全球DRAM产业格局的结构性变化：长鑫存储扩产、4F²架构与成本竞争的趋势研判（2026—2030）

独立性声明。 本文为独立的第三方产业分析，不持任何国家、企业或意识形态立场，也不构成投资建议。文中数据来自企业公告、行业研究机构和主流媒体的公开报道；凡属企业自述或分析机构估计的内容，均在文中注明。

内容摘要

2026年10月中旬，全球存储芯片行业接连出现几条受到广泛关注的消息：中国最大的DRAM（动态随机存取存储器）制造商长鑫存储（CXMT）宣布在4F²存储单元架构上取得关键进展；埃隆·马斯克发文称”中国将在2028年大量生产RAM”；英伟达首席执行官黄仁勋表示对与长鑫合作”持开放态度”。这些消息叠加在一轮由人工智能驱动的存储”超级周期”之上，引发了关于全球内存产业格局是否将发生结构性变化的讨论。本文的主要判断如下：

1. 市场结构已经出现变化，但集中度仍然很高。 2026年第二季度，三星、SK海力士和美光合计占全球DRAM营收的87.6%；长鑫的份额升至9.5%—10%，是增长最快的供应商。

2. 决定格局变化的核心变量是成本，而不只是技术。 多家分析机构估计，长鑫目前的每比特成本仍高于三大厂商20%至30%以上；当前的高利润主要来自行业价格上涨。新一代G5平台和4F²架构的意义，在于有可能缩小这一成本差距。

3. 4F²是存储单元物理结构的改变，不涉及接口标准。 4F²产品仍遵循通用的DDR5等接口规范。它对行业的影响主要体现在单位晶圆产出和制造路线上，而不是”重新定义标准”。目前该技术尚无公开的产品规格和独立验证，产业化仍需观察。

4. 市场正在分层。 高带宽内存（HBM）仍由韩美企业主导；长鑫的主要增长空间在规模更大的通用DRAM市场，包括个人电脑、手机、汽车和通用服务器。

5. 设备条件是中长期的根本约束。 国产浸没式DUV光刻机已开始小批量生产，EUV光源样机已公开展示，但距离大规模商用仍有数年时间。

6. 情景研判。 2026—2028年最可能出现的是”有序扩张”情景：长鑫的份额稳步提高，行业价格在供需偏紧的环境下保持高位，格局的显著变化更可能出现在2028年以后，具体取决于新架构的良率和成本表现。

一、事件背景

2026年10月10日，长鑫存储总裁曹堪宇在上海举行的集成电路与芯粒大会上介绍了公司的4F² DRAM架构，并表示计划在年底前推出基于该技术的服务器DDR5 RDIMM内存条。1 消息公布后，三星和SK海力士的股价当日下跌超过7%，美光下跌约4%。2

10月15日，马斯克在社交平台X上写道：“中国将在2028年大量生产RAM。”这条帖子回应的是韩国《每日经济新闻》的一篇专栏。专栏认为，新产能陆续投产可能在2029年前后导致供过于求，韩国只有两到三年时间培育新的增长产业。3 在此之前，马斯克还表示，美光在内存定价上”较为合理”，并认为高带宽内存（HBM）的技术门槛”要高得多”。4 韩国媒体分析认为，作为大型采购方，马斯克的表态可能与2028年前后的价格谈判有关。5

10月22日，黄仁勋在柏林GTC大会期间接受CNBC采访。他表示，预计到2029年不会出现内存供过于求；被问及是否考虑与长鑫合作时，他回答：“我们持开放态度。”6 10月21日，美光则在技术说明会上宣布不推进4F²架构的商业化。7

这些消息的共同背景是一轮罕见的存储景气周期。理解这一背景，是判断产业格局走向的前提。

二、市场结构：份额与超级周期

2.1 市场份额

根据TrendForce的统计，2026年第二季度全球DRAM产业营收为1547.3亿美元，环比增长59.5%。8 各厂商的份额如下：

厂商	2026年第二季度营收（亿美元）	环比增速	市场份额
三星电子	609.8	63.4%	39.4%
SK海力士	385.9	37.9%	24.9%
美光科技	360.0	65.5%	23.3%
长鑫存储	146.2	99.3%	9.5%
南亚科技	26.1	68.3%	1.7%

资料来源：TrendForce。9

韩国两家企业的合计份额约为64%。Counterpoint Research的统计口径略有不同，显示长鑫第二季度份额为10%，高于上年同期的4%。10 需要说明的是，美光是总部位于美国爱达荷州的美国企业，其重要的DRAM制造基地设在台湾地区。

2.2 超级周期与价格

这一轮景气周期的主要驱动力是人工智能。大型语言模型的训练与推理带动服务器内存需求快速增长，存储企业把更多产能转向HBM和服务器DRAM，通用内存的供应因此偏紧。TrendForce的数据显示，DRAM合约价第二季度环比上涨约80%，第三季度约40%，第四季度预计约20%，涨幅逐季收窄，但仍在上涨。11 其中，通用型DRAM第四季度合约价预计环比上涨10%—15%；PC内存的供应在2027年可能进一步收缩。12

HBM对晶圆的占用远大于对比特产出的贡献。TrendForce估计，到2026年底，HBM约占三大厂商DRAM晶圆投入的22%，但只占比特供应的约9%。13 这意味着，向HBM转移产能会明显减少通用内存的供应，这正是长鑫获得市场空间的主要背景。

三、长鑫存储的发展状况

3.1 经营业绩

长鑫科技于2026年7月27日在上海证券交易所科创板上市。14 公司业绩预告显示，2026年前三季度营业收入预计为2650亿至2700亿元，同比增长726%至741.5%；归母净利润预计为1310亿至1325亿元，上年同期亏损52.8亿元。15 第三季度营收环比增长约15%至20%，净利润环比增长约1%至4%，增速放缓主要与行业价格涨幅收窄有关。16 摩根士丹利估计，公司第二季度毛利率约为87.6%。17

3.2 产能扩张

长鑫目前在合肥运营两座12英寸晶圆厂，在北京运营一座，每座月产能约10万片；在建的上海和合肥项目全部投产后，总产能有望翻倍至每月60万片以上。18 半导体研究机构SemiAnalysis预计，长鑫的月产能将从2026年底的约35万片增至2027年底的42万片和2028年底的50万片，2027年约占全球DRAM产能的17%；但按比特出货量计算，其2027年的份额约为12%。19 产能与出货之间的差距反映了一个基本事实：从建成厂房到稳定出货，还需要经过设备安装、工艺验证、良率爬坡和客户认证等环节。

3.3 技术进展

2026年9月，长鑫宣布第五代DRAM技术平台（G5）进入量产。按8Gb产品折算，该平台单片晶圆的毛芯片数比上一代至少增加50%；同时推出两款单颗容量24Gb的LPDDR5X产品。20 公司表示，该平台的开发结合了计算机仿真以及与国内设备厂商在关键工序上的合作。21 需要注意的是，“毛芯片数”指剔除缺陷之前的潜在芯片数量，并不等同于良率提升。

四、4F²架构：标准变革还是结构改良

4.1 技术本质

DRAM的每个存储单元由一个晶体管和一个电容组成。“F”指工艺的最小特征尺寸，“6F²”和”4F²”表示单个存储单元占用的面积。目前主流产品采用6F²结构；4F²通过把晶体管由水平改为垂直，使单元面积在理论上缩小约三分之一，同样面积内可容纳的单元数量增加30%至50%。22 长鑫的方案采用自主开发的垂直环绕栅极（VSG）晶体管，公司称已拥有300多项与4F²相关的专利族。23 这一数字属于企业自述，尚未经过独立核实。

一种直观的说法是：同样大小的芯片可以容纳更多存储容量，同一片晶圆可以切出更多芯片。

4.2 物理结构与接口标准是两个层面

从这个角度看，4F²属于物理结构层面的改良，不是接口标准的变革。内存产品必须与英特尔、AMD等处理器平台兼容，接口和协议遵循行业通用的JEDEC标准。长鑫计划推出的首款4F²产品是DDR5 RDIMM，仍属于标准化的服务器内存条。24 因此，4F²的竞争意义不在于”制定新标准”，而在于在相同标准下，有可能以更高的密度和更低的单位成本生产产品，并在制造路线上形成一套自主的专利组合。

4.3 行业各方的路线选择

主要厂商对4F²的判断并不一致：

• 三星： 2026年10月14日，三星半导体部门首席技术官宋在赫确认公司正在开发4F²技术，但没有公布时间表。25 行业媒体报道的计划是2026年完成开发、2027年进行质量测试、2028年转入量产。26
• 美光： 在10月21日的技术说明会上，首席技术与产品官斯科特·德博尔表示，美光研究4F²约20年，认为该结构可以制造，但成本更高、工艺更复杂，在先进节点上存在根本性的微缩障碍。他认为，决定DRAM密度的是电容间距，4F²相对6F²的密度优势”有些言过其实”。美光将继续推进6F²微缩，并同步开发3D DRAM。27,28
• 长鑫： 计划年底前”推出”首款4F²产品。目前公司尚未披露产品的容量、速率、良率和客户等具体信息。29 行业分析指出，服务器内存从样品到完成主流平台认证，通常需要12至18个月。30

分析判断： 4F²是否能转化为成本优势，取决于量产良率和工艺复杂度带来的额外成本，而这两项目前都缺乏公开数据。各厂商的不同选择说明，这一技术路线的经济性在行业内部仍有争议。较为稳妥的判断是：4F²为长鑫提供了一条在现有设备条件下提高密度的可能路径，其实际效果需要以2027—2028年的量产数据来检验。

五、成本：格局变化的核心变量

5.1 当前的成本状况

存储芯片具有较强的大宗商品属性，长期竞争力主要取决于每比特成本。分析机构对长鑫成本状况的估计基本一致：

• SemiAnalysis估计，长鑫DDR5的每比特成本比三大厂商高30%以上；2026年第一季度，其DRAM平均售价只比三大厂商低约5%至10%。31
• Futurum的分析师认为，长鑫在工艺上落后两到三代，每比特成本高20%至30%。32

成本差距的主要来源是设备条件。由于无法使用EUV光刻机，长鑫依靠DUV设备和多重曝光技术实现精细图形，工序更多，成本也更高。33 在当前供不应求、价格高企的环境下，这一差距对利润的影响被行业价格上涨所掩盖。

5.2 缩小成本差距的路径

长鑫缩小成本差距主要有三条路径：

• 工艺平台升级： G5平台单片晶圆的毛芯片数增加50%以上。34
• 架构改良： 4F²在理论上可进一步提高单位面积密度。
• 规模效应： 产能从每月30万片向50万片以上扩张。35

此外，国产设备的成熟也会对成本产生长期影响，详见第七部分。

5.3 “价格低10%至15%即可形成优势”的条件

一种分析观点认为，长鑫不必追求比竞争对手便宜30%，只要在质量稳定的前提下价格低10%至15%，就足以对下游厂商形成很强的吸引力。这一判断符合大宗商品市场的一般规律：对年采购额以数十亿美元计的PC、手机和云服务企业来说，10%左右的成本差异已经相当可观。

但这一判断的成立需要一个前提：长鑫的每比特成本至少要接近竞争对手的水平。根据现有估计，长鑫目前的策略是跟随市场价格、保持利润率，而不是主动降价。36 因此，“便宜10%至15%”更适合理解为成本差距缩小之后的一种可能情景，而不是当前的市场状况。

5.4 价格竞争的可能性

另一个常被讨论的问题是：如果长鑫的市场份额继续扩大，现有龙头企业是否会通过大幅降价来回应。一种观点认为，龙头企业不会采取这种做法，因为这会严重损害自身利润。

从当前情况看，几方面因素确实降低了大规模价格竞争的可能性：

• 供需偏紧： 市场整体供不应求，降价缺乏必要性。
• 长期协议锁定： 三星2027年75%至80%的供应已签订长期协议。37
• 产能优先方向： 龙头企业正把产能优先投向利润更高的HBM和服务器产品。

但也需要看到，按分析机构的估计，龙头企业目前的单位成本低于长鑫，这意味着在行业下行周期中，它们的价格承受能力相对更强。DRAM行业历史上多次经历周期性价格下行和产业整合。因此，价格竞争的可能性主要取决于未来的供需周期，而不只是企业意愿。

六、市场分层：HBM与通用内存

6.1 两个市场的不同逻辑

HBM和通用DRAM正在形成两个竞争逻辑不同的市场。HBM技术门槛高、单价高，目前主要由SK海力士、三星和美光供应。马斯克本人也认为，HBM”难度要高得多”。38 通用DRAM则覆盖个人电脑、智能手机、汽车电子、游戏机和通用服务器，规模更大，接口标准化程度高，更接近大宗商品市场。

长鑫的主要增长空间在通用DRAM市场。它选择首先把4F²用于服务器DDR5，说明它也在进入数据中心的配套供应链：一台AI服务器除HBM外，还需要大量通用服务器内存。39 黄仁勋提到，英伟达已经在数据中心大规模使用手机常用的LPDDR内存，并将继续大量使用。40 这为通用内存供应商提供了新的需求来源。

6.2 AI需求的不确定性

关于AI内存需求的持续性，各方判断不一。黄仁勋认为，到2029年不会出现供过于求，“内存总是越多越好”。41 韩国专栏则担心，2029年前后可能出现供应过剩。42 也有观点认为，超大规模AI算力投资主要集中在少数国家和大型科技企业，长期需求规模存在不确定性。

从产业策略角度看，以通用内存为主的业务结构，受单一应用领域需求波动的影响相对较小，市场基础也更广泛；但通用内存的价格周期性更强，在下行周期中利润波动也更大。两种业务结构各有其风险特征。

七、设备条件与光刻机国产化

7.1 设备是中长期的根本约束

从更长远的角度看，光刻设备是影响中国存储产业成本结构的根本因素。只有在先进光刻设备上实现自主供应，多重曝光带来的额外工序和成本才有可能系统性降低。

7.2 DUV光刻机：进入小批量生产

据路透社报道，上海国资背景的上海爱盛那电子科技集团已开始量产国产浸没式DUV光刻机，并整合了宇量昇和上海微电子（SMEE）的研发团队。43 首批设备预计于2026年交付中芯国际、华虹半导体和长鑫存储，计划2026年生产约5台，2027年约20台。44 作为对比，ASML在2025年交付了131台同类设备。45 报道同时指出，国产设备在性能和可靠性上仍落后于ASML，需要进一步测试，部分关键零部件仍从日本进口。46,47 业内评论认为，首批设备主要面向28纳米制程。48

这种组织方式体现了一种”分工攻关、集成验证”的产业协同模式：各家企业和科研机构分别承担光源、物镜、工件台等子系统的研发，最终由整机企业集成，再由晶圆厂在实际产线上验证。

7.3 EUV光源：完成样机展示

2026年10月，上海朗道硅光科技在中国国际工业博览会上展示了EUV光源样机，称国产化率达到90%。公司技术人员表示，距离大规模商用还需要三到五年。49 该样机公布的EUV输出功率约为15瓦，而ASML首款用于大规模量产的光源功率为250瓦，功率提升和长期运行稳定性是主要挑战。50 光源只是EUV光刻机的子系统之一，整机研发还涉及光学系统、工件台和系统集成等多个环节。

7.4 “先易后难”的产业化路径

一种务实的产业化思路是：国产设备先用于精度要求相对宽松、需求量大的产品，在实际生产中积累数据、持续改进，再逐步进入要求更高的领域。成熟制程的家电控制芯片、电源管理芯片和部分通用存储产品，都可能成为国产设备的早期应用场景。这一路径的优势是能在真实产线中加快迭代；代价是初期的生产效率可能偏低、单位成本偏高，需要设备厂商与晶圆厂之间建立长期的协作和容错机制。

八、外部政策环境

外部政策环境是影响长鑫海外市场拓展的另一项变量。2026年6月，美国国防部更新”1260H清单”，将长鑫存储和长江存储列入其中。51 长鑫于2026年8月28日向美国哥伦比亚特区联邦地区法院提起诉讼，要求撤销这一认定，并表示公司的产品用于民用和商业用途。52

这一清单本身并不禁止私营企业之间的商业往来。根据相关法律，美国国防部自2026年6月30日起不得与清单内企业直接签约，2027年6月30日起限制间接采购；自2027年12月23日起，美国联邦机构将不得采购长鑫的半导体产品。53 截至2026年8月，长鑫未被列入美国商务部的”实体清单”。54 换言之，相关限制主要影响美国政府采购链，而全球大部分商业市场目前并无类似的法律限制。

九、情景研判（2026—2030）

情景	相对可能性（作者判断）	主要特征	关键条件
A. 有序扩张（基准）	较高，约55%	长鑫份额稳步升至12%—15%；行业价格在供需偏紧中保持高位；龙头企业继续以HBM和高端服务器产品为重点	4F²按计划小规模推出；AI需求保持增长；外部政策环境基本稳定
B. 加速替代（上行）	约25%	4F²和G5良率快速提升，成本接近行业水平；长鑫在通用内存领域具备价格竞争力，份额在2028—2030年明显提高	新架构良率达标；国产设备稳定上线；海外客户完成认证
C. 周期回调（下行）	约20%	新增产能集中释放，AI投资增速放缓，价格进入下行周期；行业利润普遍收缩，成本较高的企业承压更大	2028—2029年供给增速超过需求增速；主要客户削减资本支出

三种情景并非互相排斥。从时间上看，2026—2028年的市场格局主要受现有产能和长期协议的影响，较大的结构性变化更可能出现在2028年以后。

关键观察指标：

• 长鑫4F²产品的规格披露、客户认证进度和量产良率；
• 分析机构对长鑫每比特成本差距的最新估计；
• DRAM合约价走势，以及HBM占晶圆投入的比重；
• 长鑫实际出货份额与名义产能之间的差距；
• 国产DUV光刻机的交付数量和产线运行表现；
• 美国相关政策的变化以及诉讼进展。

十、延伸讨论：成本领先逻辑的适用范围

本文讨论的”标准化产品＋规模化制造＋成本领先”逻辑，也常被用于分析其他领域，例如成熟制程芯片、闪存，乃至操作系统和办公软件。需要注意的是，不同领域的竞争机制差异很大。

• 硬件领域： 竞争力主要取决于制造成本和产能，适用规模经济的规律。
• 基础软件领域： 边际成本很低，竞争力更多取决于软件生态、用户习惯和兼容性，即网络效应。

因此，基础软件领域的替代往往从特定场景切入，例如政府和企业办公，再逐步扩大生态。在硬件领域有效的成本策略，不能直接套用到软件领域。

从更宏观的角度看，在高度集中的市场中出现新的有效竞争者，通常有助于降低下游成本、扩大技术应用范围，并促使现有企业加快创新。这种竞争的长期效果，最终取决于新进入者能否在质量、成本和持续创新上形成稳定的能力。

十一、结论

全球DRAM产业正处在一个由AI驱动的特殊周期中。长鑫存储借助通用内存供应偏紧的市场环境，在一年内把营收份额从约4%提高到约10%，并通过G5平台和4F²架构探索在现有设备条件下提高密度的路径。

从产业经济学的角度看，格局能否发生持续的结构性变化，取决于三个条件：

• 成本： 每比特成本能否接近行业领先水平。
• 质量： 产品能否通过主流平台和全球客户的认证。
• 设备： 国产设备能否在产线上稳定运行。

这三项条件目前都处在推进之中，尚未完全得到验证。

可以较有把握地判断的是：全球DRAM市场正在从三家企业高度集中的格局，逐步走向供应来源更加多元的格局。对下游厂商和终端用户而言，供应来源的多元化有助于增强供应链的韧性和价格谈判空间。对现有龙头企业而言，竞争的重点将更多集中在HBM、先进封装和3D DRAM等技术前沿。这一演变过程的速度，将主要由2027—2028年的量产数据决定。

关于作者

殷彤博士（Dr Tong Yin） 现任 InsightBridge Global 及 InsightBridge Global Lab 首席研究官，主要从事跨领域的独立分析研究。

免责声明

本文为作者基于2026年10月23日可得公开资料作出的独立分析，仅供研究参考，不构成任何投资、法律或商业建议。文中的情景和可能性判断属于作者的分析。企业自述和分析机构的估计可能随后续信息而调整，相关数据以原始发布方为准。文中观点仅代表作者个人，不必然代表 InsightBridge Global 的立场。

Statement of independence. This is an independent third-party industry analysis. It takes no national, corporate or ideological position and does not constitute investment advice. Data come from company announcements, industry research firms and mainstream media reports; company claims and analyst estimates are identified as such.

Summary

In mid-October 2026, several closely watched developments came in quick succession:

• ChangXin Memory Technologies (CXMT), China’s largest DRAM maker, announced key progress on a 4F² memory-cell architecture.
• Elon Musk wrote that “China will be making a LOT of RAM in 2028”.
• Nvidia CEO Jensen Huang said Nvidia was “open-minded” about working with CXMT.

These developments arrived in the middle of an AI-driven memory “super-cycle”. Together they prompted debate about whether the structure of the global memory industry is about to change. The main judgements of this paper are as follows.

1. The market structure is shifting, but concentration remains high. In Q2 2026, Samsung, SK hynix and Micron together held 87.6% of global DRAM revenue. CXMT’s share rose to 9.5–10%, making it the fastest-growing supplier.

2. Cost, not technology alone, is the decisive variable. Several analyst firms estimate that CXMT’s cost per bit is still 20–30% or more above the three leaders. Its current high profits come mainly from rising industry prices. CXMT’s new G5 platform and the 4F² architecture matter because they could narrow this cost gap.

3. 4F² changes the physical structure of the memory cell, not the interface standard. 4F² products still follow common interface specifications such as DDR5. Their impact lies in output per wafer and the manufacturing route, not in “redefining the standard”. No product specifications or independent verification have yet been published, so industrialisation remains to be seen.

4. The market is segmenting. Korean and US firms lead high-bandwidth memory (HBM). CXMT’s main growth opportunity lies in the much larger conventional DRAM market for PCs, phones, vehicles and general-purpose servers.

5. Equipment is the fundamental long-term constraint. Domestic immersion DUV lithography tools have entered small-batch production, and an EUV light-source prototype has been shown publicly. Large-scale commercial use is still several years away.

6. Scenarios. The most likely scenario for 2026–2028 is “orderly expansion”: CXMT’s share rises steadily while tight supply and demand keep industry prices high. Any significant structural change is more likely after 2028 and will depend on the yield and cost performance of the new architectures.

1. Background

On 10 October 2026, CXMT President Cao Kanyu presented the company’s 4F² DRAM architecture at an integrated-circuit and chiplet conference in Shanghai. He said CXMT planned to launch a server DDR5 RDIMM module based on the technology by the end of the year.1 After the announcement, Samsung and SK hynix shares fell by more than 7% on the day and Micron by about 4%.2

On 15 October, Elon Musk wrote on X that “China will be making a LOT of RAM in 2028”. He was responding to a column in South Korea’s Maeil Business Newspaper. The column argued that new capacity could create an oversupply around 2029 and that Korea had two to three years to develop new growth industries.3 Earlier in the week, Musk had said that Micron had been “reasonable” on memory pricing and that HBM was “a much harder hill to climb”.4 Korean media read his remarks as positioning by a large buyer ahead of price negotiations around 2028.5

On 22 October, Jensen Huang spoke to CNBC on the sidelines of GTC Berlin. He said he did not expect a memory glut by 2029. Asked whether Nvidia would consider working with CXMT, he replied, “We’re open-minded.”6 The day before, on 21 October, Micron announced at a technology webinar that it would not commercialise the 4F² architecture.7

These developments share a common backdrop: an unusual memory boom. Understanding that boom is a precondition for judging where the industry structure is heading.

2. Market Structure: Shares and the Super-Cycle

2.1 Market shares

According to TrendForce, global DRAM revenue reached US$154.73 billion in Q2 2026, up 59.5% quarter on quarter.8 Supplier shares were as follows:

Supplier	Q2 2026 revenue (US$ bn)	QoQ growth	Market share
Samsung Electronics	60.98	63.4%	39.4%
SK hynix	38.59	37.9%	24.9%
Micron Technology	36.00	65.5%	23.3%
CXMT	14.62	99.3%	9.5%
Nanya	2.61	68.3%	1.7%

Source: TrendForce.9

The two Korean firms together hold about 64%. Counterpoint Research, using a slightly different method, puts CXMT’s Q2 share at 10%, up from 4% a year earlier.10 Micron is a US company headquartered in Boise, Idaho, with major DRAM manufacturing operations in Taiwan.

2.2 The super-cycle and prices

The main driver of this cycle is artificial intelligence. Training and running large language models has driven rapid growth in server memory demand. Memory makers have shifted capacity towards HBM and server DRAM, which has tightened supply of conventional memory.

TrendForce data show how prices have moved:11,12

• Contract prices: up about 80% quarter on quarter in Q2, about 40% in Q3 and an expected 20% or so in Q4. Gains are narrowing but prices are still rising.
• Conventional DRAM: Q4 contract prices are expected to rise 10–15% quarter on quarter.
• PC DRAM: supply may shrink further in 2027.

HBM uses far more wafers than its share of bit output suggests. TrendForce estimates that by the end of 2026 HBM will take about 22% of the three leaders’ DRAM wafer input but provide only about 9% of bit supply.13 Shifting capacity to HBM therefore cuts conventional memory supply sharply. This is the main reason CXMT has found room in the market.

3. CXMT’s Position

3.1 Financial performance

CXMT listed on the Shanghai Stock Exchange’s STAR Market on 27 July 2026.14 Its preliminary results for the first three quarters of 2026 show:15,16

• Revenue: 265–270 billion yuan, up 726–741.5% year on year.
• Net profit attributable to shareholders: 131–132.5 billion yuan, compared with a loss of 5.28 billion yuan a year earlier.
• Third quarter: revenue rose about 15–20% from the second quarter and net profit about 1–4%. The slowdown reflects narrowing price gains across the industry.

Morgan Stanley estimates the company’s Q2 gross margin at about 87.6%.17

3.2 Capacity expansion

CXMT currently operates two 12-inch fabs in Hefei and one in Beijing, each producing about 100,000 wafers a month. Once projects under way in Shanghai and Hefei are fully operational, total capacity could double to more than 600,000 wafers a month.18

SemiAnalysis projects CXMT’s monthly capacity as follows:19

• about 350,000 wafers at the end of 2026;
• 420,000 at the end of 2027, or about 17% of global DRAM capacity;
• 500,000 at the end of 2028.

On a bit-shipment basis, however, SemiAnalysis puts CXMT’s 2027 share at about 12%.20 The gap between capacity and shipments reflects a basic fact. Between completing a fab and shipping steadily, a company must still install equipment, qualify processes, ramp up yields and win customer certification.

3.3 Technology progress

In September 2026, CXMT announced that its fifth-generation DRAM technology platform (G5) had entered mass production. On an 8Gb-equivalent basis, the platform yields at least 50% more gross dies per wafer than the previous generation. CXMT also launched two 24Gb LPDDR5X products.21 The company said the platform was developed using computer simulation and joint work with domestic equipment makers on critical process steps.22 “Gross dies” counts potential chips before defective ones are excluded, so it does not measure an improvement in yield.

4. The 4F² Architecture: A New Standard or a Structural Refinement?

4.1 What the technology is

Each DRAM cell consists of one transistor and one capacitor. “F” is the minimum feature size of a process, and “6F²” and “4F²” describe the area a single cell occupies. Mainstream products today use a 6F² layout. A 4F² design stands the transistor up vertically instead of laying it flat. In theory this shrinks the cell area by about one-third and fits 30–50% more cells into the same area.23

CXMT’s approach uses a vertical surrounded gate (VSG) transistor that it developed itself. The company says it holds more than 300 patent families related to 4F².24 This figure is a company claim and has not been independently verified.

Put simply, the same size of chip can hold more memory, and more chips can be cut from the same wafer.

4.2 Physical structure and interface standards are different layers

4F² is therefore a refinement of physical structure, not a change in interface standards. Memory products must work with processor platforms from Intel, AMD and others, so their interfaces and protocols follow industry-wide JEDEC standards. CXMT’s first planned 4F² product is a DDR5 RDIMM, which remains a standardised server memory module.25

The competitive significance of 4F² does not lie in “setting a new standard”. It lies in the possibility of making products to the same standards at higher density and lower unit cost, and in building an independent patent portfolio around the manufacturing route.

4.3 How the main suppliers are positioned

The main manufacturers disagree about 4F²:

• Samsung: On 14 October 2026, Song Jae-hyuk, chief technology officer of Samsung’s semiconductor division, confirmed that the company is developing 4F² technology but gave no timetable. Industry media report a plan to complete development in 2026, run quality tests in 2027 and move to mass production in 2028.26
• Micron: At its 21 October technology webinar, Chief Technology and Products Officer Scott DeBoer said Micron had studied 4F² for about 20 years. He said the structure can be built, but at higher cost and with more process complexity, and that it faces fundamental scaling problems at leading-edge nodes. In his view, capacitor pitch is what determines DRAM density, and the density advantage of 4F² over 6F² is “a bit of a myth”. Micron will keep scaling 6F² while developing 3D DRAM in parallel.27,28
• CXMT: It plans to “launch” its first 4F² product by year-end. It has not yet disclosed capacity, speed, yield or customer details.29 Analysts note that server memory typically takes 12 to 18 months to move from samples to qualification on mainstream platforms.30

Assessment: Whether 4F² becomes a cost advantage depends on two things: volume-production yields and the extra cost of a more complex process. Neither is yet supported by public data. The suppliers’ differing choices show that the economics of this route are still contested within the industry. The prudent judgement is that 4F² offers CXMT a possible path to higher density with its existing equipment. Its real effect will have to be tested against volume-production data in 2027–2028.

5. Cost: The Core Variable in Structural Change

5.1 The current cost position

Memory chips behave much like commodities, so long-term competitiveness depends mainly on cost per bit. Analyst estimates of CXMT’s cost position broadly agree:

• SemiAnalysis: CXMT’s DDR5 cost per bit is more than 30% above the three leaders. In Q1 2026, CXMT’s average DRAM selling price was only about 5–10% below theirs.31
• Futurum: CXMT is two to three process generations behind and pays 20–30% more per bit.32

The gap stems mainly from equipment access. Without EUV lithography, CXMT relies on DUV tools and multiple patterning to form fine features, which adds process steps and raises cost.33 With supply tight and prices high, rising industry prices currently mask the effect of this gap on profits.

5.2 Paths to narrowing the cost gap

CXMT has three main paths to narrowing the gap:

• Platform upgrades: the G5 platform raises gross dies per wafer by more than 50%.34
• Architecture: 4F² could, in theory, further raise density per unit area.
• Scale: capacity is expanding from about 300,000 towards more than 500,000 wafers a month.35

The maturing of domestic equipment will also affect costs over the long run, as Section 7 discusses.

5.3 When is a 10–15% price advantage enough?

One analytical view holds that CXMT does not need to be 30% cheaper than its rivals. If its quality is stable, a price 10–15% lower would already be very attractive to customers. This fits the general logic of commodity markets. For PC, phone and cloud companies whose annual purchases run to billions of dollars, a cost difference of around 10% is substantial.

This view rests on one condition: CXMT’s cost per bit must at least approach that of its competitors. Current estimates suggest CXMT is following market prices and protecting margins rather than cutting prices.36 A 10–15% discount is therefore best understood as a possible scenario once the cost gap narrows, not as the present state of the market.

5.4 The likelihood of price competition

A related question is often raised: if CXMT keeps gaining share, will the incumbents respond with sharp price cuts? One view holds that they will not, because doing so would seriously damage their own profits.

Several current factors do make large-scale price competition less likely:

• Tight supply: demand still exceeds supply across the market, so price cuts are unnecessary.
• Long-term agreements: Samsung has reportedly committed 75–80% of its 2027 supply under such contracts.37
• Capacity priorities: the incumbents are directing capacity first to higher-margin HBM and server products.

However, by analyst estimates the incumbents’ unit costs are currently lower than CXMT’s, which gives them more room to absorb price pressure in a downturn. The DRAM industry has gone through repeated cyclical price declines and consolidation. Whether price competition emerges will therefore depend mainly on the future supply-and-demand cycle, not only on company intentions.

6. Market Segmentation: HBM and Conventional Memory

6.1 Two markets with different logics

HBM and conventional DRAM are becoming two markets with different competitive logics.

• HBM: high technical barriers and high prices. SK hynix, Samsung and Micron currently supply it. Musk himself called HBM “a much harder hill to climb”.38
• Conventional DRAM: used in PCs, smartphones, automotive electronics, game consoles and general-purpose servers. The market is larger, highly standardised and closer to a commodity market.

CXMT’s main growth opportunity lies in conventional DRAM. Its decision to apply 4F² first to server DDR5 shows that it is also entering the data-centre supply chain: besides HBM, an AI server needs large amounts of conventional server memory.39 Huang said Nvidia already uses LPDDR, the memory common in phones, at scale in data centres and will keep using it in large quantities.40 This creates a new source of demand for conventional memory suppliers.

6.2 Uncertainty about AI demand

Views differ on how durable AI-driven memory demand will be:

• Huang: sees no oversupply by 2029, saying “you can always use more” memory.41
• The Korean column: worries about oversupply around 2029.42
• Another view: very large AI computing investment is concentrated in a few countries and large technology firms, which makes long-term demand uncertain.

From a strategic standpoint, each business mix carries its own risk profile. A business centred on conventional memory depends less on swings in any single application and rests on a broader market base. Conventional memory prices are, however, more cyclical, so profits swing more widely in downturns.

7. Equipment and Domestic Lithography

7.1 Equipment as the fundamental long-term constraint

Over the longer term, lithography equipment is the fundamental factor shaping the cost structure of China’s memory industry. Only domestic supply of advanced lithography tools would allow the extra steps and costs of multiple patterning to fall systematically.

7.2 DUV lithography: small-batch production

According to Reuters, Shanghai Aishengna Electronic Technology Group, a Shanghai state-owned company, has begun mass-producing domestic immersion DUV lithography tools. It has absorbed development teams from Yuliangsheng and Shanghai Micro Electronics Equipment (SMEE).43

• Deliveries: the first tools are expected to go to SMIC, Hua Hong Semiconductor and CXMT in 2026.44
• Volumes: about 5 tools are planned for 2026 and about 20 for 2027. By comparison, ASML shipped 131 such systems in 2025.45,46
• Gaps: the domestic tools still trail ASML’s in performance and reliability and need further testing. Some key components still come from Japan.47,48
• Target node: industry commentary suggests the first tools mainly target the 28-nanometre node.49

This way of organising the work reflects a model of “division of labour followed by integration and validation”:

• different companies and research institutes develop subsystems such as the light source, projection optics and wafer stage;
• a system integrator assembles the complete tool;
• chipmakers then validate it on real production lines.

7.3 EUV light source: a prototype on display

In October 2026, Shanghai-based Langdao Silex showed a prototype EUV light source at the China International Industry Fair. It claimed a 90% domestic content rate. Company staff said large-scale commercial use is still three to five years away.50

The prototype’s stated EUV output is about 15 watts. By comparison, ASML’s first source for high-volume manufacturing delivered 250 watts. Raising power and achieving long-term operating stability are the main challenges.51 The light source is only one subsystem of an EUV scanner; a complete machine also requires optics, wafer stages and system integration.

7.4 An “easier first, harder later” path

One pragmatic path to industrialisation is to use domestic tools first for high-volume products with looser precision requirements. Producers can then gather data in real production, keep improving the tools, and move step by step into more demanding segments. Possible early applications include:

• mature-node control chips for household appliances;
• power-management chips;
• some conventional memory products.

The advantage of this path is faster iteration on real lines. The cost is lower productivity and higher unit costs at first. It therefore requires long-term cooperation, and tolerance for early problems, between toolmakers and chipmakers.

8. The External Policy Environment

External policy is another variable affecting CXMT’s expansion abroad. In June 2026, the US Department of Defense updated its “1260H list”, which includes CXMT and Yangtze Memory Technologies.52 On 28 August 2026, CXMT filed suit in the US District Court for the District of Columbia seeking to overturn the designation. It stated that its products are for civilian and commercial use.53

The list itself does not prohibit commercial dealings between private companies. Under the relevant laws, restrictions take effect in stages:54

• From 30 June 2026: the Department of Defense may not contract directly with listed companies.
• From 30 June 2027: indirect procurement through supply-chain intermediaries is restricted.
• From 23 December 2027: all US federal agencies will be barred from procuring CXMT’s semiconductor products.

As of August 2026, CXMT was not on the US Commerce Department’s Entity List.55 In short, the restrictions mainly affect the US government procurement chain. Most commercial markets worldwide currently impose no comparable legal restrictions.

9. Scenarios, 2026–2030

Scenario	Relative likelihood (author’s judgement)	Main features	Key conditions
A. Orderly expansion (baseline)	Higher, about 55%	CXMT’s share rises steadily to 12–15%; prices stay high amid tight supply; incumbents keep focusing on HBM and high-end server products	4F² launches on a small scale as planned; AI demand keeps growing; external policy broadly stable
B. Accelerated substitution (upside)	About 25%	Yields on 4F² and G5 improve quickly and costs approach industry levels; CXMT becomes price-competitive in conventional memory and its share rises markedly in 2028–2030	New architectures reach target yields; domestic equipment runs reliably; overseas customers complete qualification
C. Cyclical correction (downside)	About 20%	New capacity arrives in a wave while AI investment slows and prices turn down; industry profits shrink, with higher-cost producers under more pressure	Supply growth outpaces demand in 2028–2029; major customers cut capital spending

The scenarios are not mutually exclusive. In timing terms, the market in 2026–2028 will be shaped mainly by existing capacity and long-term contracts. Larger structural change is more likely after 2028.

Key indicators to watch:

• specifications, customer qualification progress and volume yields for CXMT’s 4F² products;
• updated analyst estimates of CXMT’s cost-per-bit gap;
• DRAM contract price trends and HBM’s share of wafer input;
• the gap between CXMT’s actual shipment share and its nameplate capacity;
• delivery numbers and line performance of domestic DUV tools;
• changes in relevant US policy and the progress of CXMT’s lawsuit.

10. Extension: Where Cost-Leadership Logic Applies

The logic discussed here combines standardised products, large-scale manufacturing and cost leadership. It is often applied to other fields as well, such as mature-node chips, flash memory, and even operating systems and office software. The competitive mechanisms in these fields differ greatly, however:

• Hardware: competitiveness depends mainly on manufacturing cost and capacity, and economies of scale apply.
• Foundational software: marginal costs are very low, and competitiveness depends more on software ecosystems, user habits and compatibility, that is, on network effects.

Substitution in foundational software therefore usually starts in specific settings, such as government and enterprise office work, and then gradually widens the ecosystem. A cost strategy that works in hardware cannot simply be transplanted to software.

From a broader perspective, a new and effective competitor in a highly concentrated market usually lowers downstream costs, widens the use of technology and prompts incumbents to innovate faster. The long-term effect of such competition depends on whether the new entrant builds lasting capabilities in quality, cost and continuous innovation.

11. Conclusions

The global DRAM industry is in an unusual AI-driven cycle. CXMT has used tight supply of conventional memory to raise its revenue share from about 4% to about 10% within a year. Through its G5 platform and the 4F² architecture, it is exploring ways to raise density with its existing equipment.

From an industrial-economics perspective, a lasting structural change depends on three conditions:

• Cost: whether CXMT’s cost per bit can approach the industry’s best.
• Quality: whether its products can pass qualification on mainstream platforms and with global customers.
• Equipment: whether domestic tools can run reliably on production lines.

All three are in progress and none has yet been fully demonstrated.

What can be said with reasonable confidence is that the global DRAM market is moving gradually from a structure dominated by three firms towards more diverse sources of supply.

• Downstream manufacturers and end users: more supply sources strengthen supply-chain resilience and widen room for price negotiation.
• Incumbents: competition will increasingly centre on technology frontiers such as HBM, advanced packaging and 3D DRAM.

The pace of this transition will be decided mainly by volume-production data in 2027–2028.

About the Author

Dr Tong Yin is Chief Research Officer of InsightBridge Global and InsightBridge Global Lab and conducts independent cross-disciplinary research and analysis.

Disclaimer

This paper is an independent analysis based on public information available as of 23 October 2026. It is provided for research purposes only and does not constitute investment, legal or business advice. The scenarios and likelihood judgements are the author’s own analysis. Company claims and analyst estimates may be revised as new information emerges; data are subject to the original publishers. The views expressed are the author’s own and do not necessarily represent the position of InsightBridge Global.