AiX 環球 AI 情報 Global AI Intelligence

第 27 期 2026 年 10 月 24 日（星期六） 07:30 出刊 10 條情報 · 16 原文來源連結 RSS 過往期數

今日十條 · 2026 年 10 月 24 日同一天，人手與算力同時被重新安排：國務院把就業寫進人工智能推廣條件，四家企業的設備則直接搬進已通電的場地

今日主編判斷

今日兩條主線在同一天落地。國辦印發就業友好型意見，把崗位替代風險評估、穩慎推廣節奏與失業緩衝期寫成推廣條件。另一邊，英偉達把 AI 工廠裝進集裝箱、部署期由六個月壓至最快十四天，美國商務部再向功率半導體撥款最高三點八九億美元。算力照舊搶，人手不再是可以事後頭痛的變數。Two threads landed in the same 24 hours. Beijing's State Council issued an employment-friendly AI directive that turns displacement-risk assessment, a go-slow on labour-intensive sectors and voluntary redundancy buffers into conditions of rollout. Elsewhere, Nvidia packed an AI factory into containers and cut site deployment from six months to as little as fourteen days, while the U.S. Commerce Department put up to US$389 million behind power semiconductors. Compute is still scarce; headcount is no longer an afterthought.

今日十條

本欄由 AIX Society 編輯部整理公開資料後撰稿，每條均附原始來源連結；事實與觀點分列，觀點屬本會判斷。

按受眾篩選

國務院辦公廳印發就業友好型意見，六方面十八項任務把崗位影響寫入人工智能推廣條件China's State Council issues employment-friendly AI directive with 18 tasks tying rollout to job impact

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners人力資源HR法務合規Legal & compliance

事件: 十月二十三日，國務院辦公廳印發《關於適應人工智能技術發展助力構建就業友好型發展方式的意見》，部署六方面十八項任務，包括建立人工智能應用就業風險評估機制、穩慎把握勞動密集型行業應用節奏、鼓勵企業設失業緩衝期。On 23 October, China's State Council General Office issued an opinion on building an employment-friendly development model amid AI adoption, setting 18 tasks across six areas. These include a jobs risk-assessment mechanism for AI deployment, a cautious pace for labour-intensive sectors, and encouragement for firms to set redundancy buffer periods.
背後: 這是九月《人工智能+》意見之後，首次由國務院層面把就業影響列為技術推廣的前置條件，而非事後補救。政策取向是把人工智能視為就業結構的改造者，用評估與轉崗把衝擊分散到時間軸上。Coming after September's AI Plus opinion, this is the first State Council-level text to treat job impact as a precondition of deployment rather than an after-the-fact remedy. The intent is to treat AI as a restructurer of employment and spread the shock across time through assessment and redeployment.

顧問觀點

我們認為，企業過去習慣以效率為單一理由推進自動化，此後在同一份立項文件裏要同時交代崗位去留，一把手簽署的風險由技術部門轉回經營層。In our view, firms used efficiency alone to justify automation. From now on the same project paper has to answer for the jobs affected, moving risk from the IT function back to the executive who signs.

對日常工作的影響: 人力資源部需要為每個自動化項目補一份崗位影響說明；廠務與客服主管則要交出推廣節奏的建議，而不只是效益測算。HR must attach a job-impact note to every automation project, while plant and service managers now submit rollout pace recommendations alongside benefit forecasts.
如何改善: 本周挑選一個已排期的自動化項目，補寫受影響崗位、人數與轉崗去向三欄，作為往後所有立項文件的模板。This week, pick one scheduled automation project and add three columns on affected roles, headcount and redeployment, then use it as the template for all future project papers.

來源人民日報（2026-10-24）↗證券時報（2026-10-24）↗

《國家信息化「十五五」規劃》印發，到二零三零年適時啟動 6G 商用並加快構建全國一體化算力網China publishes 15th Five-Year informatisation plan, targeting 6G launch by 2030 and a nationwide compute network

影響力✓ 已核實Verified3 年3 years企業決策者Business owners資訊科技IT

事件: 十月二十三日，中央網絡安全和信息化委員會印發《國家信息化「十五五」規劃》，提出到二零三零年適時啟動第六代移動通信商用，部署十方面任務，並明確加快構建全國一體化算力網、研製高性能人工智能芯片與高可用基礎軟件棧。China's Central Cyberspace Affairs Commission issued the 15th Five-Year Plan for national informatisation on 23 October. It targets a timely commercial launch of 6G by 2030, sets ten task areas, and calls for a nationwide integrated compute network plus high-performance AI chips and a highly available basic software stack.
背後: 規劃與前一日印發的就業友好型意見構成同一批頂層文件。把 6G、算力網與芯片軟件棧並列，等於提前鎖定下一輪基建的採購與標準時間表，企業投資周期可據此對齊。The plan forms a batch of top-level documents with the employment directive issued a day earlier. Placing 6G, the compute network and the chip-and-software stack side by side fixes procurement and standard-setting timelines, giving corporate investment cycles a reference point.

顧問觀點

我們認為，對企業而言這份規劃的實用價值不在技術願景，而在時間表：採購、認證與地方補貼的窗口會圍繞二零三零年這個節點排列。In our view, the practical value for companies lies less in the technology vision than in the timetable: procurement, certification and local subsidy windows will cluster around the 2030 marker.

對日常工作的影響: 資訊科技部門要據此重排未來三年的網絡與終端更新計劃；合規部門則需跟進新一批標準與認證目錄的出台。IT departments should re-sequence three-year network and device refresh plans accordingly, while compliance teams track the new standards and certification catalogues that will follow.
如何改善: 本周把公司現有的網絡與終端更新計劃對照規劃的十方面任務，標出三項與自身業務直接相關、可在明年爭取補貼的項目。This week, map your existing network and device refresh plan against the ten task areas and flag three items tied to your business that could qualify for next year's subsidies.

來源Wind 萬得（2026-10-24）↗

金蝶升級企業智能體操作系統靈基，開放七十個 ERP 連接器與一千一百二十七個工具，付費客戶達三千七百零六家Kingdee upgrades agentic ERP platform Lingee with 70 connectors and 1,127 tools, reaching 3,706 paying customers

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT

事件: 十月二十三日，金蝶在北京發佈靈基 Manage、企業上下文平台與 ERP 連接器。連接器首批七十個、一千一百二十七個工具，覆蓋星瀚、星空、星辰與人力資源系統；截至當日付費客戶三千七百零六家，並發布一千四百餘題的 Lingee Bench。Kingdee released Lingee Manage, an enterprise context platform and ERP connectors in Beijing on 23 October. The first batch covers 70 connectors and 1,127 tools across its Xinghan, Xingkong, Xingchen and HR systems. Paying customers reached 3,706 that day, alongside the launch of the 1,400-question Lingee Bench.
背後: 企業智能體過去卡在兩處：不懂企業語境，也無法真正寫入業務系統。把上下文與連接器同時開放，是從演示走向可交付的分界線，也是內地管理軟件商對代理時代的集體補課。Enterprise agents stalled at two points: not understanding company context and not being able to write into business systems. Opening context and connectors together marks the line between demo and deliverable, and is mainland management-software vendors catching up with the agent era.

顧問觀點

我們認為，連接器數量並非重點，真正的分水嶺是評測基準是否公開題目與評分方法，否則客戶仍只能憑銷售演示判斷成敗。In our view, the connector count is not the point. The real dividing line is whether the benchmark publishes its questions and scoring, otherwise buyers still judge success from a sales demo.

對日常工作的影響: 財務與採購主管要把代理的寫入權限收窄到特定單據類型；資訊科技部門則須為每個連接器訂明誰有權開通。Finance and procurement leads must narrow agent write access to specific document types, while IT has to specify who is authorised to enable each connector.
如何改善: 本周挑選一個重複度最高的單據流程，先只開放查詢權限試行兩周，記錄人工覆核比例再決定是否放開寫入。This week, pick your most repetitive document workflow, open read-only access for two weeks, and record the human review rate before allowing write access.

來源中國經濟網（2026-10-23）↗上海證券報（2026-10-23）↗中國經營報（2026-10-23）↗

南韓與日本在邱舉行第廿二屆經濟交流會議，一百五十名代表簽署人工智能與機械人、先進材料合作方向South Korea and Japan hold 22nd economic exchange conference in Daegu, with 150 delegates focusing on AI, robotics and advanced materials

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners資訊科技IT

事件: 十月二十三日，南韓產業通商資源部與日本經濟產業省在邱合辦第二十二屆韓日（北陸）經濟交流會議，約一百五十名政府、企業與研究機構代表出席。人工智能與機械人環節由南韓雲端軟件商 Tilon 與機械人企業 iM Robotics 介紹工廠應用案例。South Korea's trade ministry and Japan's METI held the 22nd Korea-Japan (Hokuriku) Economic Exchange Conference in Daegu on 23 October, with about 150 delegates from government, business and research bodies. In the AI and robotics session, Korean cloud software firm Tilon and robotics firm iM Robotics presented factory applications.
背後: 會議自二零零零年起連接南韓四個道市與日本北陸三縣，是地方產業層面的對話而非首腦級協議。兩國把人工智能與機械人列入議程主體，反映合作主軸正由貿易投資轉向製造現場的技術落地。Running since 2000, the conference links four Korean provinces and cities with Japan's three Hokuriku prefectures — a subnational industry forum rather than a summit. Putting AI and robotics on the main agenda shows the axis shifting from trade and investment to shop-floor technology adoption.

顧問觀點

我們認為，這種地方對地方的安排對企業的實用價值高於宣言：北陸是精密機械與材料集聚地，與地方軟件商結合，最適合做中小型自動化試點。In our view, such subnational formats matter more to companies than declarations. Hokuriku clusters precision machinery and materials, and pairing it with regional software firms suits small-scale automation pilots.

對日常工作的影響: 有意開拓日韓市場的製造商可從縣級商會入手；廠務主管則可留意同類自動搬運與智能工廠案例，作技術選型的對照。Manufacturers eyeing Japan or Korea can start with prefectural chambers of commerce, while plant managers should watch comparable AGV and smart-factory cases when comparing technology options.
如何改善: 本周挑一條重複搬運工序，記錄現時人力與節拍，作為日後與日韓方案商洽談試點時的對照基線。This week, pick one repetitive handling process and log its current labour and cycle time as a baseline for later pilot talks with Japanese or Korean vendors.

來源Aju Press（2026-10-23）↗韓國產業通商資源部（2026-10-23）↗

OpenAI 擬以一點二五萬億美元投前估值集資最多三百億美元，阿聯酋投資者與 Coatue 領投OpenAI nears US$30 billion raise at US$1.25 trillion pre-money valuation, led by UAE investors and Coatue

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT

事件: 據 CNBC 十月二十二日報道，OpenAI 接近完成最多三百億美元融資，投前估值一點二五萬億美元，由阿聯酋投資者與 Coatue 領投，目標下周內完成簽署，並同步洽談從阿聯酋取得額外算力。CNBC reported on 22 October that OpenAI is nearing a raise of up to US$30 billion at a US$1.25 trillion pre-money valuation, led by UAE investors and Coatue, targeting signatures within the following week and negotiating additional compute from the UAE in parallel.
背後: 時點選在對手啟動上市程序之前，反映前沿實驗室如今爭奪的是同一批長期資金。把股權與算力兩份協議並行談判，也說明資本與基礎設施已難分開定價。The timing, ahead of a rival's listing process, shows frontier labs now compete for the same pool of long-term capital. Negotiating equity and compute in parallel also shows capital and infrastructure can no longer be priced apart.

顧問觀點

我們認為，企業客戶不宜把這類估值當作供應商穩定性的證明；更值得追蹤的是算力協議條款，因為它決定模型交付節奏。In our view, enterprise buyers should not read such valuations as proof of vendor stability. The compute agreement terms matter more, because they set the pace of model delivery.

對日常工作的影響: 資訊科技採購在續約時應加入供應商集中度與算力來源條款；財務則需為單一供應商依賴準備替代方案的成本。IT procurement should add vendor-concentration and compute-source clauses at renewal, while finance budgets the cost of an alternative to single-supplier dependence.
如何改善: 本周在模型供應商清單上補一欄，記下每家可替代的次選模型與切換所需工時，為續約談判留後手。This week, add a column to your model-vendor list recording a fallback model and the engineer-hours needed to switch, giving you leverage at renewal.

來源International Business Times（2026-10-23）↗TechBooky（2026-10-23）↗

高盛投入五億美元成立小企業人工智能研究院，四家前沿實驗室參與，調查指八成六校友企業已用人工智能、僅一成五嵌入核心流程Goldman Sachs commits US$500 million to a Small Business AI Institute with four frontier labs as partners

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners人力資源HR市場推廣Marketing

事件: 十月二十二日，高盛宣布成立小企業人工智能研究院，屬一萬家小企業計劃新增五億美元承諾的一部分。Anthropic、Google、Meta 與 OpenAI 為合作方；其調查顯示二千四百餘家校友企業中八成六已用人工智能，僅一成五完全嵌入核心營運。Goldman Sachs announced a Small Business AI Institute on 22 October, part of a new US$500 million commitment to its 10,000 Small Businesses programme, with Anthropic, Google, Meta and OpenAI as partners. Its survey of more than 2,400 alumni found 86% use AI but only 15% have it fully integrated into core operations.
背後: 培訓與工具由前沿實驗室提供，資金與客戶網絡由銀行提供，這種分工正在成為中小企人工智能普及的標準模板。八成六與一成五之間的落差，說明瓶頸已由工具可得性轉為流程改造能力。Labs supply training and tools while the bank supplies capital and a client network, a division of labour fast becoming the template for SME adoption. The gap between 86% and 15% shows the bottleneck has moved from tool availability to process redesign.

顧問觀點

我們認為，中小企真正缺的不是模型額度，而是肯為流程重寫負責的人；培訓若只教提問，一年後使用率仍會回落。In our view, SMEs lack not model credits but someone accountable for rewriting processes. Training that only teaches prompting will see usage fall back within a year.

對日常工作的影響: 老闆要指定一名流程擁有人而非工具擁有人；市場部可先從報價與內容生產兩條線試行，會計則保留人工覆核節點。Owners should appoint a process owner rather than a tool owner. Marketing can start with quoting and content production, while accounting keeps a human review checkpoint.
如何改善: 本周選一條最花人手的內部流程，記錄改造前所需工時，一個月後用同一張表對比，作為是否擴大投入的依據。This week, pick your most labour-intensive internal process, log the hours it takes, and compare on the same sheet a month later to decide whether to expand.

來源Goldman Sachs（2026-10-22）↗NeoTeo（2026-10-23）↗

Meta 把新商業條款延至十一月二十日生效，歐盟與英國監管機構同步要求交代人工智能訓練資料摘要Meta delays new commercial terms to 20 November as EU and UK regulators press for AI training-data summaries

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance市場推廣Marketing

事件: 十月二十三日，Meta 宣布把新版商業條款與自助廣告條款生效日由十月三十日延至十一月二十日。此前出版商與廣告商向歐盟及英國競爭與市場管理局投訴，指條款賦予 Meta 就內容用於人工智能的無限權利。On 23 October Meta pushed the effective date of its revised Commercial Terms and Self-Serve Ad Terms from 30 October to 20 November, after publishers and advertisers complained to EU and UK competition authorities that the terms grant unlimited rights to use their content for AI.
背後: 延期並非讓步，而是為談判爭取時間。歐委會同時表示已向尚未公布訓練資料摘要的企業發出資訊要求，平台與內容方的爭議正被拉回《人工智能法案》的透明度義務框架內。The delay is time bought for negotiation, not a concession. The European Commission said it has sent information requests to firms yet to publish training-data summaries, pulling the platform-versus-publisher dispute back into the AI Act's transparency obligations.

顧問觀點

我們認為，企業官網與品牌內容的授權價值此後要有明確條款支撐；靠 robots.txt 表達意願，在合同層面幾乎沒有約束力。In our view, the licensing value of corporate sites and brand content now needs explicit contractual support. Expressing preferences through robots.txt carries almost no contractual force.

對日常工作的影響: 法務要檢視廣告與平台合約中的內容授權範圍；市場部則需盤點官網內容是否已被用作訓練資料。Legal should review content-licensing scope in ad and platform contracts, while marketing audits whether site content has already been drawn into training data.
如何改善: 本周在官網加入一段明確的內容使用聲明，保留授權與否的表態，並同步更新 robots.txt 與合約附件。This week, add an explicit content-use notice to your website reserving your licensing position, and align robots.txt and contract annexes with it.

來源El Fondo（2026-10-23）↗

美國商務部擬向四家企業撥款最高三點八九億美元，功率半導體與 AI 機架供電成重點U.S. Commerce proposes up to US$389 million for four firms, targeting power semiconductors and AI rack delivery

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT企業決策者Business owners

事件: 美國商務部與四家企業簽署資助意向書，擬撥款最高三點八九億美元，屬《芯片與科學法案》一百億美元研發計劃。Vertical Semiconductor 獲七千三百萬美元研發垂直氮化鎵器件，用於 AI 數據中心機架。The U.S. Commerce Department signed letters of intent with four firms for up to US$389 million under the CHIPS and Science Act's US$10 billion R&D programme. Vertical Semiconductor would receive US$73 million for vertical gallium nitride devices aimed at AI data-centre racks.
背後: 資金流向已由晶圓製造延伸到供電與材料，顯示算力擴張的瓶頸正從芯片數量轉向電力轉換效率。以少數股權換取研發資助，也是本輪產業政策的固定做法。Funding has moved beyond wafer fabrication into power delivery and materials, showing the bottleneck in compute expansion shifting from chip count to conversion efficiency. Minority equity in exchange for R&D support is now the standard instrument.

顧問觀點

我們認為，數據中心的成本模型要向後端延伸：電力轉換與散熱的設計餘量，會比採購單價更影響三年後的營運成本。In our view, data-centre cost models must extend downstream: design headroom for power conversion and cooling will shape three-year operating costs more than purchase price.

對日常工作的影響: 資訊科技與設施主管要重新核算每機櫃的供電上限；採購則需把功率器件壽命納入標書評分。IT and facilities leads should re-cost power ceilings per rack, while procurement adds power-device lifespan to tender scoring.
如何改善: 本周向供應商索取現有機櫃的功率轉換效率與散熱餘量數據，據此訂出一條新增機櫃不得超越的上限。This week, request conversion efficiency and cooling headroom figures for your current racks, then set a ceiling no new rack may exceed.

來源Analytics Insight（2026-10-23）↗

英偉達在柏林 GTC 發布 DSX POD 模組化 AI 工廠，現場部署期由約六個月縮至最快十四天Nvidia unveils DSX POD modular AI factory at GTC Berlin, cutting site deployment from six months to 14 days

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT企業決策者Business owners

事件: 十月二十一日，英偉達在柏林 GTC 發布 DSX POD 參考設計，把計算、網絡、供電與液冷預裝於集裝箱，單箱容納七個 Vera Rubin NVL72 機櫃；場地配套就緒後，現場部署由約六個月縮至最快十四天，二零二七年初起全球交付。At GTC Berlin on 21 October Nvidia unveiled the DSX POD reference design, packing compute, networking, power and liquid cooling into containers holding seven Vera Rubin NVL72 racks. With site utilities ready, deployment falls from about six months to as little as 14 days, with global delivery from early 2027.
背後: 算力競賽的瓶頸早已不是芯片交付，而是場地、電力與施工協調。把製造與驗證移出工地，等於允許設備與場地並行準備，也把交付責任轉移到合作夥伴身上。The bottleneck in the compute race long ago moved from chip delivery to sites, power and construction coordination. Moving fabrication and validation off-site lets equipment and site preparation run in parallel and shifts delivery responsibility to partners.

顧問觀點

我們認為，十四天只計算現場工序；真正的準備期取決於併網與地方審批。把它當作採購交期，會嚴重低估項目風險。In our view, the 14 days covers only on-site work; real lead time depends on grid connection and local approvals. Treating it as procurement lead time badly underestimates project risk.

對日常工作的影響: 項目經理要把併網審批與地方許可直接排進關鍵路徑；財務則須為提前供電的違約條款預留保證金。Project managers must put grid approval and local permits on the critical path, while finance reserves guarantees for early-power breach clauses.
如何改善: 本周把現有數據中心項目的關鍵路徑重畫一次，把併網審批單獨列為一條，確認它是否才是真正的交期瓶頸。This week, redraw the critical path of your data-centre project with grid approval as a separate line, and check whether it is the real bottleneck.

來源證券時報（2026-10-23）↗騰訊新聞（2026-10-23）↗

新加坡資媒局官員論銀行代理權限：人不必逐筆批准，但須對結果負責並留下可追溯紀錄Singapore IMDA official sets out how banks should delegate authority to agents, with audit trails as the evidence

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT法務合規Legal & compliance

事件: 新加坡資訊通信媒體發展局人工智能治理與安全群組總監李婉絲表示，代理系統與生成式人工智能之別在於可採取行動，機構須界定其權限邊界。資媒局今年一月推出代理式人工智能治理框架，五月更新。Lee Wan Sie, cluster director of AI governance and safety at Singapore's IMDA, said agents differ from generative AI because they can act, so institutions must define the boundaries of their authority. IMDA issued its Model AI Governance Framework for Agentic AI in January and updated it in May.
背後: 金融業正由審查模型輸出轉為審查系統權限。框架以自願性質補足現行法律，並把人類監督重新定義為例外處理而非逐筆批准，這與多國監管方向一致，只是先行一步。Finance is moving from reviewing model outputs to reviewing system authority. The voluntary framework supplements existing law and redefines human oversight as exception handling rather than per-action approval, in line with regulators elsewhere, just earlier.

顧問觀點

我們認為，關鍵是把權限寫成可審核的清單：能做什麼、不得做什麼、超過哪個金額必須停手，而非只用一句「人在迴路」交代。In our view, the point is to write authority as an auditable list: what the agent may do, may not do, and at what value it must stop, rather than settling the matter with a line about a human in the loop.

對日常工作的影響: 法務要為代理訂明授權與義務邊界；資訊科技須確保每筆代理動作可回溯到授權來源；財務則設金額門檻。Legal should define the mandate and duty boundaries, IT must make every agent action traceable to its authority, and finance sets the value thresholds.
如何改善: 本周為一個代理流程寫出三行權限表：可執行、須批准、禁止，並附上金額與頻次上限。This week, write a three-line authority sheet for one agent workflow, covering allowed actions, actions needing approval and prohibited ones, with value and frequency caps.

來源The Asian Banker（2026-10-23）↗

本週可做的三件事

把未來六個月打算自動化的工序逐一列出，標明受影響崗位與人數，並在公告前預留轉崗說明。List the processes you plan to automate in the next six months, mark the roles and headcount affected, and prepare redeployment notes before any announcement.
為已上線的代理指定一名最終問責人，並訂下三個需要人手批准的門檻。Name one accountable owner per live agent and set three thresholds that require human approval.
抽查一套跨雲依賴的代理流程，確認模型、資料庫與密鑰不在同一個地區、同一個供應商之下。Pick one agent workflow that spans clouds and verify that models, databases and secrets do not all sit in one region under one vendor.

常見問題

今日（2026-10-24）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、國務院辦公廳印發就業友好型意見，六方面十八項任務把崗位影響寫入人工智能推廣條件；2、《國家信息化「十五五」規劃》印發，到二零三零年適時啟動 6G 商用並加快構建全國一體化算力網；3、金蝶升級企業智能體操作系統靈基，開放七十個 ERP 連接器與一千一百二十七個工具，付費客戶達三千七百零六家；4、南韓與日本在邱舉行第廿二屆經濟交流會議，一百五十名代表簽署人工智能與機械人、先進材料合作方向；5、OpenAI 擬以一點二五萬億美元投前估值集資最多三百億美元，阿聯酋投資者與 Coatue 領投；6、高盛投入五億美元成立小企業人工智能研究院，四家前沿實驗室參與，調查指八成六校友企業已用人工智能、僅一成五嵌入核心流程；7、Meta 把新商業條款延至十一月二十日生效，歐盟與英國監管機構同步要求交代人工智能訓練資料摘要；8、美國商務部擬向四家企業撥款最高三點八九億美元，功率半導體與 AI 機架供電成重點；9、英偉達在柏林 GTC 發布 DSX POD 模組化 AI 工廠，現場部署期由約六個月縮至最快十四天；10、新加坡資媒局官員論銀行代理權限：人不必逐筆批准，但須對結果負責並留下可追溯紀錄。每條均附本會顧問觀點、對日常工作的影響與一項可即時執行的建議。

本會對今日 AI 局勢的整體判斷是甚麼？

今日兩條主線在同一天落地。國辦印發就業友好型意見，把崗位替代風險評估、穩慎推廣節奏與失業緩衝期寫成推廣條件。另一邊，英偉達把 AI 工廠裝進集裝箱、部署期由六個月壓至最快十四天，美國商務部再向功率半導體撥款最高三點八九億美元。算力照舊搶，人手不再是可以事後頭痛的變數。

這些情報的資料來源是甚麼？如何核實？

本欄每日由 AIX Society 編輯部檢索公開資料後撰稿，每條新聞均附原始來源連結（本期包括：人民日報、證券時報、Wind 萬得、中國經濟網、上海證券報、中國經營報、Aju Press、韓國產業通商資源部），並將事實與本會觀點分列。讀者可按來源連結自行核實。

訂閱每週電子報

每週一封：AI 轉型的實戰觀察、工具實測與案例拆解。