EE Times原文发布 09-18 20:00Radar 收录 09-18 20:01查看原文 ↗

AI需求将使DRAM市场持续承压关键事件

AI Demand Will Keep DRAM Market Under Pressure

中文摘要

AI基础设施支出正推动DRAM短缺，该短缺预计将持续至2027年，并使消费电子厂商在优先级排序中进一步靠后。

原文深度解读

本文讨论AI基础设施支出推动DRAM市场持续短缺，涉及TechInsights的Mike Howard、Circular Technology的Brad Gastwirth等分析师，以及Dell'Oro Group和Omdia的报告。文章指出DRAM短缺可能持续至2027年甚至更久，AI相关应用可能占DRAM总收入的75%，内存约占产品物料清单的30%，消费电子厂商在供应优先级中靠后。与电子元器件供应链的关系是：DRAM作为AI硬件栈中的关键半导体组件，其供应紧张和产能分配将影响消费电子等下游行业的元器件获取和成本。

TechInsights的Mike Howard认为当前DRAM市场是数十年来供应最不足的时期，供应限制预计将持续到2027年，因为需求增长超过行业增加制造产能的能力。
Circular Technology的Brad Gastwirth表示明年产能似乎已完全被预定，未来18个月内看不到供应过剩的情景。
与以往周期不同，内存制造商在支出上更加谨慎和有纪律，通过长期协议和客户资助的扩张计划来降低风险，而非过度建设产能。
Howard估计AI相关应用可能在几年内占DRAM总收入的75%；Gastwirth指出内存可占产品物料清单的约30%，短缺对消费设备制造商尤其痛苦。
Dell'Oro Group报告显示2026年第二季度全球数据中心IT半导体和组件收入激增182%，DRAM和存储收入以三位数速度增长，每比特平均售价同比翻倍以上。
Omdia将2026年半导体收入预测上调至同比增长94.1%，预计内存IC将占2026年半导体总收入的50%以上，瓶颈在HBM、先进封装和节点产能，预计至少持续到2027年。

供应链影响

DRAM供应紧张可能持续至2027年，取决于AI基础设施投资是否维持当前增速以及产能扩张的实际落地情况。
消费电子制造商可能因内存占物料清单约30%而面临成本上升和产品数量减少，取决于其能否获得足够的DRAM分配。
AI相关应用可能占据DRAM总收入的75%，可能导致非AI应用在供应优先级中进一步靠后，取决于供应商的长期协议覆盖比例。
HBM、先进封装和节点产能瓶颈可能限制整体内存供应，取决于这些环节的扩产进度和技术突破。
机器人技术和物理AI可能成为下一个内存需求驱动因素，可能进一步延长当前内存周期，取决于其商业化落地时间。

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。

正文内容 · AI翻译

中文英文

AI基础设施支出的看似无止境的激增将继续造成DRAM供应短缺，行业分析师认为这种短缺可能持续到2027年甚至更久。

Mike Howard

最新一轮周期与以往的DRAM繁荣与萧条显著不同，当时内存需求主要与PC和智能手机相关。AI已使内存成为数据中心基础设施的关键组件，导致供应约束更加严重和持久。

“这是数十年来DRAM市场供应最不足的时期，”TechInsights的DRAM和内存市场总监Mike Howard在接受EE Times采访时表示，并补充说随着AI部署加速，定价和收入增长已经激增。他预计供应约束将持续到2027年，因为需求增长超过了行业增加制造产能的能力。

Circular Technology的全球研究 and 市场情报主管Brad Gastwirth持有类似观点。“我看不出市场如何在短期内真正放松，”他在接受EE Times采访时表示。“每个人都知道明年到来的产能似乎已完全被预定，所以我看不到未来18个月内会出现供应过剩的情景。”

由上海永铭电子有限公司（YMIN）2026年9月17日

由RAMXEED 2026年9月16日

由Analog Devices 2026年9月15日

尽管需求高涨，内存制造商仍在保持纪律。他们没有像以往周期那样过度建设产能——最终导致严重供应过剩和价格崩溃——而是采取更为审慎的方式。“由于一轮又一轮周期发生的事情，他们在花钱方面非常有条不紊和非常谨慎，”Gastwirth说。

他指出长期协议和客户资助的扩张计划是生产商以比过去小得多的风险增加产能的证据。

Howard也看到结构性变化降低了经典内存崩溃的可能性。长期供应协议现在覆盖了行业产出的很大一部分，帮助供应商更好地将生产与需求对齐，避免大量库存积累。

仍有待观察的是，当前的AI需求能否在长期内保持足够强劲，以证明持续扩张的合理性，以及AI泡沫是否濒临破裂。

Howard和Gastwirth都预计长期轨迹将保持其方向，无论任何挫折。“确实存在对AI泡沫破裂的担忧，而且这种担忧是有道理的，”Howard说。然而，他说，AI的经济前景和对越来越强大的计算系统的需求支持持续投资。

Gastwirth表示，鉴于仅LLM已经带来的所有好处，他不认为会出现AI泡沫破裂。他承认围绕工作替代和采用时间表等问题存在不确定性，但他相信AI的影响已经太大而无法逆转。

除了训练和推理工作负载，Gastwirth将机器人和物理AI视为内存需求的下一个主要驱动力。“当机器人真正开始加速时，你需要为机器人复制和三重系统才能真正工作。”他说，所有传感器收集的信息量将需要指数级更多的DRAM。

Gastwirth表示，物理AI带来的数据增长可能创造新的需求流，将当前内存周期延长到远远超出今天的数据中心建设。

“内存已成为AI硬件栈中最重要的组件之一，往往是门控半导体组件，”Howard说，他认为内存现在是系统部署的瓶颈，而不是支持性元素。

后果已经在AI市场之外感受到。由于超大规模客户愿意为高带宽内存（HBM）和先进DRAM支付溢价，消费电子制造商发现自己在供应商的优先级列表中排名较低。

Gastwirth表示，他预计更高的组件成本将转化为更少的消费产品和更高的价格，因为内存可占产品物料清单的约30%，使短缺对消费设备制造商尤其痛苦。

Brad Gastwirth

两位分析师都同意，只要供应保持紧张，AI客户将继续获得优惠待遇。Howard估计，AI相关应用可能在几年内占DRAM总收入的75%。

Howard和Gastwirth并不是唯一持相同观点的分析师。

根据Dell'Oro Group最近发布的一份报告，2026年第二季度全球数据中心IT半导体和组件收入（用于服务器和存储系统）激增182%，DRAM和存储收入以三位数速度增长，每比特平均售价同比翻倍以上。加速的服务器部署也推动了对互补组件的强劲需求，包括HBM、CPU、内存和高速后端NIC。

Omdia也将2026年半导体收入预测上调至同比增长94.1%，原因是AI基础设施投资继续超过全球供应，对DRAM的需求空前，内存IC现在预计将占2026年半导体总收入的50%以上。

该研究公司表示，AI需求已超过行业当前生产和封装芯片的能力，HBM、先进封装和节点产能的瓶颈预计至少持续到2027年。

另见：

纪律将保持内存市场紧张

两个事件说明AI如何改写内存规则

Majestic Labs为内存池化AI服务器筹集1亿美元

DRAM无法跟上AI需求

AI和大数据、DRAM、高带宽内存、内存、半导体

CIRCULAR TECHNOLOGY、TECHINSIGHTS

Gary Hilson

Gary Hilson是一名自由撰稿人和编辑，为北美各地的印刷和像素出版物撰写了数千字。他的兴趣领域包括软件、企业和网络技术、研究和教育、可持续交通以及社区新闻。他的文章已由Network Computing、InformationWeek、Computing Canada、Computer Dealer News、Toronto Business Times、Strategy Magazine和Ottawa Citizen发表。

在LinkedIn上关注Gary

The seemingly unending surge in AI infrastructure spending will continue to create DRAM supply shortages, which industry analysts believe could persist well into 2027 and potentially beyond.

Mike Howard

This latest cycle differs significantly from previous DRAM booms and busts, when memory demand was tied largely to PCs and smartphones. AI has made memory a critical component of data center infrastructure, leading to more acute and long-lasting supply constraints.

“This is the most undersupplied the DRAM market has been in decades,” Mike Howard, director of DRAM and memory markets at TechInsights, said in an interview with EE Times, adding that pricing and revenue growth have surged as AI deployments accelerate. He expects supply constraints to continue through 2027 because demand growth is exceeding the industry’s ability to add manufacturing capacity.

Brad Gastwirth, global head of research and market intelligence at Circular Technology, shares a similar view. “I don’t see how markets really loosen up anytime soon,” he said in an interview with EE Times. “Everybody knows that the capacity that comes in next year seems to be fully committed, so I don’t see a scenario where you have excess supply in the next 18 months.”

By Shanghai Yongming Electronic Co., Ltd. (YMIN) 09.17.2026

By RAMXEED 09.16.2026

By Analog Devices 09.15.2026

Despite the high demand, memory makers are exercising discipline. Rather than overbuilding capacity as they have in previous cycles—which ultimately led to significant oversupply and collapsing prices—they are taking a more measured approach. “They’ve been very methodical and very careful to spend money because of what’s happened cycle after cycle,” Gastwirth said.

He pointed to long-term agreements and customer-funded expansion plans as evidence that producers are growing capacity with considerably less risk than in the past.

Howard also sees structural changes reducing the likelihood of a classic memory crash. Long-term supply agreements now cover a significant portion of industry output, helping suppliers better align production with demand and avoid large inventory accumulations.

What remains to be seen is whether the current AI demand can remain strong enough over the longer term to justify continued expansion and whether the AI bubble is on the verge of popping.

Both Howard and Gastwirth expect the long-term trajectory to maintain its course, regardless of any setbacks. “There is real concern about an AI bust, and it is warranted,” Howard said. However, the economic promise of AI and the need for increasingly capable computing systems support ongoing investment, he said.

Gastwirth said he doesn’t believe there will be an AI bust given all the benefits that have already been seen from LLMs alone. He acknowledged uncertainty around issues such as job displacement and adoption timelines, but he believes AI’s impact is already too significant to reverse.

Beyond training and inference workloads, Gastwirth sees robotics and physical AI as the next major driver of memory demand. “When you get the robotics really starting to pick up, you need duplicate and triplicate systems for robots to really work.” He said the amount of information that’s gathered with all the sensors will require exponentially more DRAM.

Gastwirth said the resulting data growth from physical AI could create new demand streams that extend the current memory cycle well beyond today’s data center buildouts.

“Memory has become one of the most important components in the AI hardware stack, often the gating semiconductor component,” Howard said, arguing that memory now serves as a bottleneck for system deployment rather than a supporting element.

The consequences are already being felt outside the AI market. With hyperscale customers willing to pay premium prices for high-bandwidth memory (HBM) and advanced DRAM, consumer electronics makers find themselves lower on suppliers’ priority lists.

Gastwirth said he expects higher component costs to translate into fewer consumer products and increased prices as memory can account for roughly 30% of a product’s bill of materials, making shortages particularly painful for consumer device makers.

Brad Gastwirth

Both analysts agree that AI customers will continue receiving preferential treatment as long as supply remains constrained. Howard estimates that AI-related applications could account for as much as 75% of total DRAM revenue within a few years.

Howard and Gastwirth aren’t the only analysts singing from the same songbook.

According to a recently published report by Dell’Oro Group, worldwide data center IT semiconductor and component revenue for servers and storage systems surged 182% in the second quarter of 2026, with DRAM and storage revenue growing at a triple-digit pace, as average selling prices per bit more than doubled year over year. Accelerated server deployments also drove strong demand for complementary components, including HBM, CPUs, memory, and high-speed back-end NICs.

Omdia has also raised its 2026 semiconductor revenue forecast to 94.1% year-over-year growth, driven by unprecedented demand for DRAM as AI infrastructure investment continues to outpace global supply, with memory ICs now expected to account for more than 50% of total semiconductor revenue in 2026.

The research firm said AI demand has exceeded the industry’s current ability to produce and package chips, with bottlenecks across HBM, advanced packaging, and node capacity expected to persist until at least 2027.