
摩根士丹利報告:代理式AI涵蓋約30兆美元消費支出,AI基礎設施電力容量預估2028年達145GW
代理式AI正從回答問題進化為任務執行,Meta旗下Muse正與Walmart、Best Buy及Expedia等企業合作,整合商品搜尋、比較與旅遊預訂服務。摩根士丹利估計代理式AI潛在應用涵蓋約30兆美元消費支出,AI機櫃記憶體成本比重從Blackwell不到1成提升至Vera Rubin約25%,大型雲端與新興AI雲端業者設施容量預估由2025年約36GW增至2028年約145GW。
AI-generated summary
人工智慧正從單純回答問題進化為任務執行,代理式AI可依使用者目標規劃步驟、調用工具,並在授權範圍內完成工作。Meta旗下Muse正拓展與Walmart、Best Buy及Expedia等企業合作,整合商品搜尋、比較與旅遊預訂服務。摩根士丹利報告估計代理式AI潛在應用涵蓋約30兆美元消費支出,AI基礎設施電力容量預估由2025年約36GW增至2028年約145GW。
從回答問題到代辦任務:AI應用正在融入日常生活
人工智慧(AI)正在從單純的回答問題,進一步走向任務執行。代理式AI(Agentic AI)可依使用者目標規劃步驟、調用工具,並在授權範圍內完成工作。根據摩根士丹利近期報告指出,Meta旗下Muse正拓展與Walmart、Best Buy及Expedia等企業合作,整合商品搜尋、比較與旅遊預訂服務。使用者未來可望透過AI完成原本需要切換多個網站的流程,而服務整合程度與任務執行能力,也將成為電子平台是否能吸引並留住用戶的重要因素。
消費服務加速整合 運算需求延伸至基礎設施
依據一份針對Muse的研究報告估算,代理式AI的潛在應用涵蓋約30兆美元的消費支出。消費者若透過AI比較商品、挑選飯店並完成預訂,零售商與旅遊業者可能需要串接平台,讓商品與房源納入推薦及預訂流程,部分行銷預算也可能隨之移轉。不過,商品供應、配送與住宿仍由合作業者提供,因此,即使消費者改透過AI下單,業者的價格競爭力、供貨能力與服務品質,仍是能否取得訂單的關鍵。
隨著零售與旅遊服務深度串接,AI 的多工具調用與端到端任務執行力,將重塑平台獲客與用戶留存關鍵。(路透)
由於這類服務需要AI反覆推理、調用工具,並保留前後任務內容,運算負荷因而高於單次問答,也提高對處理器與記憶體的需求。摩根士丹利發布的《Mapping AI’s Rate of Change: From Buildout to Broadening》模型估計,記憶體占AI機櫃總成本的比重,在前一代Blackwell機櫃中不到1成,在新一代Vera Rubin機櫃中則約為25%。綜合上述運算需求與成本考量,記憶體容量、存取效率及成本控制,均是系統設計的重要因素。此外,運算需求擴大,也促使雲端業者擴建資料中心,並提高對電力供應及相關設施的需求。同份報告預估,大型雲端與新興AI雲端業者將持續擴充算力供給,其設施容量以電力容量衡量,將由2025年約36GW增至2028年約145GW,約達原來的四倍。此為整體AI基礎設施的擴建預估,實際進度仍取決於電力供應、機房建置與設備交付等因素。
從產業趨勢到資產配置 理解報酬與風險來源
代理式AI帶起跨產業之需求,卻不必然為每個產業帶來相同幅度的投資報酬。平台須將使用量轉化為收入,設備商受採購進度影響,基礎設施業者則須承擔建置與融資成本。隨著AI相關產業之技術快速演進,也可能改變既有競爭優勢,使需求成長未必同步反映在企業獲利與股價表現。
因此,資產配置須同時考量成長潛力、收益來源與風險承受能力。分散至不同產業與資產,可降低對單一企業或題材的依賴,但分散效果仍取決於資產間的連動程度。隨著AI商業化進度、企業評價及利率環境改變,投資人可定期檢視並適時調整股債比重與產業配置,有助於分散風險,並保留參與產業發展的彈性。
(作者為元大證券國際金融業務部專業襄理)
AI outlook — possibilities, not facts
大型雲端與新興AI雲端業者設施容量將由2025年約36GW增至2028年約145GW
Likely · Within years

Huang Renxun, CEO of Huida, was involved in a relationship dispute with Huang Le, a woman suspected of assisting Chinese espionage activities. His eldest daughter Huang Minshan was originally scheduled to visit South Korea on the 8th and meet with South Korean Minister of Science, Technology, Information and Communications Bae Kyung-hun, but the trip was suddenly canceled and the meeting was cancelled. South Korean officials did not specify the reason, but sources said it may be due to family reasons. Huida AI Day Seoul 2026 will be held as usual.

The 2026 Information Month and Taiwan Education Technology Exhibition will be held at Taipei World Trade Center Hall 1 from November 12 to 15. The Ministry of Interior has set up a national base map area in the Smart Nation Pavilion and planned six scenario training bases to allow the public to understand the application value of spatial information in areas such as land planning, disaster prevention, transportation and conservation through interactive experiences.

The new generation communication network includes 5G, 5G-A, 6G, 10G optical network and satellite Internet, realizing the integration of communication, perception, computing and intelligence. At present, my country has built more than 26,000 industrial 5G private networks and 1,260 5G factories. The industrial Internet covers all 41 industrial categories. Nanjing Purple Mountain Laboratory is conducting 6G multi-base station collaborative sensing testing.

Residents in Batam, Indonesia are protesting data centres amid a water shortage caused by drought and El Niño, as the facilities use water to cool servers powering AI tools.

Google will automatically convert Gem to "skills" and enter read-only status starting from November 17, 2026. It will be officially removed on December 17. Some files may not be completely retained. Users are recommended to back up as soon as possible. Workspace Business, Enterprise and Nonprofit accounts will transition in March 2027, and Education Edition accounts will transition in June 2027.

Technology giants such as Anthropic, Meta, Microsoft, OpenAI, and Tencent have intensively launched personal intelligence products. AI is evolving from a chat tool to a work assistant that can actively handle tasks such as email, payment, and scheduling. However, issues such as privacy leaks, fraud risks and platform blocking are still obstacles that need to be overcome before these AI butlers can enter daily life.