Veteran Says TSMC Can’t Do 3D DRAM, Flags CXMT’s 10 New Fabs

Charles Kau (高啟全), the Taiwanese memory veteran who now chairs 兆捷科技, said on 10 November that TSMC will not move into DRAM manufacturing and that the bigger threat to memory makers is China’s CXMT, which he said is building 10 wafer fabs, TechNews reported. DigiTimes quoted him as saying “台積電做不來3D DRAM”, which we translate as: TSMC can’t do 3D DRAM.

He said CXMT’s new capacity is expected to come fully online within two years and could push the market into oversupply. CXMT’s monthly DRAM capacity is expected to rise from about 320,000 wafers to 420,000 by 2027, Commercial Times reported in September; that is about 31% more, by our arithmetic.

Business Pill · The Memory Wall

An evergreen Business Engineer explainer on why memory bandwidth limits AI chips. It teaches the general idea only and does not discuss Charles Kau, TSMC or CXMT.

Why He Says TSMC Will Not Make DRAM

Kau was responding to market talk, which we reported last week, that TSMC could move into 3D DRAM at Nvidia’s request. According to TechNews, he said logic foundry and memory process technology and experience are completely different, and recalled that Vanguard International Semiconductor (世界先進) once tried memory and returned to foundry work.

On the technology, he said DRAM has only reached the 6F² cell layout and no maker has achieved 4F². United Daily News reported that 4F² could in theory cut cell area by about a third, and that Kau said the main obstacle is leakage: “會有很多漏電”, which we translate as: there will be a lot of leakage. DigiTimes reported that TSMC said it does not comment on market speculation.

What Was Posted on X

Dan Nystedt relayed the Taiwanese media reports early on 11 November UTC, describing Charles Kau as a former chairman of YMTC and of Taiwan’s Inotera Memories, and giving the timing of the expected flood of Chinese memory as about two years (late 2028). This piece works from the Chinese-language reports themselves (TechNews, United Daily News and DigiTimes).

The Two-Year Warning on CXMT

TechNews reported that Kau said CXMT is expanding 10 wafer fabs in Beijing, Shanghai and elsewhere. He said the view held by some Taiwanese firms, “中國產能僅供內需，不致影響全球市場”, was a short-term impression formed in a good market. We translate that view as: Chinese capacity is only for domestic demand and does not touch the global market.

According to TechNews, he said that even on mature processes, without EUV tools, the volume would leave the market in severe oversupply, and that Taiwan’s memory makers would be hit first once Chinese capacity fully arrives in two years. He also said international customers such as Apple would still buy from China when supply is short.

What the Capacity Numbers Show

CXMT’s global DRAM revenue share rose to 9.5% in the second quarter from 7.6% in the first, putting it fourth behind Samsung, SK hynix and Micron, TrendForce reported. Against TrendForce’s second-quarter industry revenue of nearly $154.73 billion, that implies about $14.7 billion for CXMT, by our arithmetic.

The September Commercial Times report put the timing differently from Kau. It said the industry sees China’s expansion affecting global supply and demand possibly after 2027, with CXMT’s new supply going first to China’s PC, phone and server markets.

Where He Credits China

Kau did not dismiss Chinese memory makers. TechNews reported that he said YMTC, whose team he led in its early days, has world-leading flash technology and licenses patents to Samsung and Micron. Our YMTC piece covers its rise in China’s NAND market. On X, Dan Nystedt described Kau as a former chairman of YMTC and of Taiwan’s Inotera Memories.

The Structural Read

As we read it, Kau moves the question from who can build next-generation DRAM to who will have the most ordinary DRAM capacity in two years. Whether that capacity reaches the global market or stays in China is the point on which he and the September Commercial Times report differ.

Three Implications

FOR TSMC Kau said logic foundry and memory process technology are completely different and that TSMC will not move into DRAM manufacturing, TechNews reported; DigiTimes reported that TSMC said it does not comment on market speculation.

FOR MEMORY PRICES Kau said CXMT’s capacity, even on mature processes, could leave the market in severe oversupply once it fully comes online in two years, according to TechNews.

FOR MARKET SHARE CXMT’s share of DRAM revenue rose to 9.5% in the second quarter from 7.6% in the first, TrendForce reported, about $14.7 billion of quarterly revenue by our arithmetic.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Memory Chokepoint.

The framework says: “Memory is becoming the new compute. HBM growth is outpacing accelerator demand, making memory the rate-limiting factor in the AI scaling equation.”

As we read it, Kau’s warning concerns the other end of that market: if HBM is the scarce layer, conventional DRAM capacity added in China is where he expects prices to come under pressure. The framework predates his remarks and does not discuss them.

What Is Not Established

The 10 fabs and the two-year timing are Kau’s statements as reported by TechNews; we read press reports of his remarks, not a transcript. Past its opening, the DigiTimes article sits behind a member login. Nystedt’s ‘late 2028’ is his reading of ‘about two years’. The 31%, the $14.7 billion and the English translations are ours.

The Bottom Line

Charles Kau said TSMC will not enter DRAM manufacturing and that CXMT’s 10 wafer fabs could flood the memory market within two years, according to TechNews. CXMT held 9.5% of DRAM revenue in the second quarter, TrendForce reported, and its monthly capacity is expected to reach 420,000 wafers by 2027, according to Commercial Times.