“存储热潮将持续3至4年”；中国技术追赶成关键变量
AI驱动的存储半导体行业热潮预计还将延续三到四年。存储厂商虽在积极扩充产能，但需求增长迅速。报道将中国的技术追赶列为影响这轮周期的关键变量。
原文深度解读
本文是对TechInsights半导体部门高级副总裁Jeongdong Choe的访谈，主题为AI驱动的存储半导体行业景气周期。文章涉及存储需求、供给缺口、价格走势、各厂商资本开支、HBM竞争格局以及中国存储厂商CXMT和YMTC的技术追赶。关键数字包括：当前供给仅满足约85%的需求，价格已上涨超过80%，DRAM和NAND均短缺约10-12%，AI存储需求短期增长4-6倍，供给预计在明年第三或第四季度达到需求的90-95%，行业高景气预计持续3-4年。与电子元器件供应链的关系在于：存储芯片是电子元器件供应链的核心品类，其供需失衡、价格高企和产能扩张将直接影响下游电子产品的物料成本和供应稳定性，而中国厂商的技术追赶可能在中长期改变存储供应格局。
- TechInsights的Jeongdong Choe判断，当前存储需求为100时实际供给能力约为85，DRAM短缺约10-12%，NAND面临类似缺口，价格已上涨超过80%，高价位预计持续3-4年。
- 供给短缺难以在3-6个月内解决，因为过去两到三年行业低迷期资本开支受限，新厂投资产出至少需要一年半到两年；预计明年第三或第四季度供给可达到需求的90-95%，但真正平衡需要供给达到102-103%。
- AI存储需求不再以历史1.2倍或1.4倍增长，而是短期激增4-6倍，每家公司仅建一两个晶圆厂不足以满足需求；全球资本开支加速约始于两年前，早期设施产出将从明年下半年开始影响供给。
- CXMT已宣布第五代（G5）DRAM工艺平台量产，G5将Active Area半间距缩小至11.95纳米；Choe分析称CXMT原路线图目标为14纳米级，但近期产品披露显示缩小至12-13纳米，且无需EUV，可通过DUV双重或四重 patterning实现，相当于跳过了另一到两代。
- YMTC在NAND领域早期商业化基于混合键合的'Xtacking'架构，并应用'Stair-Free'架构减少高堆叠NAND外围面积开销；Choe强调，仅从技术角度看，中国制造商已不再是跟随者而是竞争者。
- HBM3E中SK海力士占据超过70%市场份额，三星为12-15%，美光为10%；预计明年下半年HBM4采用加速后三星份额可能升至25-30%，SK海力士明年仍保持HBM领导地位，竞争领导格局可能在2028年发生变化。
供应链影响
- 存储芯片供给可能持续低于需求，导致DDR、LPDDR、HBM及NAND等品类价格维持高位，下游电子元器件采购成本可能承压，具体取决于AI需求增长与新增产能释放的节奏。
- CXMT和YMTC的产能扩张可能在中长期增加全球存储供给来源，但实际影响取决于其良率、经济可行性及设备获取情况，可能改变存储供应格局并加剧竞争。
- HBM产能挤占传统DRAM产能，可能使标准DDR和LPDDR供应紧张状况延续，影响依赖这些存储器的消费电子和工业电子供应链。
- 新晶圆厂建设周期较长，早期产出可能到明年下半年才开始影响供给，短期内供应链紧张局面可能难以缓解，但若AI需求持续强劲，新增产能可能被吸收而不引发结构性过剩。
- 美国等地的海外建厂面临更高资本成本、更长建设时间和劳动力成本差异，可能影响存储厂商的全球产能布局和供应链区域化趋势，具体取决于补贴力度和运营效率。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
TechInsights半导体部门高级副总裁Jeongdong Choe。（照片：THE ELEC）
AI驱动的存储半导体行业热潮预计将在未来三到四年内延续。尽管存储制造商正在积极扩大产能，但以AI数据中心为中心的快速增长需求预计将在可预见的未来超过供应。
然而，中国存储制造商的快速技术追赶仍然是一个关键变量。长鑫存储技术有限公司（CXMT）正在通过跳过工艺世代来缩小差距。在NAND闪存方面，长江存储科技有限责任公司（YMTC）已在行业领导者之前将先进技术商业化。随着中国企业同时扩大产能，存储供应竞争可能在中长期加剧。
TechInsights半导体部门高级副总裁Jeongdong Choe表示：“假设当前存储需求为100，实际供应能力约为85。虽然可能出现小幅价格波动，但高价水平将持续，支撑未来3-4年的繁荣。”
存储供应短缺预计将从明年第三或第四季度开始逐渐缓解。Choe高级副总裁预计，到明年供应水平将升至需求的90-95%。鉴于平衡市场通常需要供应超过需求2-3%，卖方市场可能会持续。
“AI存储需求不再以历史1.2倍或1.4倍的速度增长；相反，它在短期内激增4倍至6倍，”Choe指出。“每家公司仅建一两个晶圆厂不足以满足这一需求。”
与此同时，随着DRAM微缩达到物理极限，后来者缩小技术差距的空间已经扩大。CXMT最近正式宣布其第五代（G5）DRAM工艺平台量产。CXMT将其内部DRAM技术世代分类为G1、G3、G4和G5。G1对应19纳米，G3对应18纳米，G4对应16纳米级。G5工艺将Active Area（AA）半间距——一个关键微缩指标——缩小至11.95纳米。
“CXMT最初在其内部路线图中将G5目标定为14纳米级，但近期产品披露表明已缩小至12-13纳米，”Choe分析道。“即使没有极紫外（EUV）光刻设备，使用基于深紫外（DUV）的双重 patterning（DPT）或四重 patterning（QPT）技术也可以实现微缩。”他补充道，“虽然操作特性、单元尺寸和裸片尺寸仍需进一步分析，但如果他们的公告属实，CXMT又跳过了一到两代。”
在NAND闪存方面，中国的技术追赶被评估为更快。YMTC早期商业化了基于混合键合的“Xtacking”架构，在单独晶圆上制造NAND单元阵列和外围电路，然后将其键合。它还应用了“Stair-Free”架构，以最小化高堆叠NAND中的外围面积开销。
“撇开经济可行性或良率不谈，纯粹从技术角度来看，中国制造商不再是跟随者——他们是竞争者，”Choe强调。“随着DRAM和NAND的技术差距迅速缩小，中国企业的技术能力值得密切关注。”
问答
问：好久不见。今天我们邀请到TechInsights的Jeongdong Choe博士。欢迎。
答：你好。很高兴回来。
问：存储市场最近正经历显著繁荣。虽然下游市场存在一些争论，但AI投资依然强劲，有很多可讨论的内容。让我们从下游市场开始。历史上，智能手机是存储的最大市场。这种向AI的转变是否已经绝对化？
答：超过50%的需求已转向AI。
问：除了存储，功率半导体和其他组件也在转向AI和服务器。您如何看待当前的下游市场？存在相互矛盾的论点——一些人建议放缓发展速度，而另一些人主张在管理风险的同时继续积极发展。这些争论直接影响了股价。
答：市场需求正在急剧飙升，尤其是在数据中心和AI领域。消费者和企业对ChatGPT和云基础设施等服务的使用量每天都在指数级增长。存储需求已达到可用性严重受限的水平，导致服务提供商限制会话时间或提高定价。此外，随着代理式AI、高级推理和物理AI的出现，数据处理需求继续超出预期地升级。由于数据中心无法完全满足需求，传统存储库存已经耗尽。结果，DDR、LPDDR和HBM应用的价格上涨了超过80%。今天的现实是，即使以100%产能运营生产设施，市场也供不应求。DRAM供应短缺约10-12%，NAND面临类似赤字。这种情况将持续一段时间。
问：所以需求为100时，供应约为85%？
答：供应目前仅满足约85%的需求。这种短缺无法在3到6个月内解决，因为过去两到三年行业低迷期间资本支出（CAPEX）受到严重限制。新设施投资要产生产出，至少还需要一年半到两年。因此，许多分析师预计，数据中心和一般存储应用的高价水平将在未来约两年内持续。此外，AI训练中大量使用的HBM在过去一年中获得了溢价定价。像SK海力士这样的制造商已将传统DRAM产能重新分配给HBM生产，加剧了标准DDR和LPDDR存储器的短缺。简而言之，需求仍然巨大，而供应商无法完全满足，这种动态将在可预见的未来持续。
问：一些研究公司预测，这个超级周期可能延续到2030年或2031年。
答：价格动态和物理供应必须分开看待。虽然价格可能波动，但它们将稳定在比以前高得多的水平，维持未来3-4年的有利环境。目前在建晶圆厂的产出将在大约一年后进入市场，这可能导致温和的价格回落。然而，由于基准定价仍然很高，行业盈利能力应保持稳健。
问：您对需求侧有何看法？像Anthropic和OpenAI这样的软件模型开发者进展迅速，导致一些人主张适度步伐。相反，像Jensen Huang和政策制定者这样的人认为，必须在管理风险的同时继续投资。您如何评估需求侧动态？
答：我完全同意Jensen Huang的观点。虽然放缓发展的呼吁来自OpenAI等实体，但这些担忧主要源于AI向能够自主决策的代理式和物理形式过渡时的潜在风险。这些担忧应通过监管和制度框架来解决。人类对技术进步和便利的需求在历史上不断扩展；它不能被人为阻止。正如社会没有从铁器时代回到石器时代，技术进步将在更强治理的支持下继续，无需人为放缓发展。
问：批评者指出，虽然AI数据中心产生收入，但它们需要天文数字的持续资本投资，引发可持续性担忧。此外，后来者认为，领跑者呼吁监管放缓相当于拉上梯子。行业能否维持这种投资规模？
答：股票估值、私人投资者和全球政府举措——包括美国和韩国——正在积极支持AI基础设施建设。由于AI基础设施被全球公认为国家优先事项，资本投资支持可能会与这些扩张计划保持一致。
问：转向供应侧，存储制造商在过去三年中避免激进投资，但现在正在加速扩张。国内生产商的目标是到2030年将晶圆投入翻倍。美光、三星、SK海力士和铠侠的资本支出情况如何？
答：资本支出强度异常高，接近收入的30-40%。我最近访问美光博伊西工厂时，两座大型晶圆厂的建设正在快速推进。三星和SK海力士也在采取类似举措，例如SK海力士的龙仁集群和通过Solidigm大连工厂扩大固态硬盘产能，以及潜在的美国设施。铠侠也在扩张。除了西方和韩国公司，必须关注中国。在政府支持自力更生的背景下，CXMT和YMTC正在积极扩建晶圆厂以提高国内产量。CXMT的目标是将市场份额从12-13%提高到20-25%，计划在上海和北京建设多个晶圆厂，而YMTC在武汉扩张。全球资本支出加速大约始于两年前，这意味着早期设施的初始产出将在明年下半年开始影响供应。虽然后期设施需要3-4年，但强劲的AI需求应能吸收这些新产能，而不会在短期内引发结构性过剩。潜在的供应驱动型低迷可能在5-6年后重新出现，但这是长期风险，目前并未影响短期前景。
问：您预计当前供需失衡——供应仅满足85%的需求——何时会缓解至95%？
答：到明年第三或第四季度，供应应达到需求的90-95%。
问：90-95%是否被视为市场均衡？
答：90-95%的范围通常代表正常运营条件。然而，由于累积的积压，市场紧张可能再持续一两个季度，使定价在明年保持弹性。
问：随着未来几年计划持续增加晶圆产能，如果需求为100，供应达到100，是否被视为平衡？
答：供应达到需求的100%在技术上仍然紧张；真正的均衡需要供应达到102-103%。然而，关键问题是供应商能否在2-3年内达到需求的100%。AI存储需求，尤其是HBM，不是线性增长而是指数增长——与历史年增长1.2倍至1.4倍相比，短期内激增4倍至6倍。每家公司一两个新晶圆厂不足以覆盖这种需求转变。
问：长期供应协议（LTA）似乎在存储供应商和关键客户之间越来越普遍。LTA如何影响存储周期？
答：LTA有积极影响。历史上，短期季度合同导致突然取消订单或重复预订，造成极端价格波动。多年期LTA稳定定价，为客户确保可预测的供应，并允许集成设备制造商（IDM）高效规划生产计划。
问：哪些客户层级正在获得这些LTA？
答：主要的AI半导体设计商和部分超大规模企业正在签订这些多年期协议。
问：韩国存储制造商预计今年和明年将有强劲的盈利表现。您预计明年营业利润率会进一步扩大吗？
答：营业利润率应保持当前水平或明年可能略有上升，因为供应短缺将持续。然而，不同产品类别的表现各异——如HBM、商品DRAM和企业级SSD。在HBM3E中，SK海力士占据超过70%的市场份额，三星为12-15%，美光为10%。这种分布可能会持续到今年年底和明年年初。即使HBM4在2026年底或2027年初出现，HBM3E仍将是明年上半年的主要出货驱动力，维持SK海力士的优势。明年下半年，随着HBM4采用加速，三星的市场份额可能升至25-30%。三星特别针对HBM4E进行了大量研发投资，目标瞄准英伟达和超大规模客户。因此，SK海力士将在明年保持HBM领导地位，而竞争领导动态可能在2028年发生变化。
问：保持HBM市场领导地位在商业上意味着什么？
答：领导地位在供应不足的市场中赋予定价权和分配控制权。虽然SK海力士在过去两年中占据该位置，并可能在明年保持，但三星可能在2028年左右重新获得战略杠杆。
问：您如何看待美光目前在存储和HBM中的定位？
答：美光从标准DDR5和LPDDR5价格上涨中显著受益，与三星类似。然而，其HBM出货量仍然相对温和。此外，SK海力士已与台积电就基础裸片建立了紧密合作关系，而三星利用其内部代工厂生产先进逻辑基础裸片。美光面临更复杂的处境：虽然它为HBM3E和部分HBM4产品内部制造基础裸片，但最终必须依赖台积电的先进节点。由于台积电代工产能分配紧张，确保足够的基础裸片产能对美光来说是一个运营挑战。
问：今天韩国、美国和中国存储供应商之间的技术能力如何比较？
答：在DRAM方面，韩国和美国生产商之间的工艺差距已基本缩小。在10纳米后期，美光落后近一代，但在13纳米、12纳米和11纳米节点，三星、SK海力士和美光的单元尺寸、特征尺寸和设计规则几乎相同——差异小于0.1纳米。所有三家主要供应商从1b/1β节点开始使用DUV/EUV光刻和高K金属栅（HKMG）技术。
这种技术趋同使后来者能够快速缩小差距。值得注意的是，中国的CXMT已实施了相当于前三家供应商的先进功能，但不包括EUV设备。虽然CXMT目前的量产集中在1z（G3），但其最近宣布的G5工艺平台使用DUV双重 patterning（DPT）或四重 patterning（QPT）技术，无需EUV即可缩小至12-13纳米，代表跳过了一到两代工艺。
在NAND闪存方面，YMTC在多项架构创新中领先。YMTC是第一个通过其“Xtacking”架构商业化混合键合的公司，这种方法现在正被铠侠、三星和SK海力士采用。YMTC也是第一个商业化2TB QLC NAND模块并应用“Stair-Free”接触结构以减少超高堆叠NAND中外围电路开销的公司。从纯技术角度来看，YMTC作为直接竞争者而非后期跟随者运营。
问：激烈的供应竞争是否可能导致未来再次出现存储市场斗鸡博弈？
答：传统价格战不太可能，因为AI存储架构正变得高度分层。需求从HBM扩展到企业级SSD（eSSD）、CXL内存模块、GDDR以及HBF（高带宽闪存）等高密度NAND解决方案。由于AI工作负载需要跨训练、推理和代理系统的多样化、专用存储层级，扩大的市场需求应能吸收不断增长的全球供应。
问：您如何看待存储制造商在美国建设晶圆厂的地缘政治压力？
答：美国政策鼓励国内半导体晶圆厂建设，对全球供应商施加压力。然而，海外晶圆厂建设涉及显著更高的资本成本、更长的建设时间——通常是韩国的两倍——以及巨大的劳动力成本差异。即使有政府补贴，抵消这些成本差异对制造商来说仍然具有挑战性。
关于Solidigm在美国的潜在NAND投资，其目前的大连工厂面临物理扩张限制以及进口低温蚀刻机等先进设备的限制。为了扩大其高需求的192层和下一代240/250层QLC NAND（用于AI数据中心）的产量，正在战略性地评估美国或韩国国内等替代制造地点。
Jeongdong Choe, Senior Vice President of the Semiconductor Division at TechInsights. (Photo: THE ELEC)
The AI-driven boom in the memory semiconductor industry is projected to extend for the next three to four years. Although memory manufacturers are actively expanding production capacity, rapidly growing demand centered around AI data centers is expected to outpace supply for the foreseeable future.
However, the rapid technological catch-up by Chinese memory manufacturers remains a critical variable. ChangXin Memory Technologies (CXMT) is narrowing the gap by skipping process generations. In NAND flash, Yangtze Memory Technologies Corp (YMTC) has commercialized advanced technologies ahead of industry leaders. As Chinese firms simultaneously expand production capacity, competition in memory supply could intensify over the medium to long term.
Jeongdong Choe, Senior Vice President of the Semiconductor Division at TechInsights, stated, "Assuming current memory demand is at 100, actual supply capability stands at around 85. While there may be minor price fluctuations, elevated price levels will persist, sustaining the boom for the next 3–4 years."
The memory supply shortage is expected to ease gradually starting in the third or fourth quarter of next year. SVP Choe anticipates that supply levels will rise to 90–95% of demand by next year. Given that a balanced market typically requires supply to exceed demand by 2–3%, a seller's market is likely to persist.
"Demand for AI memory is no longer growing at historical rates of 1.2x or 1.4x; instead, it is surging by 4x to 6x in the short term," Choe noted. "Simply building one or two fabs per company will not be enough to meet this demand."
At the same time, as DRAM scaling hits physical limits, the margin for latecomers to close the technological gap has widened. CXMT recently officially announced the mass production of its 5th-generation (G5) DRAM process platform. CXMT classifies its internal DRAM technology generations as G1, G3, G4, and G5. G1 corresponds to 19nm, G3 to 18nm, and G4 to the 16nm class. The G5 process scales down the Active Area (AA) half-pitch—a key scaling metric—to 11.95 nanometers.
"CXMT originally targeted the 14nm class for G5 on its internal roadmap, but recent product disclosures indicate scaling down to 12–13nm," Choe analyzed. "Even without extreme ultraviolet (EUV) lithography equipment, scaling is achievable using deep ultraviolet (DUV) based double-patterning (DPT) or quadruple-patterning (QPT) techniques." He added, "While operational characteristics, cell size, and die size still require further analysis, if their announcements hold true, CXMT has skipped another one or two generations."
In NAND flash, China's technological pursuit is evaluated as even faster. YMTC commercialized its hybrid-bonding-based 'Xtacking' architecture early on, manufacturing the NAND cell array and peripheral circuits on separate wafers before bonding them. It also applies a 'Stair-Free' architecture to minimize peripheral area overhead in high-stack NAND.
"Setting aside economic viability or yield, purely from a technological standpoint, Chinese manufacturers are no longer followers—they are competitors," Choe emphasized. "As the technological gap narrows rapidly in both DRAM and NAND, the technical capabilities of Chinese firms warrant close attention."
Q&A
Q: It has been a while. We are joined today by Dr. Jeongdong Choe from TechInsights. Welcome.
A: Hello. It is great to be back.
Q: The memory market is experiencing a significant boom recently. While there is some debate regarding downstream markets, AI investments remain robust, providing much to discuss. Let's start with downstream markets. Historically, smartphones were the largest market for memory. Has that shift toward AI become absolute?
A: More than 50% of demand has shifted toward AI.
Q: Beyond memory, power semiconductors and other components are also pivoting toward AI and servers. How do you view the downstream market at this point? There are conflicting arguments—some suggest slowing down development speed, while others advocate managing risks while continuing aggressive development. These debates have directly impacted stock prices.
A: Market demand is surging dramatically, particularly in data centers and AI. Usage of services like ChatGPT and cloud infrastructure by both consumers and enterprises is growing exponentially day by day. Demand for memory has reached levels where availability is severely constrained, leading service providers to limit session times or increase pricing. Furthermore, as agentic AI, advanced inference, and physical AI emerge, the demand for data processing continues to escalate beyond expectations. Because data centers cannot fully satisfy demand, conventional memory inventories have been depleted. As a result, prices across DDR, LPDDR, and HBM applications have risen by over 80%. The reality today is that even operating production facilities at 100% capacity results in an undersupplied market. DRAM supply is short by approximately 10–12%, with NAND facing a similar deficit. This condition will persist for some time.
Q: So with demand at 100, supply sits at around 85%?
A: Supply currently meets only about 85% of demand. This shortage cannot be resolved in three to six months because capital expenditures (CAPEX) were severely constrained during the industry downturn over the past two to three years. For new facility investments to yield output, it will take at least another year and a half to two years. Consequently, many analysts project that elevated price levels will persist for roughly the next two years across data centers and general memory applications. Furthermore, HBM—which is heavily used in AI training—commanded premium pricing over the past year. Manufacturers like SK Hynix have reallocated conventional DRAM capacity toward HBM production, exacerbating shortages in standard DDR and LPDDR memory. In short, demand remains enormous while suppliers cannot fully meet it, a dynamic that will continue for the foreseeable future.
Q: Some research firms project that this supercycle could extend until 2030 or 2031.
A: Price dynamics and physical supply must be viewed separately. While prices may fluctuate, they will settle at significantly higher levels than before, sustaining a favorable environment for the next 3–4 years. Output from fabs currently under construction will hit the market in about a year, which may cause mild price pullbacks. However, because baseline pricing remains high, industry profitability should remain solid.
Q: What is your perspective on the demand side? Software model developers like Anthropic and OpenAI are progressing rapidly, leading some to advocate for a measured pace. Conversely, figures like Jensen Huang and policymakers argue that investments must continue while managing risks. How do you evaluate demand side dynamics?
A: I fully agree with Jensen Huang's perspective. While calls for pacing development originate from entities like OpenAI, those concerns stem largely from potential risks as AI transitions into agentic and physical forms capable of autonomous decision-making. Those concerns should be addressed through regulatory and institutional frameworks. Human demand for technological advancement and convenience has continuously expanded throughout history; it cannot be artificially halted. Just as society did not revert from the Iron Age to the Stone Age, technological progress will continue, supported by stronger governance, without needing to artificially slow down development.
Q: Critics point out that while AI data centers generate revenue, they require astronomical, ongoing capital investments, raising sustainability concerns. Furthermore, latecomers argue that calls for regulatory slowdowns by frontrunners amount to pulling up the ladder. Can the industry sustain this investment scale?
A: Stock valuations, private investors, and government initiatives globally—including in the US and Korea—are actively supporting AI infrastructure builds. Because AI infrastructure is recognized as a national priority worldwide, capital investment support will likely remain aligned with these expansion plans.
Q: Turning to the supply side, memory makers refrained from aggressive investments over the past three years but are now accelerating expansion. Domestic producers aim to double wafer input by 2030. How are capital expenditures shaping up across Micron, Samsung, SK Hynix, and Kioxia?
A: CAPEX intensity is exceptionally high, approaching 30–40% of revenue. During my recent visit to Micron's Boise site, construction of two massive fabs was progressing rapidly. Samsung and SK Hynix are making similar moves, such as SK Hynix's Yongin cluster and solid-state drive capacity expansion via Solidigm's Dalian fab, alongside potential US facilities. Kioxia is also expanding. Beyond Western and Korean firms, attention must be paid to China. Backed by government support for self-reliance, CXMT and YMTC are aggressively expanding fabs to boost domestic output. CXMT aims to raise its market share from 12–13% to 20–25%, planning multiple fabs across Shanghai and Beijing, while YMTC expands in Wuhan. Global CAPEX acceleration began roughly two years ago, meaning initial output from early facilities will begin impacting supply by the second half of next year. While later facilities will take 3–4 years, strong AI demand should absorb this new capacity without triggering structural oversupply in the near term. A potential supply-driven downturn might re-emerge in 5–6 years, but that remains a long-term risk not currently weighing on short-term outlooks.
Q: When do you expect the current supply-demand imbalance—where supply meets only 85% of demand—to ease toward 95%?
A: By the third or fourth quarter of next year, supply should reach 90–95% of demand.
Q: Is 90–95% considered market equilibrium?
A: A range of 90–95% typically represents normal operating conditions. However, due to accumulated backlogs, market tightness could extend for another quarter or two, keeping pricing resilient through next year.
Q: With continuous wafer capacity additions planned over the coming years, if demand is 100 and supply reaches 100, is that considered balanced?
A: Supply at 100% of demand is still technically tight; true equilibrium requires supply at 102–103%. However, the key question is whether suppliers can even reach 100% of demand in 2–3 years. AI memory demand, particularly for HBM, is not growing linearly but exponentially—surging by 4x to 6x in the short term compared to historical 1.2x to 1.4x annual growth. One or two new fabs per company will be insufficient to cover this demand shift.
Q: Long-term supply agreements (LTAs) appear to be increasingly common between memory suppliers and key customers. How do LTAs impact memory cycles?
A: LTAs have a positive impact. Historically, short-term quarterly contracting led to abrupt order cancellations or double-booking, creating extreme price volatility. Multi-year LTAs stabilize pricing, secure predictable supply for customers, and allow integrated device manufacturers (IDMs) to plan production schedules efficiently.
Q: Which customer tiers are securing these LTAs?
A: Major AI semiconductor designers and select hyperscalers are entering into these multi-year arrangements.
Q: Korean memory makers are projecting strong earnings performance this year and next. Do you expect operating margins to expand further next year?
A: Operating margins should maintain current levels or potentially tick slightly higher next year because supply shortages will persist. However, performance varies by product category—such as HBM, commodity DRAM, and enterprise SSDs. In HBM3E, SK Hynix holds over 70% market share, with Samsung at 12–15% and Micron at 10%. This distribution will likely hold through the end of this year and early next year. Even as HBM4 emerges in late 2026 or early 2027, HBM3E will remain the primary volume driver in the first half of next year, maintaining SK Hynix's advantage. In the second half of next year, as HBM4 adoption ramps, Samsung's market share could rise to 25–30%. Samsung is particularly targeting HBM4E with heavy R&D investments aimed at Nvidia and hyperscale clients. Thus, SK Hynix will maintain HBM leadership through next year, while competitive leadership dynamics may shift by 2028.
Q: What does holding HBM market leadership signify commercially?
A: Leadership grants pricing power and allocation control in an undersupplied market. While SK Hynix held that position over the past two years and will likely retain it through next year, Samsung could regain strategic leverage around 2028.
Q: How do you view Micron's current positioning in memory and HBM?
A: Micron benefits significantly from standard DDR5 and LPDDR5 price increases, similar to Samsung. However, its HBM volume remains relatively modest. Furthermore, SK Hynix has established a tight partnership with TSMC for base dies, while Samsung utilizes its internal foundry for advanced logic base dies. Micron faces a more complex position: while it manufactures base dies internally for HBM3E and select HBM4 products, it must ultimately rely on TSMC for advanced nodes. With TSMC foundry capacity tightly allocated, securing sufficient base die capacity presents an operational challenge for Micron.
Q: How do technical capabilities compare between Korean, US, and Chinese memory suppliers today?
A: In DRAM, the process gap between Korean and US producers has largely closed. In the late 10nm era, Micron trailed by nearly a generation, but at 13nm, 12nm, and 11nm nodes, cell sizes, feature sizes, and design rules across Samsung, SK Hynix, and Micron are practically identical—differing by less than 0.1 nanometers. All three major suppliers utilize DUV/EUV lithography and High-K Metal Gate (HKMG) technology starting from 1b/1β nodes.
This technical convergence allows latecomers to narrow the gap quickly. Notably, China's CXMT has implemented advanced features equivalent to the top three suppliers, excluding EUV equipment. While current volume production at CXMT centers on 1z (G3), its recently announced G5 process platform scales down to 12–13nm using DUV double-patterning (DPT) or quadruple-patterning (QPT) techniques without EUV, representing a jump of one to two process generations.
In NAND flash, YMTC leads in several architectural innovations. YMTC was the first to commercialize hybrid bonding via its 'Xtacking' architecture, a method now being adopted by Kioxia, Samsung, and SK Hynix. YMTC was also first to commercialize 2TB QLC NAND modules and apply 'Stair-Free' contact structures to reduce peripheral circuit overhead in ultra-high-stack NAND. From a purely technological perspective, YMTC operates as a direct competitor rather than a late follower.
Q: Could intense supply competition lead to another memory market chicken game in the future?
A: A traditional price war is unlikely because AI memory architectures are becoming highly stratified. Demand extends beyond HBM to enterprise SSDs (eSSD), CXL memory modules, GDDR, and high-density NAND solutions like HBF (High-Bandwidth Flash). Because AI workloads require diverse, specialized memory tiers across training, inference, and agentic systems, expanding market demand should absorb growing global supply.
Q: How do you view geopolitical pressures on memory makers to construct fabs in the US?
A: US policy encourages domestic semiconductor fab construction, placing pressure on global suppliers. However, overseas fab construction entails significantly higher capital costs, longer construction timelines—often double those in Korea—and substantial labor cost differentials. Even with government subsidies, offsetting these cost disparities remains challenging for manufacturers.
Regarding Solidigm's potential NAND investments in the US, its current Dalian fab faces physical expansion limits and restrictions on importing advanced equipment like cryogenic etchers. To expand production of its high-demand 192-layer and next-generation 240/250-layer QLC NAND for AI data centers, alternative manufacturing locations like the US or domestic Korean sites are being evaluated strategically.