CXMT Memory Chip Advance Jolts Global Stocks as Investors Reassess China's Challenge

11 Min Read
CXMT Memory Chip Advance Jolts Global Stocks as Investors Reassess China's Challenge

A smaller memory cell triggers a global selloff

China's ChangXin Memory Technologies plans to introduce a new server memory product before the end of 2026, using an architecture that could increase memory density by about 30%. The announcement sent shares of established chipmakers sharply lower, even though turning the design into reliable mass production will take further work.

Contents
  1. A smaller memory cell triggers a global selloff
  2. What 4F² means in plain language
  3. How CXMT is changing the manufacturing process
  4. A launch is followed by a longer production test
  5. Why the market reaction reached beyond memory makers
  6. Why HBM remains a different contest
  7. A rising market share and a crowded development calendar
  8. Export controls shape the route to expansion
  9. Customers and valuation raise the stakes
  10. Key Points

CXMT president Cao Kanyu presented the progress at an industry conference in Shanghai on October 10. The company is pursuing a design called 4F², which reduces the space occupied by each memory cell. Its planned product targets data centre servers, placing it in more direct competition with Samsung Electronics, SK Hynix and Micron Technology.

On October 13, SK Hynix shares closed about 10% lower and Samsung fell 7.4%, with both reaching their weakest levels since August. South Korea's Kospi index dropped 5.5%. Micron lost 3.62% in Tuesday trading after falling roughly 3.9% on Monday.

The reaction spread across chip equipment suppliers, storage companies and other technology stocks. Investors were reassessing how quickly a Chinese manufacturer could challenge businesses that have benefited from demand for artificial intelligence computing.

CXMT's immediate target is conventional server memory. High-bandwidth memory, or HBM, the specialised memory used alongside powerful AI processors, remains a stronger position for its established rivals. The commercial test for CXMT will be producing enough dependable chips at a competitive cost.

What 4F² means in plain language

DRAM, short for dynamic random-access memory, holds information that a computer is actively using. A server needs this working memory to keep data close to its processors while they carry out tasks. Each chip contains a vast number of tiny memory cells.

The industry's common 6F² layout gives each cell an area equal to six times the square of a minimum feature size, represented by F. A 4F² layout reduces that area to four times the same square. At an unchanged feature size, the cell itself occupies about one third less space.

Advertisement

Smaller cells can allow more information to fit on a chip. The gain across a complete product also depends on the space occupied by other circuitry and on manufacturing choices. Morningstar analyst Jing Jie Yu estimates that successful commercialisation of CXMT's architecture could increase its memory density by 30%.

Zhang Xiaolu, a deputy director at consultancy CIC, estimates a potential increase of 30% to 50% in the number of bits produced across a wafer, alongside a roughly 33% reduction in individual cell area. These estimates describe the potential of the design rather than measured output from established mass production.

Samsung and SK Hynix are also exploring 4F² designs. SK Hynix presented a development roadmap in 2025, making the timing of CXMT's proposed introduction especially sensitive for investors assessing the technological lead of existing suppliers.

How CXMT is changing the manufacturing process

Cao described an approach called Vertical Surrounding Gate, or VSG, which changes the arrangement of the transistors that control memory cells. The company expects the structure to reduce electrical leakage and other problems that become harder to manage as cells shrink.

Electrical leakage means current escapes where it is not wanted. In a memory chip, controlling that current helps cells retain information reliably while limiting wasted power. A smaller design is useful only if its cells continue to function consistently.

CXMT also plans to manufacture the memory cells and their peripheral circuitry on separate silicon wafers, then join them through hybrid bonding. Peripheral circuitry handles functions such as selecting cells and moving information into and out of them. Hybrid bonding connects the separately manufactured parts through extremely precise electrical and physical connections.

Cao said separating those manufacturing steps was central to the company's approach. Zhang said the combination of VSG and bonding at wafer level could raise density without depending on extreme ultraviolet lithography, or EUV, an advanced technology used to print tiny chip patterns.

A launch is followed by a longer production test

CXMT is expected to introduce a DDR5 RDIMM using the new architecture by the end of 2026. DDR5 is a generation of computer memory, while an RDIMM is a memory module commonly used in servers. The product must work reliably with the machines and systems that customers already operate.

Advertisement

Production yields, the proportion of manufactured chips that work well enough to sell, remain unclear. Cao did not disclose detailed yield figures or a firm schedule for commercial deliveries at scale. A process that produces many defective chips can be expensive even when the underlying design is promising.

Jing Jie Yu, the Morningstar analyst, described both the competitive opportunity and the manufacturing challenge:

A successful and earlier implementation of 4F² architecture in DRAM production would vastly improve CXMT's competitive position versus peers,

Yu also expects a meaningful gap between the product unveiling and commercialisation at viable yields, pointing to the difficulty of maintaining uniform results in mass production.

Zhang said server memory validation typically takes 18 to 24 months. During that process, customers check whether products meet demanding requirements for compatibility, reliability and operation. A product introduction therefore begins a further stage of testing and adoption before broad deployment.

Why the market reaction reached beyond memory makers

Japan's SoftBank Group fell 8.5%, semiconductor equipment supplier Tokyo Electron dropped 7.2%, and testing equipment maker Advantest declined 2.9%. The Nikkei 225 index lost 1.8%, showing that the selling extended beyond companies competing directly with CXMT.

European shares also retreated. ASML fell 3.3%, while STMicroelectronics and Infineon each lost about 5%. Storage companies SanDisk, Seagate and Western Digital also declined. CXMT itself closed 2.8% lower in Shanghai on Tuesday.

The selloff followed an already weak Monday session in the United States, when the Nasdaq Composite fell 0.44%. Vey-Sern Ling, a senior equity adviser at Union Bancaire Privee, said investors were taking profits in AI stocks that had performed strongly, especially memory companies.

Phillip Wool, chief market strategist at Rayliant Global Advisors, considered the scale of the selling out of proportion to the immediate competitive threat. He pointed to the work still needed for CXMT to manufacture the new design in volume.

The next day's trading showed a smaller decline for SK Hynix, which fell 1.66% on October 14. Investors were weighing a possible longer term shift in competition against the much slower practical process of qualifying chips and expanding production.

Why HBM remains a different contest

HBM moves large amounts of data rapidly between memory and a processor. It is particularly valuable in AI systems because powerful processors need a continuous flow of information to avoid sitting idle. Producing it involves demanding manufacturing, stacking and packaging processes.

Advertisement

CXMT's proposed DDR5 server product addresses another part of a data centre's memory requirements. General server memory supports the wider system, while HBM supplies especially fast access for demanding processors such as Nvidia's AI accelerators.

Wool described the established suppliers' position in HBM:

More importantly, we're not even talking about HBM here, and the Korean chipmakers are still unquestionably dominating that space, with very little threat at all from the likes of CXMT or anyone else for the foreseeable future,

Zhang said major overseas manufacturers have concentrated capacity on HBM3e and HBM4, while conventional server DDR5 supply remains tight. That creates a potential opening for another dependable supplier of general server memory, particularly for Chinese customers seeking a domestic source.

Ling also sees a difference between the longer term competitive challenge and current manufacturing capabilities. Chinese memory companies could affect the balance of supply and demand over time, he said, while continuing to face technology difficulties in the short term.

A rising market share and a crowded development calendar

Counterpoint Research estimates that CXMT's share of global DRAM revenue rose from 4% in the second quarter of 2025 to 10% in the second quarter of 2026, making it the world's fourth largest DRAM supplier. An HSBC analysis also put its share at around 10%, up from about 4% a year earlier.

The increase amounts to six percentage points, or two and a half times its earlier share. It gives CXMT a growing commercial base while it works on its next architecture. The company is also progressing towards LPDDR6 smartphone memory on a timetable broadly similar to Samsung and SK Hynix.

CXMT's recent announcements show several projects advancing together:

  • September 20, 2026: The company announced mass production on its fifth generation DRAM process platform, called G5, and introduced high capacity LPDDR5X memory.
  • September 28, 2026: It announced plans for a research and development project costing 24.1 billion yuan and a second phase testing base costing 10.8 billion yuan.
  • October 10, 2026: Cao presented progress on DRAM devices, 4F² architecture and hybrid bonding in Shanghai.
  • By the end of 2026: CXMT aims to introduce its new server memory product.

The G5 platform uses four exposures during patterning and reaches an active area half pitch of 11.95 nanometres, a measure of spacing within the memory array. The two planned investment projects total 34.9 billion yuan and cover process development, packaging, testing and module assembly.

Advertisement

Export controls shape the route to expansion

CXMT, founded in 2016 and headquartered in Hefei, benefits from China's large domestic electronics market. Its access to advanced manufacturing equipment is constrained by export restrictions, including limits affecting ASML's most advanced EUV systems.

The proposed architecture places more weight on the arrangement of devices and the joining of separately manufactured wafers. Zhang said it could achieve higher density without EUV, while creating demand for precision etching, deposition, bonding and polishing equipment.

Competitors are investing on a different equipment and production timetable. Japan has committed up to 536 billion yen to Micron's Hiroshima expansion, supporting advanced DRAM development and production, with initial shipments targeted as early as 2028. That project includes EUV equipment and other manufacturing tools.

Competition also carries legal tensions. South Korean prosecutors brought charges against ten people over an alleged Samsung technology leak involving manufacturing knowledge said to have reached CXMT. Prosecutors alleged that the transferred information helped accelerate its DRAM production and supported HBM development. The allegations remain subject to criminal proceedings.

Customers and valuation raise the stakes

CXMT's July listing on Shanghai's STAR Market produced a debut share price surge of nearly 466%, briefly making it the most valuable company listed in mainland China. Its shares later stood about 4% below the quoted debut day price of 49 yuan.

The company was valued at about 3.46 trillion yuan, compared with Tencent's 3.87 trillion yuan. Those figures place CXMT at roughly 89% of Tencent's valuation, showing the scale of investor expectations attached to a memory manufacturer founded only a decade earlier.

Customer adoption will determine how far its commercial reach extends. Apple has been testing CXMT DRAM for devices sold in China and seeking permission from the US government for broader use. No completed supply agreement or wider approval was disclosed.

Citi analysts have argued that sales to data centres in regions such as Europe could affect Micron even if US restrictions limit direct sales into the American market. A supplier can exert competitive pressure through customers elsewhere, changing which manufacturers win orders and how much they can charge.

For server operators, the practical prospect is another source of conventional memory if CXMT delivers dependable products in sufficient quantities. For established manufacturers, the pressure is to protect their technical lead while competing in both conventional DRAM and more specialised memory. The next milestones are the planned product introduction, customer validation and stable production.

Advertisement

Key Points

  • CXMT aims to introduce server memory using a denser 4F² architecture before the end of 2026.
  • SK Hynix fell about 10% and Samsung 7.4% on October 13, while selling spread across global technology stocks.
  • The design combines vertical transistor structures with separately manufactured wafers joined through hybrid bonding.
  • Analysts see potential density gains, but production yields and customer qualification will determine commercial progress.
  • Counterpoint estimates CXMT's DRAM revenue share increased from 4% to 10% between the second quarters of 2025 and 2026.
  • The planned product targets conventional server memory, while established suppliers retain a stronger position in HBM for AI systems.
Share This Article

You May also Like