AiX 環球 AI 情報 Global AI Intelligence

第 37 期 2026 年 11 月 3 日（星期二） 07:30 出刊 10 條情報 · 17 原文來源連結 RSS 過往期數

今日十條 · 2026 年 11 月 3 日制度先於能力到位：電網、監管與資金條件同時收緊，模型競賽讓位給算力與合規的帳單

今日主編判斷

今日十條指向同一件事：制約人工智能的不再是模型分數，而是電、錢與規則。OpenAI 把讓電寫成全美計劃，Anthropic 首披八千九百五十四億美元算力承諾，韓國加碼二十一萬三千億韓圜押注國產算力；香港證監會與英國監管同時把問責指向人。企業要算的是能否長期供電、供資與負責。What constrains AI is no longer benchmark scores but power, capital and rules. OpenAI made curtailment a national programme, Anthropic disclosed a $895.4 billion compute commitment, and Korea pledged KRW 21.3 trillion to domestic compute, while Hong Kong's SFC and UK regulators pushed accountability back onto people. Companies must now show power, funding and responsibility for years.

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

OpenAI 啟動「電網守護者計劃」，承諾美國園區高峰時讓出一吉瓦用電OpenAI starts its Grid Protectors Program, extending a one-gigawatt curtailment pledge nationwide

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: OpenAI 十一月二日啟動「電網守護者計劃」，由佐治亞州一園區推至全美，電網告急時最多減一吉瓦。OpenAI launched its Grid Protectors Program on 2 November, extending a pledge made at one Georgia campus to company-built campuses nationwide, with up to one gigawatt of load shed in grid emergencies.
背後: 選址已由供電審批轉為社區許可。美國今年多個項目因居民反對延誤，削峰寫進並網條件，成為談判籌碼。Site approval now hinges on community consent more than on utility supply. Several US projects were delayed by local opposition this year, so curtailment written into interconnection terms became a bargaining chip.

顧問觀點

我們認為算力正由稀缺資產變成可調度負荷。當電網可要求園區讓電，自建或租用的算力不再保證全天可用，可用率將寫入服務協議。We read this as compute turning from scarce asset into dispatchable load. Once a grid can require a campus to shed power, the compute a firm builds or rents is no longer guaranteed around the clock, and availability will move into service agreements.

對日常工作的影響: IT 主管要分層關鍵與非關鍵任務，並在雲端合約追問限流與賠償；設施主管須預留現場儲能預算。IT leads should tier critical against non-critical workloads and ask cloud contracts about throttling and credits; facilities leads must budget for on-site storage.
如何改善: 本週列出三個對延遲最敏感的業務流程，向主要雲端供應商索取高峰限流的書面說明，並測試一次降載演練。This week, name the three most latency-sensitive processes, request written throttling terms from your main cloud provider, and run one load-shedding drill.

來源新浪財經（環球市場播報）（2026-11-02）↗IT時代網（2026-11-03）↗

Anthropic 修訂招股書：上季轉盈為虧，算力承諾達八千九百五十四億美元Anthropic's revised filing: a return to operating loss and $895.4 billion of compute commitments

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners法務合規Legal & compliance

事件: Anthropic 十一月二日提交修訂招股書，第三季營收一百六十九點五六億美元，經營虧損五點六七億。Anthropic filed a revised prospectus on 2 November: third-quarter revenue of $16.956 billion against a $567 million operating loss.
背後: 長約鎖定算力，但需求超出額度，須臨時加購五十億美元短期算力，每元算力收入因而下降。Long-term contracts reserved capacity, but demand outran it, forcing about $5 billion of pricier short-term compute and lowering revenue per dollar of compute spending.

顧問觀點

我們認為這是本輪最清楚的成本披露。模型價格持續下行，成本卻被多年合約鎖死，利潤將取決於誰能把短期算力換成長期產能。We see this as the clearest cost disclosure of the cycle. Prices keep falling while costs are locked into multi-year contracts, so margin depends on converting expensive short-term compute into long-term capacity.

對日常工作的影響: 財務主管要把長約承諾視同租賃負債納入風險評估；採購須爭取算力價格隨市價調整，而非只談折扣。Finance leads should treat a vendor's long-term commitments as lease-like liabilities in risk review; procurement should seek market-indexed pricing, not just discounts.
如何改善: 本週把主要模型供應商的承諾金額、還款年期與價格掛鈎方式整理成一張表，交風險委員會審閱一次。This week, tabulate your main model vendors' committed amounts, payment horizons and price-indexing terms, and take the table to your risk committee once.

來源華爾街見聞（2026-11-03）↗中央社（2026-11-03）↗

韓國明年預算投向半導體與實體 AI，三大項目加碼至二十一萬三千億韓圜Korea's 2027 budget puts KRW 21.3 trillion into chips, physical AI and data centres

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

事件: 李在明十一月二日向國會提交二〇二七年度預算，半導體、實體 AI 與數據中心獲二十一萬三千億韓圜。President Lee Jae-myung submitted the 2027 budget on 2 November, with KRW 21.3 trillion for semiconductors, physical AI and AI data centres.
背後: 韓國十月半導體出口按年增九成三七，為加碼提供財力。政府以合約學系綁定企業用人。Korean semiconductor exports rose 93.7% year on year in October, funding the expansion, while contract academic departments tie talent supply to employers.

顧問觀點

我們認為這是亞洲版產業政策的轉向：由補貼研發走向直接購買算力與人才。對區內企業而言，韓國的公共算力與合約學系會成為招聘與選址的競爭變數。We read this as an Asian shift from subsidising R&D to buying compute and talent directly. Korea's public compute and contract departments will become variables in regional hiring and siting.

對日常工作的影響: HR 主管要把韓國校園招聘與當地合約學系對齊；老闆須評估把推理節點放在韓國的電價與補貼條件。HR leads should align Korean campus hiring with local contract departments; owners should test whether Korean power prices and subsidies beat current locations.
如何改善: 本週向韓國當地合作夥伴索取公共算力申請條件與合約學系名單，作為下年度招聘與選址的比較依據。This week, ask a Korean partner for public-compute application terms and the list of contract departments, to benchmark next year's hiring and siting.

來源科創板日報（騰訊新聞）（2026-11-02）↗鳳凰網科技（騰訊新聞）（2026-11-02）↗

香港證監會首份調查：三成大機構已用代理型 AI，將訂原則為本指引Hong Kong's SFC finds 30% of large institutions run agentic AI and will issue principles-based guidance

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資訊科技IT

事件: 香港證監會十一月二日發表首份大型機構 AI 調查，三成已用代理型 AI，兩成半試點，九成五擬加快。Hong Kong's SFC published its first large-institution AI survey on 2 November: 30% run agentic AI, 25% are piloting, and 95% expect faster adoption within a year.
背後: 監管承認逐步人手審批不切實際，改按風險相稱。可撤回的後台操作放寬，交易執行仍從嚴。The regulator accepted that reviewing every step by hand is impractical and moved to proportionality: reversible back-office work gains autonomy, while advice and execution stay strict.

顧問觀點

我們認為以人問責將由每步批核，改為權限設計。企業要交的不是更多簽名，而是代理可讀哪些資料、可動多少金額、誰在何時介入。We expect accountability to move from per-step approval to authority design. Firms must produce not more signatures but records of which data an agent may reach, what amounts it may move and who intervenes.

對日常工作的影響: 合規主管要按可撤回程度為代理分級並設金額上限；IT 須記錄每個工具調用，以便還原責任鏈。Compliance leads should tier agents by reversibility with monetary caps; IT must log every tool call so the chain of responsibility can be reconstructed.
如何改善: 本週挑一條已上線的代理流程，補上權限清單與金額上限兩個欄位，並指定一名具名的問責主管。This week, take one live agent workflow, add a permissions list and a monetary cap, and name a single accountable manager.

來源央廣網（2026-11-02）↗上海證券報（2026-11-02）↗

延伸相關課程Related course

美團開源路線再進一步：一萬六千億參數模型全程以國產算力訓練Meituan releases LongCat-2.5, a 1.6-trillion-parameter model trained entirely on domestic compute

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: 美團十一月二日發布 LongCat-2.5，一萬六千億參數，全程以國產算力訓練，面向長程代理任務。Meituan released LongCat-2.5 on 2 November: 1.6 trillion parameters, trained entirely on domestic compute and built for long-horizon agent tasks.
背後: 國產卡顯存小，瓶頸在互聯。美團把主要流量收進超節點，並以四比特量化換取四成吞吐提升。Domestic accelerators hold less memory per card, so interconnect is the bottleneck. Meituan moved main traffic onto super-node links and used four-bit inference for about 40% more throughput.

顧問觀點

我們認為值得注意的不是分數，而是全鏈路可在非美系算力上跑通。這意味着供應鏈受限地區的企業，終於有一個可在本地部署、可審計的萬億級選項。What matters is not the scores but a full pipeline running on non-US compute. Firms in constrained jurisdictions now have a trillion-scale option they can deploy locally and audit.

對日常工作的影響: IT 主管要為本地部署評估機房電力與顯存配置；數據主管須確認權重與推理日誌留在境內。IT leads should size data-centre power and memory for local deployment; data leads must confirm weights and inference logs stay in-country.
如何改善: 本週在一條內部知識庫問答上試用 LongCat-2.5 接口，記錄準確度與單次成本。This week, test the LongCat-2.5 API on one internal knowledge-base Q&A task, recording accuracy and cost per inference.

來源美團（2026-11-02）↗環球網（騰訊新聞）（2026-11-02）↗

土耳其競爭管理局就谷歌搜尋 AI 功能立案，並先行實施臨時監管措施Turkey opens a probe into Google's AI search features, citing exclusion of publishers

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

事件: 土耳其競爭管理局十一月二日立案查谷歌「AI 概覽」與「AI 模式」，指其對出版商有排除之嫌。Turkey's Competition Authority opened an investigation on 2 November into Google's AI Overviews and AI Mode, citing exclusionary treatment of publishers.
背後: 監管焦點由排序轉向摘要即答案。歐盟與英國同月處理同類爭議，臨時措施往往先行於裁決。Enforcement is shifting from ranking to answer summaries and their effect on traffic. The EU and UK are handling parallel cases, and interim measures often precede rulings.

顧問觀點

我們認為內容供應方首次取得程序上的槓桿。但臨時措施只改變談判位置，不改變商業模式；出版商要的是可計價的引用，而非更多聲明。We see content suppliers gaining procedural leverage for the first time. Interim measures shift bargaining position, not the business model; publishers need priced citations, not statements.

對日常工作的影響: 市場主管要重估自然流量跌幅，把預算移向可歸因渠道；法務須檢查內容授權是否涵蓋 AI 摘要用途。Marketing leads should re-forecast organic traffic decline and shift budget to attributable channels; legal must check whether content licences cover AI summaries.
如何改善: 本週抽三日比對官網自然流量與 AI 摘要出現次數，估算被引用未付費的比例，作為談判依據。This week, compare three days of organic traffic against the number of AI summaries citing you, estimating the unpaid citation share as a negotiating basis.

來源IT之家（騰訊新聞）（2026-11-02）↗

AI 就業文件進入行業細化階段，廣電主管部門提出先評估再推廣China's AI employment policy reaches sector level, with an assessment duty before rollout

影響力✓ 已核實Verified6–12 個月6–12 months人力資源HR企業決策者Business owners法務合規Legal & compliance

事件: 廣東廣電局十一月解讀國辦發二十九號文件：部署十八項任務，高衝擊場景須同步做就業風險評估。Guangdong's broadcast regulator published a reading of State Council document 29 in November: 18 tasks, with employment-risk assessment required for high-impact applications.
背後: 文件十月下旬公布，本月起由行業部門逐個落地。內容生產首個被點名拆解，替代與新增崗位同時出現。Issued in late October, the document is now being unpacked sector by sector. Content production is the first named case, where replacement and new roles appear together.

顧問觀點

我們認為企業將要交出「就業影響說明」，而不只是效率報告。這會把人力資源部門拉進 AI 項目立項階段，改變過去由 IT 單獨決策的慣例。We expect firms to submit employment-impact statements, not just efficiency reports. That pulls HR into project approval and breaks the habit of IT deciding alone.

對日常工作的影響: HR 主管要在立項表加一欄受影響崗位與轉崗安排；部門主管須在推廣前書面告知員工，避免爭議。HR leads should add an affected-roles and redeployment field to project forms; line managers should notify affected staff in writing before rollout.
如何改善: 本週選一個準備上線的 AI 流程，補寫一頁就業影響說明，列出受影響職位與轉崗路徑，存檔備查。This week, pick one AI workflow due for rollout, write a one-page employment-impact note listing affected roles and redeployment paths, and file it.

來源人民日報（騰訊新聞轉載）（2026-11-02）↗

延伸相關課程Related course

摩根大通與亞馬遜限制第三方 AI 代理接入，另有三十七家零售商無意中攔截JPMorgan and Amazon restrict third-party AI agents, while 37 retailers block them by accident

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

事件: 據 CNBC 報道，摩根大通禁第三方 AI 代理操作零售賬戶，亞馬遜要求 Meta 移出購物鏈路。CNBC reported that JPMorgan bars third-party AI agents from retail accounts and that Amazon asked Meta to drop its marketplace from Muse's shopping path.
背後: 代理代用戶點擊，令曝光、數據與詐騙責任失焦。沃爾瑪與 Shopify 開放接入，金融機構選擇關門。Agents clicking for users unsettle impressions, data and fraud liability at once. Walmart and Shopify open their platforms; financial institutions close the door.

顧問觀點

我們認為企業官網將成為下一個戰場。若網站無法辨識並授權代理，企業會同時失去代理帶來的成交與監管要求的可追溯性，兩邊都不討好。We think corporate websites are the next battleground. Without identifying and authorising agents, firms lose both the sales agents bring and the traceability regulators expect.

對日常工作的影響: IT 主管要為網站加上代理識別與日誌；市場主管須重算流量結構，區分代理詢盤，避免轉化率失真。IT leads should add agent identification and action logs to the website; marketing must re-segment traffic between humans and agents so conversion rates stay meaningful.
如何改善: 本週檢查網站的機械人防護規則，確認沒有誤攔主流代理，並為代理流量另設一條帶日誌的入口。This week, audit your site's bot rules for accidental blocking of mainstream agents and open a separate, logged entry path for agent traffic.

來源環球網（2026-11-02）↗

度小滿與港科大發布軒轅 4.5 金融模型，稱成本僅通用模型十分一Du Xiaoman and HKUST release Xuanyuan 4.5, a finance model pitched at a tenth of general-model cost

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners法務合規Legal & compliance資訊科技IT

事件: 度小滿與港科大、香港生成式人工智能研發中心十一月二日發布軒轅 4.5 金融模型。Du Xiaoman, HKUST and the Hong Kong Generative AI R&D Center released the Xuanyuan 4.5 finance model on 2 November.
背後: 金融機構受制於資料出境，難直接用海外通用模型。與本地高校共建行業模型，成為折中路徑。Banks are bound by data-residency and explainability rules, which makes overseas general models hard to adopt; co-building local industry models is the middle path.

顧問觀點

我們認為行業模型的分水嶺不在分數，而在可否解釋與審計。若供應商能出示每一決策所依據的因子與版本，金融機構的採購才會由試點轉為採購。The dividing line for industry models is explainability and auditability, not scores. Only vendors who can show the factors and version behind each decision will move banks from pilot to purchase.

對日常工作的影響: 風控主管要把模型版本與因子清單納入審批檔案；採購須在合約要求供應商每年提供外部評測結果。Risk leads should file model versions and factor lists with approvals; procurement should require an annual external evaluation in the contract.
如何改善: 本週要求供應商示範一次審批決策的完整還原流程，記錄其可解釋程度，作為下輪採購的評分項。This week, ask a vendor to reconstruct one approval decision end to end and record its explainability as a scoring item for the next procurement round.

來源鳳凰網科技（2026-11-02）↗經濟觀察網（2026-11-02）↗

世界互聯網大會烏鎮峰會開幕，主題定為開源開放、共建共享World Internet Conference opens in Wuzhen under an open-source, shared-development theme

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: 世界互聯網大會烏鎮峰會十一月二日至五日在浙江舉行，主題為開源開放、共建共享。The 2026 World Internet Conference runs in Wuzhen, Zhejiang, from 2 to 5 November under the theme of open source, openness and shared development.
背後: 開源被推上主場，反映競爭由單點突破轉向生態之爭。國產廠商把峰會當作旗艦模型、標準與生態的發佈場。Open source on the main stage reflects a shift from single-model breakthroughs to ecosystem competition, with domestic vendors timing flagship releases to the summit.

顧問觀點

我們認為對企業最有用的不是在場展示，而是隨之而來的授權與部署條件。開源權重能帶來的，是可審計與可私有化，而非免費。The useful part for enterprises is the licensing and deployment terms that follow. Open weights buy auditability and private deployment, not zero cost.

對日常工作的影響: IT 主管要把開源授權條款、可商用範圍與維護成本並列比較；老闆須把內部模型維護人力計入三年預算。IT leads should compare open-weight licences, commercial-use scope and maintenance cost side by side; owners should budget three years of maintenance headcount.
如何改善: 本週把兩個候選開源權重的授權條款與部署成本並列成一頁，交給技術與法務共同評分。This week, put the licences and deployment costs of two candidate open weights on one page and have technology and legal score them together.

來源世界互聯網大會（2026-11-02）↗新浪財經（錢江晚報）（2026-11-02）↗

本週可做的三件事

把三份最貴的模型或雲端合約上的承諾金額、到期日與退出罰則抄成一頁，交給財務與法務共同簽字。Copy the committed amounts, expiry dates and exit penalties of your three most expensive model or cloud contracts onto one page, and have finance and legal sign it jointly.
為已在跑的一個代理寫一頁權限與問責表：它可以做什麼、誰批准、出事由誰簽名，並設一個可撤回的操作上限。For one agent already in production, write a one-page authority and accountability sheet: what it may do, who approves, who signs when it fails, and set a reversible limit.
向業務部門收集未來十二個月的用電、算力與人力需求，據此調整預算與採購時序，而非按上半年增速外推。Collect next-twelve-month demand for power, compute and headcount from business units, then re-sequence budgets and procurement instead of extrapolating from first-half growth.

常見問題

今日（2026-11-03）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、OpenAI 啟動「電網守護者計劃」，承諾美國園區高峰時讓出一吉瓦用電；2、Anthropic 修訂招股書：上季轉盈為虧，算力承諾達八千九百五十四億美元；3、韓國明年預算投向半導體與實體 AI，三大項目加碼至二十一萬三千億韓圜；4、香港證監會首份調查：三成大機構已用代理型 AI，將訂原則為本指引；5、美團開源路線再進一步：一萬六千億參數模型全程以國產算力訓練；6、土耳其競爭管理局就谷歌搜尋 AI 功能立案，並先行實施臨時監管措施；7、AI 就業文件進入行業細化階段，廣電主管部門提出先評估再推廣；8、摩根大通與亞馬遜限制第三方 AI 代理接入，另有三十七家零售商無意中攔截；9、度小滿與港科大發布軒轅 4.5 金融模型，稱成本僅通用模型十分一；10、世界互聯網大會烏鎮峰會開幕，主題定為開源開放、共建共享。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？

今日十條指向同一件事：制約人工智能的不再是模型分數，而是電、錢與規則。OpenAI 把讓電寫成全美計劃，Anthropic 首披八千九百五十四億美元算力承諾，韓國加碼二十一萬三千億韓圜押注國產算力；香港證監會與英國監管同時把問責指向人。企業要算的是能否長期供電、供資與負責。

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：新浪財經（環球市場播報）、IT時代網、華爾街見聞、中央社、科創板日報（騰訊新聞）、鳳凰網科技（騰訊新聞）、央廣網、上海證券報），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。