises threefold, fourfold or tenfold, it is no longer a cost item that can be accepted over the long term, but becomes a technical problem that must be solved.

She further noted that recently the free cash flow of chip stocks and that of hyperscalers have moved in completely opposite directions 鈥?chipmakers benefit from price increases, while cloud providers come under pressure. But she stressed that this state is temporary, because the party under pressure has a strong incentive to change it.

Wood mentioned that the demand side is dissolving its reliance on HBM through "engineering." She named two companies 鈥?Cerebras and Groq 鈥?both focused on AI inference chips, whose architecture designs do not require external HBM and instead rely on on-chip high-speed memory. She likened this process to Tesla "engineering out cobalt" in battery design 鈥?when a raw material is too expensive or its supply is risky, engineers will find every way to bypass it.

However, the judgments of the vast majority of sell-side institutions and industry players stand in sharp contrast to Wood's views. Their core argument is that this memory cycle is not a traditional cycle driven by inventory rhythms in phones and PCs, but a structural shortage driven by AI computing investment. The memory chip price increase trend will last at least until 2027, and supply and demand are only expected to ease substantially after significant new capacity is released in 2028.

A more important structural change is the widespread adoption of long-term agreements (LTAs). Samsung, SK Hynix and Micron have reserved large amounts of capacity in advance for major North American cloud providers through 2-5 year long-term contracts, with some contracts even removing price caps and adding prepayment clauses. Bulls view this as institutional evidence that "the memory industry is moving from a strong cycle to a long boom."

Comparing Wood with mainstream market views, it can be seen that Wood is warning against treating a highly cyclical commodity as a core asset that can be held permanently. Yet the market's current mainstream view also has solid foundations: the supply-demand gap is the largest in 15 years, new capacity will not be released until 2028, and long-term agreements have pushed visibility beyond 2027. These are facts determined by capacity construction cycles and contract terms.

The tension between the two is precisely what makes this memory cycle most interesting: the cycle is being prolonged because prices are too high; and prices being too high will ultimately spawn the forces that end the cycle. For investors, the memory super-cycle may still be here, but the phase of "buying with eyes closed" is most likely over.