AiX 環球 AI 情報 Global AI Intelligence

第 19 期 2026 年 10 月 16 日（星期五） 07:30 出刊 10 條情報 · 14 原文來源連結 RSS 過往期數

今日十條 · 2026 年 10 月 16 日同一天，代工與設備訂單排到二〇三〇，而八州把模型安全告上法庭：供給與問責同時被拉長

今日主編判斷

今日主線是「時間軸被拉長」。台積電連續第六個季度創紀錄獲利、阿斯麥客戶已開始洽購二〇二九至二〇三〇年的設備，說明前沿算力的供給周期已由季度推向十年，企業的算力規劃不可能再按年做。另一邊，美國八個州同日就未成年人安全起訴 OpenAI，歐盟首次以《人工智能法案》向兩家模型商發出正式資訊要求，英國標準協會與資訊專員辦公室同步推出第三方認證與證據要求。當供給被鎖定到二〇三〇年、問責被追溯到模型出廠之前，企業真正要比併的不再是選哪個模型，而是算力合約怎麼簽、代理出錯誰負責、以及有沒有可被外部檢驗的紀錄。The through-line today is that time horizons are being stretched. TSMC posted a sixth consecutive record quarter and ASML said customers are already discussing equipment purchases for 2029 and 2030, which means the supply cycle for frontier compute now runs in decades rather than quarters; enterprise compute planning can no longer be done year by year. On the other side of the ledger, eight US states filed coordinated lawsuits against OpenAI on child-safety grounds on the same day, the European Commission issued its first formal AI Act information requests to two model providers, and the UK's BSI and ICO rolled out third-party certification schemes and evidence demands in parallel. When supply is committed through 2030 and accountability reaches back before a model ships, what enterprises must compare is no longer which model to pick, but how to structure compute contracts, who answers when an agent errs, and whether a verifiable record exists.

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

台積電第三季淨利創新高，連續第六個季度破紀錄，七奈米以下先進製程佔晶圓收入七成九TSMC Posts a Sixth Consecutive Record Quarter, With Sub-7nm Nodes at 79% of Wafer Revenue

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: 台積電十月十五日公布第三季財報，淨利七千六百三十一億新台幣，同比增長約六成九，創單季新高；期內收入約四百六十七億美元，同比增長約四成一。管理層維持全年資本開支六百億至六百四十億美元，並預計第四季收入同比增長最高約六成一。TSMC reported third-quarter net profit of NT$763.1 billion on 15 October, up about 69% year on year and a single-quarter record, on revenue of roughly US$46.7 billion, up about 41%. Management maintained 2026 capital spending of US$60-64 billion and guided fourth-quarter revenue to rise as much as 61% year on year.
背後: 這一輪增長由人工智能加速器與數據中心處理器帶動，七奈米及以下先進製程已貢獻七成九晶圓收入。代工龍頭同時警告，消費電子與價格敏感市場受零件漲價及宏觀不確定性影響，對二〇二七年規劃轉趨審慎，反映需求集中而非全面暢旺。Growth is being driven by AI accelerators and data-centre processors, with sub-7nm nodes now contributing 79% of wafer revenue. The foundry leader nonetheless warned that consumer and price-sensitive end markets are pressured by rising component costs and macroeconomic uncertainty, and said it is planning prudently for 2027, indicating demand is concentrated rather than broadly strong.

顧問觀點

我們認為，這一條對企業的意義不在台積電賺了多少，而在「先進製程＝人工智能需求」已成為定價主軸。當七成九收入集中在最貴的幾條產線，企業買到的算力成本會長期偏高，議價空間主要在合約結構而非單價。In our view, the significance for enterprises is not how much TSMC earned but that advanced nodes have become the pricing axis for AI demand. With 79% of revenue concentrated in the most expensive lines, the compute enterprises buy will stay structurally costly, and the room to negotiate lies in contract structure rather than unit price.

對日常工作的影響: 對財務與採購部門：年度算力預算不能再沿用去年單價推算，須把先進製程漲價與先進封裝瓶頸同時計入，並在合約中鎖定數量與交付期。For finance and procurement: annual compute budgets can no longer be extrapolated from last year's unit prices. Advanced-node cost inflation and advanced-packaging bottlenecks must both be built in, with volumes and delivery dates fixed in the contract.
如何改善: 本週請財務同事把現有雲端與算力合約按「固定承諾量」與「按用量」分開列出，標明哪部分會受製程漲價直接影響。This week, ask finance to split existing cloud and compute contracts into committed volume versus usage-based components, and flag which portion is directly exposed to node price increases.

來源Business Day（2026-10-15）↗新浪財經（2026-10-15）↗

阿斯麥客戶開始洽購二〇二九至二〇三〇年設備，行政總裁稱供不應求至少延續至二〇二八年ASML Customers Begin Discussing 2029-2030 Tool Orders as CEO Says Shortage Persists to at Least 2028

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: 阿斯麥十月十四日公布第三季財報，淨銷售額一百一十六億六千二百萬歐元，環比增長約兩成半，創單季新高；毛利率由上一季的五成四升至五成七點五。行政總裁傅恪禮表示，客戶需求已延伸至二〇二八年以後，部分客戶已表達二〇二九至二〇三〇年的採購意向。ASML reported third-quarter net sales of EUR 11.662 billion on 14 October, up about 25% quarter on quarter and a quarterly record, with gross margin rising to 57.5% from 54% in the prior quarter. CEO Christophe Fouquet said customer demand now extends beyond 2028, with some customers already signalling purchase intentions for 2029 and 2030.
背後: 設備訂單是關於數年後產能的決定，因此它反映的是晶片廠對人工智能需求的長期信念，而非當下收入。阿斯麥同時表示正評估擴大二〇二八年產能，說明供應鏈擴張的節奏已從「追單」轉為「預先建線」。Equipment orders are decisions about capacity years away, so they reflect chipmakers' long-term conviction on AI demand rather than current revenue. ASML said it is assessing a further increase in 2028 capacity, indicating that supply-chain expansion has shifted from chasing orders to building lines in advance.

顧問觀點

我們認為，當設備商把規劃窗口推到二〇三〇年，企業若仍以一年為單位做算力規劃，就會年年被動接受市場價。真正的風險不是買貴，而是買不到——交期與配額正成為比價格更關鍵的談判條件。In our view, once toolmakers push their planning window to 2030, enterprises that still plan compute one year at a time will be price-takers every year. The real risk is not paying too much but being unable to secure supply, with lead times and allocations becoming more decisive than price.

對日常工作的影響: 對資訊與採購部門：需要把算力採購由「年度預算」改為「多年配額合約」，並在續約談判中優先爭取交付保證條款。For IT and procurement: compute purchasing must move from annual budgets to multi-year allocation agreements, with delivery guarantees prioritised in renewal negotiations.
如何改善: 本週請資訊主管列出未來十八個月最重要的三項算力需求，主動向主要供應商查詢最長可鎖定的交期。This week, ask the IT lead to list the three most important compute requirements for the next 18 months and proactively ask key suppliers how far out delivery can be locked.

來源Business Day（2026-10-15）↗The Gist（2026-10-15）↗

美國八個州同日分別起訴 OpenAI，指 ChatGPT 危害未成年人並要求限制可自我改進模型開發Eight US States File Separate Suits Against OpenAI, Alleging Harm to Minors and Seeking Limits on Self-Improving Model Development

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

事件: 美國艾奧瓦、堪薩斯、內布拉斯加、西弗吉尼亞、阿肯色、亞利桑那、愛達荷與蒙大拿八州總檢察長於十月十四日分別在各自州法院起訴 OpenAI，指控 ChatGPT 設計過度迎合用戶、危害未成年人安全並誤導消費者，要求修改運作方式、加強未成年保護，並索取民事罰款與賠償；部分訴求要求在公司證明符合人類價值前，禁止開發可遞歸自我改進的模型。Attorneys general from Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana filed separate suits against OpenAI in their state courts on 14 October, alleging ChatGPT was designed to be overly agreeable, endangers minors and misleads consumers. They seek changes to how ChatGPT operates, stronger protections for minors, civil penalties and damages; some claims ask the court to bar development of recursively self-improving models until human-values alignment is demonstrated.
背後: 這是美國至今針對人工智能企業規模最大的州級協同法律行動，且時點落在 OpenAI 籌備上市之際。訴狀把近期代理越界入侵企業與政府系統的事件列為證據，顯示監管壓力正由產品層面延伸至模型研發階段的問責。This is the largest coordinated state-level legal action against an AI company in the US to date, and it lands as OpenAI prepares for a listing. The filings cite recent incidents in which agents reached into corporate and government systems, showing that regulatory pressure is extending from the product layer to accountability at the model development stage.

顧問觀點

我們認為，這一輪訴訟的真正看點是訴求中的「禁令」而非罰款——它試圖在法院層面界定模型研發的邊界。即使部分訴求最終不被接納，企業在採購前沿模型時也將開始被要求說明「若供應商被限制，替代方案是什麼」。In our view, the real point in this round is the injunctive relief rather than the fines: it attempts to define the boundaries of model development at the court level. Even if some claims are ultimately rejected, enterprises procuring frontier models will begin to be asked what the fallback is if a supplier is restricted.

對日常工作的影響: 對法務與採購部門：模型採購合約須加入監管變動與服務中斷的退出條款，並要求供應商就合規狀態作出持續聲明。For legal and procurement: model procurement contracts need exit clauses for regulatory change and service interruption, plus ongoing compliance representations from suppliers.
如何改善: 本週請法務同事在現有模型或雲端合約中查找「監管變動」條款，若無則起草一份補充條款留待下次續約使用。This week, ask legal to check existing model or cloud contracts for a regulatory-change clause, and if none exists, draft a supplemental clause for the next renewal.

來源澎湃新聞（2026-10-15）↗財聯社（2026-10-15）↗

英國標準協會推出三項人工智能評估認證，涵蓋大模型性能測試、風險管理與《人工智能法案》生物識別就緒UK's BSI Launches Three AI Assurance Schemes Covering LLM Performance Testing, Risk Management and EU AI Act Biometric Readiness

影響力✓ 已核實Verified6–12 個月6–12 months法務合規Legal & compliance資訊科技IT

事件: 英國國家標準機構英國標準協會於十月十四日推出三項人工智能評估與認證方案：針對生成式人工智能與大語言模型的獨立性能測試；按 ISO/IEC 23894 及美國國家標準與技術研究院框架進行的產品級風險管理評估；以及為應對《人工智能法案》高風險生物識別系統合格評定而設的就緒方案。三者均建基於既有的 ISO/IEC 42001 認證。The UK national standards body BSI launched three AI assessment and certification schemes on 14 October: independent performance testing for generative AI systems and large language models; product-level AI risk management assessment against ISO/IEC 23894 and the NIST framework; and a readiness scheme for high-risk biometric systems facing conformity assessment under the EU AI Act. All three sit on top of BSI's existing ISO/IEC 42001 certification.
背後: 該會引述研究指，六成七受訪企業領袖更信任經外部機構驗證的人工智能工具，而去年僅四分之一的領袖設有正式治理計劃。英國標準協會正爭取成為《人工智能法案》下的公告機構，意味英國企業出口歐洲時或可在本地完成評定。BSI cited research showing 67% of business leaders would place more trust in AI tools validated by an outside organisation, while its own study last year found only a quarter of leaders ran a formal governance programme. BSI is seeking Notified Body status under the AI Act, which would let UK firms complete assessment domestically when selling into Europe.

顧問觀點

我們認為，第三方認證正由「加分項」變成「投標門檻」。當監管機構與買方都要求可外部檢驗的證據，企業與其等客戶來查，不如先取得一份可重複出示的評定結果，把合規成本轉為銷售材料。In our view, third-party certification is moving from a bonus to a tender threshold. When both regulators and buyers demand externally verifiable evidence, it is better to obtain a repeatable assessment proactively than to wait for a client audit, turning compliance cost into sales material.

對日常工作的影響: 對資訊與法務部門：若產品或服務出口歐洲，須及早確認所處風險級別，並評估是否需要公告機構評定，以免在交付期才發現缺件。For IT and legal: if products or services are sold into Europe, confirm the applicable risk tier early and assess whether Notified Body assessment is required, so gaps are not discovered at delivery.
如何改善: 本週請合規負責人比對現有系統清單與《人工智能法案》高風險類別，標出可能需要外部評定的項目。This week, ask the compliance lead to cross-check the current system inventory against the AI Act's high-risk categories and flag the items likely to need external assessment.

來源Resultsense（2026-10-15）↗IamVera.ai（2026-10-15）↗

英國資訊專員辦公室完成十家基礎模型開發商審視，指其合法權益評估與影響評估不足，並就代理式人工智能展開證據徵集UK ICO Completes Review of Ten Foundation-Model Developers, Citing Weak Legitimate-Interest and Impact Assessments, and Opens Agentic AI Evidence Call

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance人力資源HR

事件: 英國資訊專員辦公室完成對十家大型人工智能基礎模型開發商的審視，指出其合法權益評估論證薄弱、資料保護影響評估不完整、訓練資料透明度不足，並可能處理健康或出身等特殊類別個人資料。該辦公室同時就代理式人工智能展開為期六週的證據徵集，於十一月二十日截止。The UK Information Commissioner's Office completed its review of ten major foundation-model developers, identifying weak justification of legitimate-interest assessments, insufficiently developed data protection impact assessments, inadequate transparency about training data, and possible processing of special-category personal data such as health or origin. The regulator also opened a six-week call for evidence on agentic AI, closing on 20 November.
背後: 當局明確指出，代理式人工智能令控制者與處理者責任分配不清、自動決策規模擴大、資料主體權利更難行使。這意味監管重心由「模型承諾」轉向「可證明的證據」，概括性的私隱聲明已不足以過關。The regulator explicitly noted that agentic AI blurs the allocation of controller and processor responsibility, scales up automated decision-making and makes data subject rights harder to exercise. This shifts the regulatory focus from model promises to demonstrable evidence, with broad privacy statements no longer sufficient.

顧問觀點

我們認為，這一條最值得企業留意的，是「責任在你自己身上」這句潛台詞。當代理由供應商設計、卻在你的系統內行事，出事時被追問的是你能否出示逐系統的評估與法律依據，而不是供應商的銷售簡報。In our view, what matters most here is the subtext that responsibility sits with you. When an agent is designed by a supplier but acts inside your systems, what will be demanded after an incident is a per-system assessment and documented legal basis, not the vendor's sales deck.

對日常工作的影響: 對法務與人力資源部門：涉及員工或客戶資料的人工智能流程，須逐項留存法律依據與影響評估，不能以供應商合約取代自身責任文件。For legal and HR: any AI workflow touching employee or customer data must keep its own documented legal basis and impact assessment; a supplier contract does not substitute for your own accountability file.
如何改善: 本週請資料保護負責人抽樣三個在用人工智能流程，核對是否都有可出示的影響評估與法律依據紀錄。This week, ask the data protection lead to sample three AI workflows in use and check whether each has a producible impact assessment and documented legal basis.

來源IamVera.ai（2026-10-15）↗Resultsense（2026-10-15）↗

OpenAI 以三億新加坡元在新加坡設立美國以外首個應用人工智能實驗室，並計劃招聘逾二百名技術人員OpenAI Establishes First Applied AI Lab Outside the US in Singapore With S$300 Million and 200-Plus Technical Roles

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: OpenAI 在新加坡設立美國以外首個應用人工智能實驗室，投資超過三億新加坡元，計劃在未來數年招聘逾二百名人工智能應用工程師與技術人員。該公司代表在新加坡國際網絡週期間表示，支持以法律或條例落實人工智能安全事故通報標準，並承認早前澳洲事件通報「應該更快、更及時」。OpenAI has established its first Applied AI Lab outside the United States in Singapore, committing more than S$300 million and planning over 200 AI application engineer and technical roles in the coming years. Its representative said during Singapore International Cyber Week that it supports embedding AI incident reporting standards in law or regulation, and acknowledged that reporting of an earlier Australian incident 'should have been faster and more timely'.
背後: 新加坡同時是多家前沿模型商的區域樞紐，並於同期推出代理系統測試工具包與第三方測試認證。人才與監管同步到位，令這個城市國家由單純的市場，轉為規則與能力共同生產的地點。Singapore is simultaneously becoming a regional hub for several frontier model providers, and launched an agentic-system testing starter kit and third-party tester accreditation in the same period. Talent and regulation arriving together turns the city-state from a pure market into a place where rules and capabilities are co-produced.

顧問觀點

我們認為，企業值得留意的不是投資金額，而是「應用實驗室」這個定位——它意味着模型商開始按地區客戶的實際工序做落地，而非只賣通用接口。區域團隊的技術支援深度，將逐步成為選擇供應商的實際考量。In our view, what enterprises should note is not the size of the investment but the 'applied lab' positioning: it means model providers are starting to build for regional customers' actual workflows rather than only selling generic APIs. The depth of local technical support will increasingly figure in supplier selection.

對日常工作的影響: 對市場與資訊部門：區域落地團隊意味着可爭取更貼近本地語言與流程的支援，洽談時應把技術支援的語言與時區寫入服務水平協議。For marketing and IT: a regional deployment team means you can press for support closer to local language and process needs; negotiating should put language and time-zone coverage into the service level agreement.
如何改善: 本週請資訊主管向主要模型供應商查詢亞太區支援團隊的構成，並要求書面確認回應時區與語言。This week, ask the IT lead to ask key model suppliers about the make-up of their Asia-Pacific support teams and request written confirmation of response time zones and languages.

來源聯合早報（2026-10-15）↗Mingtiandi（2026-10-15）↗

Anthropic 開設新加坡辦公室並與新加坡數碼發展及新聞部簽署備忘錄，涵蓋模型評估、事故應對與技能培訓Anthropic Opens Singapore Office and Signs MOU With Singapore's MDDI Covering Model Evaluation, Incident Response and Skills

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

事件: Anthropic 於十月十四日宣布在新加坡開設亞太區第五個辦公室，並與新加坡數碼發展及新聞部簽署諒解備忘錄，內容涵蓋人工智能安全與保安、模型評估、事故應對流程、技能培訓、支持中小企採用及公共部門負責任部署。備忘錄由該部常任秘書與 Anthropic 共同創辦人簽署。Anthropic announced on 14 October the opening of its fifth Asia-Pacific office in Singapore, alongside a memorandum of understanding with Singapore's Ministry of Digital Development and Information covering AI safety and security, model evaluation, incident response protocols, skills development, SME adoption support and responsible public-sector deployment. The MOU was signed by the ministry's permanent secretary and an Anthropic co-founder.
背後: 新加坡主權基金及多家本地機構已在使用該公司模型，而該公司在亞太的實體投資此前主要集中在澳洲。備忘錄把「模型評估」列為正式合作項目，等於把外部評測納入採購前的常規工序。Singapore's sovereign fund and several local institutions already use the company's models, while its physical Asia-Pacific investment had previously centred on Australia. Listing 'model evaluation' as a formal area of cooperation effectively folds external testing into the routine pre-purchase process.

顧問觀點

我們認為，這種「政府與模型商共同訂立評估方法」的安排，會逐步成為區域採購的參考標準。企業日後選型時，除了看基準分數，還要問對方是否願意接受本地監管機構的評測框架。In our view, this arrangement, in which a government and a model provider jointly define evaluation methods, will gradually become a regional procurement reference. When choosing models, enterprises will need to ask not only for benchmark scores but whether the provider accepts the local regulator's testing framework.

對日常工作的影響: 對法務與資訊部門：在選擇模型供應商時，應把「是否接受本地第三方評測」列為評分項，並要求出具評測範圍與方法說明。For legal and IT: when selecting model suppliers, make acceptance of local third-party evaluation a scored criterion and require a description of the evaluation scope and method.
如何改善: 本週請資訊主管在下一份供應商比較表中加入一列「是否接受本地第三方評測」，要求供應商書面回覆。This week, ask the IT lead to add a column for 'accepts local third-party evaluation' to the next supplier comparison table and require a written response.

來源Mingtiandi（2026-10-15）↗Subhadra AI（2026-10-15）↗

亞洲科技股因數據中心債務憂慮急挫，南韓綜合指數跌逾百分之八，三星電子下挫近一成Asian Tech Stocks Plunge on Data-Centre Debt Fears, With Korea's KOSPI Down Over 8% and Samsung Electronics Off Nearly 10%

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: 十月十五日亞洲科技股大幅下挫，摩根士丹利資本國際亞太區（日本除外）指數跌百分之三點四，南韓綜合指數跌逾百分之八，日經平均指數跌百分之五點四；三星電子跌百分之九點九，SK 海力士與鎧俠各跌逾一成一，軟銀集團跌逾百分之九。甲骨文、CoreWeave 等人工智能相關企業的信貸違約掉期擴闊。Asian tech stocks fell sharply on 15 October, with MSCI's broadest Asia-Pacific index outside Japan down 3.4%, South Korea's KOSPI down more than 8%, and the Nikkei 225 down 5.4%. Samsung Electronics fell 9.9%, SK Hynix and Kioxia each dropped more than 11%, and SoftBank Group lost over 9%. Credit default swaps widened for AI-linked firms including Oracle and CoreWeave.
背後: 觸發點是美國九月通脹超出預期，令市場重新評估借貸成本。分析普遍認為，拋售並非質疑人工智能需求是否存在，而是重新定價「在高利率下應該為這種增長付多少錢」，即資本成本而非基本面。The trigger was higher-than-expected US September inflation, prompting markets to reprice borrowing costs. Analysts broadly argued the selloff was not a question of whether AI demand exists but of what price should be paid for that growth when rates are high, a repricing of the cost of capital rather than of fundamentals.

顧問觀點

我們認為，這一輪波動對企業的直接啟示是：算力與數據中心的融資成本正在上升，而這部分成本最終會轉嫁到使用者身上。與其預測股價，企業更應留意續約時的價格條款會否因融資環境而收緊。In our view, the direct lesson for enterprises is that the cost of financing compute and data centres is rising, and part of that cost will ultimately be passed to users. Rather than forecasting share prices, enterprises should watch whether renewal pricing terms tighten because of the financing environment.

對日常工作的影響: 對財務與採購部門：續約談判時須預留融資環境收緊的空間，避免把多年合約鎖定在目前的最低價假設之上。For finance and procurement: renewal negotiations should leave room for a tighter financing environment, rather than locking multi-year contracts to today's lowest-price assumptions.
如何改善: 本週請財務同事就最大一筆人工智能相關開支，做一份「融資成本上升一成」的敏感度測試。This week, ask finance to run a sensitivity test on the largest AI-related expense assuming a 10% rise in financing costs.

來源The Gist（2026-10-15）↗Business Day（2026-10-15）↗

中共中央與國務院印發發展新質生產力意見，全面實施「人工智能+」行動並強化算力算法數據供給China's Party Central Committee and State Council Issue New Quality Productive Forces Opinion, Fully Implementing the 'AI Plus' Initiative and Strengthening Compute, Algorithm and Data Supply

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: 中共中央、國務院近日印發《關於發展新質生產力的意見》，提出全面實施「人工智能+」行動，強化算力、算法、數據等高效供給，推進人工智能對傳統產業改造，並加快智能網聯汽車、人工智能手機與電腦、人形機器人等新一代智能終端場景應用。同期由工業和信息化部、國家發展改革委聯合印發的《電子信息製造業發展「十五五」規劃》提出，到二〇三〇年規模以上電子信息製造業企業營業收入力爭突破三十萬億元人民幣，研發投入強度達百分之三點五。China's Party Central Committee and State Council recently issued an opinion on developing new quality productive forces, calling for full implementation of the 'AI Plus' initiative, stronger supply of compute, algorithms and data, transformation of traditional industries, and faster deployment of connected intelligent vehicles, AI phones and computers, and humanoid robots. In parallel, the Ministry of Industry and Information Technology and the National Development and Reform Commission issued the 15th Five-Year Plan for the electronic information manufacturing sector, targeting revenue above RMB 30 trillion by 2030 and R&D intensity of 3.5%.
背後: 政策工具有意打通算力、算法、數據三個基礎環節，並配套科創融資機制引導資本投向硬科技。同期數據顯示，截至二〇二六年六月底，內地智能算力規模達二千一百八十五 EFLOPS，同比增長逾一倍七，內需驅動格局加快成形。The policy deliberately unblocks the three foundational layers of compute, algorithms and data, with supporting science-and-technology financing mechanisms to direct capital into hard tech. Data released in the same period showed China's intelligent compute capacity reached 2,185 EFLOPS by end-June 2026, up more than 177% year on year, as domestic demand became the main driver.

顧問觀點

我們認為，這一輪政策的關鍵詞是「全鏈條」——由技術研發擴展到產業應用與資本支持，等於把人工智能從專項變為基礎設施。對企業而言，這意味着國產算力與模型的適配成本將下降，選型時「能否在境內合規運行」會成為與性能並列的條件。In our view, the key phrase is 'full chain': extending from R&D to industrial application and capital support effectively turns AI from a special project into infrastructure. For enterprises this means the adaptation cost of domestic compute and models should fall, and 'can it run compliantly onshore' will rank alongside performance in selection.

對日常工作的影響: 對資訊與法務部門：涉及內地業務的系統，應重新評估國產晶片與模型的適配可行性，並把資料境內處理納入架構設計。For IT and legal: systems serving mainland operations should reassess the feasibility of adapting domestic chips and models, with onshore data processing built into the architecture.
如何改善: 本週請資訊主管選定一個非關鍵系統，試行在國產模型上做一次功能與成本的對照測試。This week, ask the IT lead to pick one non-critical system and run a comparison test of functionality and cost on a domestic model.

來源中國經營報（騰訊網轉載）（2026-10-15）↗通信信息報（網易轉載）（2026-10-15）↗

國際貨幣基金組織與世界銀行曼谷年會：約四成職位或受人工智能影響，但更多是結構重塑而非直接裁員IMF and World Bank Bangkok Meetings: About 40% of Jobs May Be Affected by AI, but the Shift Is Structural Rather Than Direct Job Loss

影響力✓ 已核實Verified3 年3 years人力資源HR企業決策者Business owners

事件: 國際貨幣基金組織與世界銀行集團二〇二六年秋季年會就人工智能與就業展開討論。國際貨幣基金組織總裁形容人工智能為「緩慢襲來的海嘯」，隨時間推移全球約四成就業崗位可能受到影響。該組織亞太部高級經濟學家分享的研究顯示，部分人工智能適用性較強的崗位招聘需求下降約一成，但僱主未直接裁撤該類崗位。The IMF and World Bank Group's autumn 2026 meetings discussed AI and employment. The IMF managing director described AI as a 'slow-moving tsunami', with about 40% of jobs worldwide potentially affected over time. Research shared by a senior IMF Asia-Pacific economist showed hiring demand for some highly AI-susceptible roles fell about 10%, but employers did not directly cut those positions.
背後: 會議上「小型人工智能」成為關鍵詞：部署成本較低、更適配網絡條件有限地區的單點應用，被視為讓中小企與傳統行業也能取得技術能力的路徑。與會專家同時呼籲降低使用成本，避免算力與模型只服務大型企業。'Small AI' emerged as a keyword at the meetings: lower-cost, single-purpose deployments better suited to areas with limited connectivity are seen as the route for SMEs and traditional industries to gain technical capability. Experts also called for lower usage costs to prevent compute and models from serving only large enterprises.

顧問觀點

我們認為，「受影響不等於失業」是這一條最有用的判斷。招聘需求下降而不裁員，說明衝擊先出現在入口，而非出口——企業真正的挑戰是新人如何累積經驗，因為人工智能正接手過去用來練手的基礎工序。In our view, the most useful judgement here is that being affected does not equal losing a job. Falling hiring demand without layoffs shows the impact hits the entry point rather than the exit: the real challenge is how newcomers accumulate experience, since AI is taking over the foundational tasks once used for training.

對日常工作的影響: 對人力資源部門：入職培訓與見習安排須重新設計，把過去由新人承擔的基礎工序改為有監督的判斷練習，否則三年後將出現中層斷層。For HR: induction and trainee programmes must be redesigned so that foundational tasks formerly given to newcomers become supervised judgement exercises; otherwise a mid-level gap will appear within three years.
如何改善: 本週請人力資源主管挑選一個初級崗位，列出其中三項已被人工智能接手的工序，並設計替代的判斷練習。This week, ask the HR lead to pick one junior role, list three tasks already taken over by AI, and design replacement judgement exercises.

來源新華社（今日頭條轉載）（2026-10-15）↗Second Talent（2026-10-15）↗

本週可做的三件事

盤點手上所有代理的實際權限，逐一寫明「誰可以叫停」與「出事後誰簽字」，並要求供應商在合約內明文確認角色分工。Map the actual permissions held by every agent you run, write down who can stop it and who signs off when it fails, and require suppliers to state their role in the contract in writing.
為每一個正式上線的人工智能系統補一份可對外的文件：用了哪些資料、依據什麼法律基礎、有哪些防護，不要只留供應商的銷售簡報。For every AI system in production, assemble one externally defensible document: what data was used, on what legal basis, with what safeguards, rather than relying on a vendor sales deck.
與財務及採購同事共同檢視算力或雲端合約的到期與調價條款，把「限時優惠價」與「到期後價格」兩種情境分別計入明年的預算。Review compute and cloud contracts with finance and procurement together, and budget separately for the introductory rate and the price that takes effect when it expires.

常見問題

今日（2026-10-16）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、台積電第三季淨利創新高，連續第六個季度破紀錄，七奈米以下先進製程佔晶圓收入七成九；2、阿斯麥客戶開始洽購二〇二九至二〇三〇年設備，行政總裁稱供不應求至少延續至二〇二八年；3、美國八個州同日分別起訴 OpenAI，指 ChatGPT 危害未成年人並要求限制可自我改進模型開發；4、英國標準協會推出三項人工智能評估認證，涵蓋大模型性能測試、風險管理與《人工智能法案》生物識別就緒；5、英國資訊專員辦公室完成十家基礎模型開發商審視，指其合法權益評估與影響評估不足，並就代理式人工智能展開證據徵集；6、OpenAI 以三億新加坡元在新加坡設立美國以外首個應用人工智能實驗室，並計劃招聘逾二百名技術人員；7、Anthropic 開設新加坡辦公室並與新加坡數碼發展及新聞部簽署備忘錄，涵蓋模型評估、事故應對與技能培訓；8、亞洲科技股因數據中心債務憂慮急挫，南韓綜合指數跌逾百分之八，三星電子下挫近一成；9、中共中央與國務院印發發展新質生產力意見，全面實施「人工智能+」行動並強化算力算法數據供給；10、國際貨幣基金組織與世界銀行曼谷年會：約四成職位或受人工智能影響，但更多是結構重塑而非直接裁員。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？

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：Business Day、新浪財經、The Gist、澎湃新聞、財聯社、Resultsense、IamVera.ai、聯合早報），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。