随着人工智能数据中心建设持续升温，全球内存市场正面临日益严峻的供需失衡。韩国存储芯片巨头SK海力士在最新财报中表示，受人工智能基础设施投资增长、先进制程技术难度提升以及新建产能所需时间较长等因素影响，全球内存供应不足的局面预计将持续至2030年。这意味着，个人电脑、智能手机及其他消费电子产品所使用的内存，未来数年仍可能面临价格上涨和供应紧张的压力。
SK海力士于2026年10月28日公布第三季度财务业绩。公司当季营收达到104.6155万亿韩元，营业利润为83.3338万亿韩元，净利润为39.7756万亿韩元，首次实现单季度营收突破100万亿韩元。营业利润率达到80%，反映出人工智能相关存储芯片需求旺盛，以及高附加值产品价格上涨带来的强劲盈利能力。
公司表示，第三季度业绩增长主要得益于DDR5服务器内存、高带宽内存HBM4出货，以及321层TLC企业级固态硬盘的销售增长。随着人工智能模型规模不断扩大，数据中心不仅需要更多存储容量，还需要更高的数据传输带宽和更低的访问延迟，这使得内存成为AI基础设施建设中的关键硬件之一。
与传统个人电脑内存相比，HBM的制造工艺更加复杂，需要将多层存储芯片进行堆叠，并通过先进封装技术与计算芯片协同工作。随着英伟达等企业持续推出新一代AI加速器，HBM需求迅速增长，也使存储厂商不得不将更多生产资源投入利润率较高的产品。
SK海力士表示，当前人工智能投资正在形成一个持续扩大的循环：企业利用AI提高软件开发、客户服务和搜索等业务的效率，并从中获得新的收入，继而继续扩大人工智能应用和基础设施投入。与此同时，AI投资主体也不再局限于少数大型科技企业，越来越多普通企业和国家级项目开始参与其中，进一步扩大了数据中心建设需求。
公司还预计，未来内存需求将不再仅仅由大型语言模型和云计算数据中心推动。随着自动驾驶汽车、机器人等实体人工智能应用发展，设备需要处理越来越多来自传感器、摄像头和实时控制系统的数据，对内存容量、速度和可靠性的要求也将进一步提高。
然而，内存生产能力无法随着需求同步增长。先进DRAM制程的研发和量产难度不断提升，而建设一座新的半导体工厂，还涉及厂房建设、洁净室部署、生产设备安装、工艺调试以及良率爬坡等多个环节，往往需要数年时间。即使厂商决定大幅增加投资，新增产能也无法立即转化为市场供应。
基于这些因素，SK海力士预计，全球内存市场供不应求的结构性局面将持续至2030年。公司此前也曾警告，2027年可能成为内存供应最紧张的年份之一。即使未来数年不断扩大产能，新增供应仍可能难以完全满足AI基础设施持续扩张所带来的需求。
此次财报还显示，SK海力士正在加快扩大生产能力。公司计划在年底前将韩国M15X工厂的产能提升至最高水平，以增加HBM4的生产。公司预计，到2026年第四季度，HBM4的出货量和销售收入都将超过上一代HBM3E，使新一代产品逐步成为主流。
在传统DRAM领域，SK海力士将提高采用第六代10纳米级制程技术，也就是1c制程的产品占比，并扩大DDR5以及面向AI服务器的LPDDR5 SOCAMM产品销售。与此同时，公司计划提升321层NAND闪存的生产比例，以增加高性能、大容量企业级固态硬盘的供应。
这些投资虽然有助于缓解未来供应压力，但短期内仍难以彻底改变市场格局。SK海力士预计，2027年DRAM和NAND闪存需求仍将保持较快增长。由于供应能力受到工厂建设周期和先进制造工艺的限制，新增产能很可能继续被旺盛需求吸收。
为了降低需求波动带来的经营风险，SK海力士还在扩大与主要客户签订的长期供应协议。目前，公司已与14家主要客户签署长期协议，其中包括全球大型云服务提供商。按照公司披露的信息，已经签订协议所覆盖的数量，相当于2027年和2028年总产能的50%以上。
这些协议通常具有较长的合同期限，有助于客户提前锁定供应，也让SK海力士能够更准确地规划生产设备、工厂建设和资本支出。公司同时表示，正在与更多客户讨论延长合同期限的可能性，部分协议的常见期限约为五年。
不过，长期供应协议并不意味着所有消费市场客户都能获得稳定供货。大型云服务商和AI芯片企业拥有较强的采购能力，能够通过长期合同锁定大量产能；个人电脑厂商、手机厂商以及零售渠道则可能受到剩余供应量和市场价格波动的影响。
对于普通消费者而言，这种供需结构可能产生多重影响。首先，内存和闪存价格上涨会直接提高电脑、手机及固态硬盘的制造成本。其次，厂商可能通过减少基础配置、提高高容量版本售价，或者调整不同产品之间的价格差距来应对成本压力。部分厂商也可能推迟升级计划，或者优先将有限的存储芯片分配给利润较高的产品。
与此同时，服务器内存需求的持续增长，还可能改变传统内存产品的供给结构。当存储厂商将更多晶圆和先进封装资源用于HBM等高附加值产品时，传统DRAM产品的供应也可能受到间接影响。即使普通电脑内存与HBM并非完全相同的产品，它们仍然会受到生产资源、投资方向和整体半导体市场供需关系的共同影响。
需要注意的是，SK海力士预测供应短缺持续至2030年，并不等于所有内存产品的价格都将在未来四年持续上涨，也不意味着每一种产品都会始终处于缺货状态。不同类型内存的供需情况可能存在显著差异，厂商扩产进度、竞争对手的投资、客户采购计划以及AI行业自身的发展变化，都可能影响实际价格走势。
SK海力士此次发布的业绩表明，人工智能热潮已经深刻改变全球存储芯片产业的投资方向。过去，内存市场经常因个人电脑和智能手机需求变化而经历明显的周期性涨跌；如今，大型AI数据中心正在成为新的重要需求来源，而且相关基础设施投资具有较长的建设周期。
如果AI应用继续扩大，存储芯片厂商将有动力维持较高资本支出，并优先发展HBM、服务器DRAM和企业级固态硬盘等产品。但如果AI投资增速放缓，或者新增产能在未来集中释放，市场也可能重新出现供需变化。
目前，SK海力士的判断是，先进内存制造能力的扩张速度仍然赶不上需求增长。对于消费电子产业而言，这意味着内存可能不再只是整机成本中一个相对稳定的组成部分，而将成为未来几年影响产品定价、硬件配置和供应链规划的重要变量。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
As the construction of artificial smart data centres continues to warm, global memory markets are facing growing supply and demand imbalances. In its latest financial report, Korea ' s storage chip giant, SK Hercules, indicated that global memory supply shortfalls are expected to continue until 2030, influenced by increased investment in artificial intelligence infrastructure, increased technological difficulties in advanced process preparation and the longer time required to build new capacity. This means that the memory of personal computers, smartphones and other consumer electronics can still be subject to rising prices and tight supply pressures in the coming years.
SK Hercules issued financial performance for the third quarter on 28 October 2026. The company received 10,461.5 trillion won for the season, with an operating profit of 83,333.8 trillion won, a net profit of 39,775.6 trillion won, and, for the first time, a single quarterly profit of over 10 trillion won. The 80 per cent turnover rate reflects the strong demand for artificial intelligence-related storage chips and the strong profitability of higher prices for higher value-added products.
According to the company, the growth in performance in the third quarter was mainly due to the growth of the DDR5 server memory, the high-bandwidth memory of HBM4 deliveries and the sale of 321-storey TLC firm-level hard drives. With the growing size of artificial intelligence models, data centres need not only more storage capacity but also higher data transfer bandwidth and lower access delays, making memory one of the key hardware in the construction of AI infrastructure.
The HBM manufacturing process is more complex than traditional personal computer memory, requiring stacking of multiple layers of storage chips and working with computing chips through advanced containment techniques. With the continued introduction of a new generation of AI accelerators in enterprises such as Weida, the demand for HBM has grown rapidly and has forced storage manufacturers to invest more of their production resources in products with higher profitability.
According to SK Hercules, current investments in artificial intelligence are creating a cycle of continuous expansion: enterprises use AI to improve the efficiency of their operations, such as software development, customer services and search, and to derive new revenues from them, and continue to expand artificial intelligence applications and infrastructure inputs. At the same time, AI's portfolio is no longer limited to a small number of large technology enterprises, and a growing number of ordinary enterprises and national projects have begun to participate, further expanding the demand for data centres.
The company also expects that future memory requirements will no longer be driven solely by large language models and cloud computing data centres. With the development of artificial intelligence applications such as auto-driving cars and robots, the equipment will need to process an increasing number of data from sensors, cameras and real-time control systems, and the requirements for memory capacity, speed and reliability will be further increased.
However, memory production capacity cannot grow simultaneously with demand. The development and volume of advanced DRAM processes is becoming more difficult, while the construction of a new semiconductor plant involves the construction of a plant, the deployment of clean houses, the installation of production equipment, the commissioning of processes and the upgrading of the slopes, often taking several years. Even when producers decide to invest significantly, new capacity cannot be immediately converted into market supply.
On the basis of these factors, SK Hercules predicts that the structural situation in the global memory market will last until 2030. The company had previously warned that 2027 might be one of the most stressful years of memory supply. Even if production capacity continues to expand in the coming years, it may be difficult to fully meet the demands of the continued expansion of AI infrastructure.
The financial report also shows that SK Hercules is accelerating the expansion of production capacity. By the end of the year, the company plans to increase the production of HBM4 by raising the capacity of the Korea M15X plant to its highest level. By the fourth quarter of 2026, the company expects HBM4 to outnumber the previous generation of HBM3E and gradually mainstream the new generation of products.
In the area of traditional DRAM, SK Hercules will increase the proportion of products using the sixth generation of 10 nanoscale technology, i.e. 1c, and expand the distribution of DDR5 and LPDDR5 SOCAMM products for the server at AI. At the same time, the company plans to increase the production ratio of the 321-storey NAND flash drive to increase the supply of high-performance, high-capacity firm-level solid disks.
While these investments have helped to ease future supply pressures, it remains difficult to radically change market patterns in the short term. SK Hercules expects that the flash demand for DRAM and NAND will continue to grow relatively rapidly in 2027. Since supply capacity is limited by plant construction cycles and advanced manufacturing processes, it is likely that new capacity will continue to be absorbed by robust demand.
SK Hercules is also expanding long-term supply agreements with key clients in order to reduce the operational risks associated with fluctuations in demand. Currently, companies have signed long-term agreements with 14 major clients, including large global cloud service providers. According to information disclosed by the company, the amount covered by the agreement is equivalent to more than 50 per cent of total capacity in 2027 and 2028.
These agreements usually have a longer duration of contracts, which helps clients to lock up supplies in advance and allows SK Hercules to plan production equipment, plant construction and capital expenditure more accurately. At the same time, the company indicated that discussions were under way with more clients on the possibility of extending the contract term, with some agreements having a common duration of about five years.
Long-term supply agreements, however, do not mean that all consumer market customers have stable supply. Large cloud service providers and AI chip companies have a strong purchasing capacity to lock in large volumes of capacity through long-term contracts; personal computer manufacturers, mobile phone manufacturers and retail outlets may be affected by remaining supply and market price fluctuations.
This supply and demand structure can have multiple effects for ordinary consumers. First, higher memory and flashback prices directly increase the manufacturing costs of computers, mobile phones and solid hard drives. Second, manufacturers may respond to cost pressures by reducing base allocation, increasing the price of high-volume versions, or adjusting the price gap between different products. Some manufacturers may also defer upgrading plans or prioritize the allocation of limited storage chips to more profitable products.
At the same time, the continued growth of server memory demand may also alter the supply structure of traditional memory products. The supply of traditional DRAM products may also be indirectly affected when storage manufacturers devote more of their resources to higher-value-added products such as HBM. Even though general computer memory is not exactly the same as HBM products, they are still subject to the combined effects of production resources, investment orientation and overall semiconductor market supply and demand.
It should be noted that the SK Hercules forecast that the supply shortfall will continue until 2030 does not mean that the price of all memory products will continue to rise over the next four years or that each product will remain under stock. There may be significant differences between supply and demand for different types of memory, and the pace of plant expansion, the investment of competitors, customer procurement plans and the evolution of the AI industry itself may affect real price trends.
The performance of publish by SK Hercules this time shows that the energy boom of artificial intelligence has profoundly changed the direction of investment in the global storage chip industry. In the past, the memory market had often experienced a marked cyclical rise and decline due to changes in the demand for PCs and smartphones; today, large AI data centres are becoming an important new source of demand, and related infrastructure investments have a longer construction cycle.
If AI applications continue to expand, storage chip manufacturers will have the incentive to maintain higher capital expenditures and prioritize the development of products such as HBM, server DRAM and enterprise-level solid disks. But if investment in AI slows down, or if new capacity is released centrally in the future, there may be a re-emergence of changes in supply and demand.
At present, SK Hercules judges that the expansion of advanced memory manufacturing capacity continues to lag behind demand growth. For the consumer electronics industry, this means that memory may no longer be a relatively stable component of the overall cost of the machine, but will be an important variable affecting product pricing, hardware configuration and supply chain planning in the coming years.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.