Author: Climber,CryptoPulse Labs
近日,美光科技宣佈其位於日本廣島的先進記憶體晶片工廠正式舉行動工儀式,總投資額約1.5兆日圓(約93億美元)。該廠將集中生產HBM(High Bandwidth Memory)等關鍵AI晶片,預計2028年左右開始大量出貨。
與此同時,包括美國、日本、新加坡、台灣(中國)在內的多地區新工廠也在同步推進,涵蓋DRAM、HBM、NAND Flash等核心產品。
這不是一場普通的產能擴張,而是對未來十年AI基礎設施話語權的爭奪戰。未來幾年,誰擁有更先進的記憶體產能,誰就可能掌握AI產業鏈中最核心的話語權。
一、 從消費電子到AI算力:記憶體市場迎來結構性變革
在過去幾十年裡,儲存行業一直是半導體中週期性波動最劇烈的領域之一。DRAM和NAND產品高度標準化,行業長期遵循需求上升、產能擴張、供應過剩、價格暴跌、減產的經典週期。
因此,美光過去很少進行如此激進的大規模資本支出,更多是根據市場狀況調節產能。然而,AI的出現徹底改變了這一邏輯。
過去,無論是個人電腦還是智慧型手機,對記憶體容量的需求增長都相對緩慢,設備增加幾個GB的記憶體就足以滿足升級需求。然而,大模型時代的資料處理方式完全不同。
以GPT、Gemini、Claude等大語言模型為代表的新一代AI,需要數萬甚至數十萬張GPU進行協同訓練,而每一張GPU都需要大量HBM作為高速快取。如果沒有足夠的HBM,GPU算力再強也無法充分發揮效能。
正是因為如此,HBM已逐漸成為整個AI伺服器生態中最稀缺的零組件之一。
業界數據顯示,一台高階AI GPU所需的HBM價值已接近甚至超過一顆傳統伺服器CPU的價格。隨著GPU持續升級,每一代產品所需的HBM容量也不斷增加,推動HBM市場進入高速成長階段。
不僅如此,AI還帶動了傳統DRAM和NAND需求的同步提升。
AI PC、AI手機、自動駕駛、邊緣AI伺服器、智慧機器人等新應用不斷湧現,推動整個儲存市場進入結構升級週期。美光預計,由AI驅動的供需緊張局面至少會持續到2026年以後,而真正的大規模新增產能要到2027至2028年左右才會釋放出來。
這意味著在未來兩到三年內,全球高階儲存市場仍將處於供不應求的狀態。正是基於此,美光選擇提前佈局,希望在下一輪行業週期中搶占更多市占率,而非等到需求爆發後再倉促擴產。
從商業邏輯視角來看,這更像是為AI時代提前修建高速公路,而不是簡單增加幾條生產線。
二、重塑全球製造業格局:美光打造AI時代供應鏈
仔細看這一輪產能擴張的佈局,一個非常鮮明的特徵是,美光幾乎對其全球所有主要製造基地都進行了重新規劃,美國是這一輪擴張的核心。
在維吉尼亞州馬納薩斯的工廠,美光已實現1α(1-alpha)奈米製程的量產,並透過一項20億美元的擴建項目,將DDR4晶圓供應能力翻了兩番,側重於汽車電子、工業控制和國防市場。
更重要的是,該公司正在美國境內重新建立先進DRAM製造能力。
位於愛達荷州博伊西的新工廠,投資規模高達500億美元,是近年美國最大的半導體製造項目之一。同時,美光宣布將其整體在美投資擴大至約2,000億美元,包括先進製造和研發體系構建,並增設第二座晶圓廠。
此外,紐約州的Clay巨型晶圓廠項目,投資金額更是超過1,000億美元,未來將興建多座晶圓廠,目標在2030年左右形成全球最大的DRAM製造基地之一。
在這些投資背後,一個關鍵詞是供應鏈本地化。
近年來,美國持續推動晶片製造環節回流,以降低對亞洲供應鏈的依賴。《晶片與科學法案》的出台,為包括美光在內的企業提供了大量財政補貼和稅收優惠,也壓低了企業赴美建廠的成本壓力。
與此同時,美光並未放棄亞洲。
日本的廣島工廠,未來將重點生產HBM,被認為是本輪擴建中戰略價值最高的項目之一。日本擁有成熟的半導體材料、設備和封裝產業鏈,加上政府的補貼扶持,使其再次成為先進存儲製造的重要基地。
而新加坡,則將承擔先進NAND製造任務,其擁有完善的電子製造生態、穩定的政策環境和國際物流優勢,適合承接全球供應中心的角色。
收購位於中國台灣銅鑼的力積電(PSMC)晶圓廠,將幫助美光快速擴充DRAM製造產能,並縮短新建晶圓廠的建設週期。
可以看出,美光的全球佈局,並非簡單地追求產能最大化,而是要在不同區域部署不同產品,實現風險分散、供應鏈安全和成本優化。
這種佈局模式,正日益成為全球半導體企業的新常態。
三、AI時代,存儲行業的最大贏家會是美光嗎?
資本市場最關心的問題,其實只有一個:如此巨大的投入,未來能否收回?答案很大程度上取決於AI需求能否持續;目前看,偏向樂觀。
Global cloud computing vendors are continuing to expand their AI capital expenditure. Tech giants like Microsoft, Google, Meta, and Amazon are investing tens of billions of dollars annually in building AI data centers.
At the same time, an increasing number of traditional enterprises are also beginning to deploy AI computing infrastructure.
As new applications such as AI Agents, multimodal models, video generation, autonomous driving, and humanoid robots continue to be implemented, the demand for high-performance storage is still growing rapidly.
Compared to the GPU market, which is mainly dominated by a few manufacturers, the HBM market is still in a phase of rapid expansion.
In the past, the high-end HBM market was long dominated by SK Hynix, with Samsung Electronics following closely, while Micron entered relatively late. However, in recent years, Micron has successfully entered the mainstream AI supply chain with products like HBM3E and obtained certifications from key customers, and its market share is rapidly increasing.
In the coming years, HBM may become one of Micron's fastest-growing profit businesses.
Of course, risks still exist.
The biggest characteristic of the semiconductor industry is cyclical fluctuations. If the pace of AI investment slows down in the future, or if new global capacity is released intensively, DRAM and NAND prices could still re-enter a downward cycle.
Furthermore, Samsung Electronics and SK Hynix are also continuously expanding capacity, making future market competition even fiercer. Advanced process technology, packaging capabilities, yield control, and customer certification will all determine who can truly win orders in the AI era.
However, unlike the past reliance on consumer electronics, AI demand is now changing the business model of the entire storage industry.
More and more high-end storage products are adopting long-term supply agreements, forming more stable cooperative relationships with cloud computing vendors and AI chip companies. This means that future industry price fluctuations may be somewhat moderated compared to the past, and corporate profitability may also become more stable.
For Micron, this round of capacity expansion is not just about increasing capacity, but also about redefining its role in the global semiconductor industry — transforming from a traditional memory chip manufacturer into a key supplier of AI infrastructure.
結論
GPUs determine the computing power of AI models, while HBM, DRAM, and NAND determine whether data can be quickly stored, accessed, and transferred. Without high-performance storage, even the most powerful AI models would struggle to reach their full potential.
Therefore, Micron's global expansion spanning the United States, Japan, Singapore, and Taiwan is not merely a manufacturing investment plan, but a strategic layout for AI infrastructure over the next decade.
It is foreseeable that in the coming years, the global storage industry will usher in a new wave of capital investment, technological upgrades, and market competition. Micron's global expansion plan, launched this time, may only be the beginning of this AI storage war.

