登入 預約諮詢
AiX 環球 AI 情報 Global AI Intelligence
第 53 期 2026 年 11 月 20 日(星期五) 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 20 日算力有了帳單,模型有了價目:今日十條都在重算同一條成本線

今日主編判斷

今日十條圍繞同一問題:智能愈便宜,帳愈難算。英偉達單季營收 1,158 億美元、九成三來自數據中心;國際能源署指數據中心投資將由 5,650 億增至 4 萬億美元,不到一百項目已抵得上英國峰值用電。同日四家大廠再壓詞元價,新加坡金管局則要為代理發身份證。Today’s ten items circle one question: the cheaper intelligence gets, the harder the bill is to read. Nvidia posted $115.8bn in quarterly revenue, 93 per cent from data centres; the IEA sees data-centre investment rising from $565bn to $4trn, with fewer than 100 projects matching Britain’s peak power. Four vendors cut token prices again as Singapore moved to issue agents an identity.

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

英偉達第三季營收 1,158 億美元,第四季指引 1,300 億且不計中國數據中心收入Nvidia posts $115.8bn quarterly revenue and guides $130bn, with no China data-centre revenue assumed

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
英偉達季度營收 1,158.44 億美元、同比增 103%,數據中心 1,076 億佔九成三,下季指引 1,300 億。Nvidia reported $115.844bn of quarterly revenue, up 103 per cent, with $107.6bn from data centres, 93 per cent of the total. It guided the next quarter to $130bn.
背後
這是黃仁勳十一月十二日辭任後首份財報,管理層以訂單與採購承諾代替人事敘事,並稱下季指引不計任何中國數據中心收入。It is the first report since Jensen Huang resigned on 12 November; management leaned on orders and purchase commitments rather than the leadership question, and said guidance assumes no China data-centre revenue.
顧問觀點

我們認為,讀這份財報要看三個數字:九成三收入靠數據中心、毛利下季降至七成一至七成二、採購承諾增至 3,900 億。需求強,但集中度與毛利受壓。We watch three numbers: 93 per cent from data centres, gross margin falling to 71–72 per cent next quarter, and $390bn of purchase commitments. Demand is strong, yet concentration and margin are squeezed.

對日常工作的影響
財務要把單一供應商與單一地區寫進風險附註;IT 採購要為 GPU 報價與交期同時做兩套假設,不能只按今日價格排預算。Finance should disclose single-supplier and single-region risk; IT procurement needs parallel assumptions on price and lead time rather than budgeting at today’s rates.
如何改善
本週向主要雲端服務商索取明年 GPU 與託管報價的區間,按低、中、高三檔重算一次系統容量預算。Ask your main cloud provider for next-year GPU and hosting quote ranges, then re-run your capacity budget at low, middle and high.

國際能源署:數據中心投資 2030 年達 4 萬億美元,不到一百個項目抵得上英國峰值用電IEA: data-centre spending to reach $4trn by 2030 as under 100 projects match Britain’s peak demand

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
國際能源署指,數據中心投資將由 2025 年 5,650 億美元升至 2030 年約 4 萬億,百個項目需 50 吉瓦。The IEA sees data-centre investment rising from $565bn in 2025 to about $4trn by 2030, with fewer than 100 projects needing 50GW.
背後
該署以衛星影像追蹤九十三個 AI 大型數據中心,兩年容量增十倍;電網擴容與變壓器交期以年計,慢過機房建成。The agency tracked 93 large AI data centres by satellite and found capacity up tenfold in two years; grid reinforcement and transformer lead times run to years, behind the pace of construction.
顧問觀點

我們認為,AI 基建的成本正從晶片帳轉到電網帳。在部分增長市場,已公布項目若全數落地,用電已超過當地峰值,這會令上線時間表延後。We see AI infrastructure cost shifting from the chip bill to the grid bill. In some growth markets, announced projects already exceed local peak demand, which pushes commissioning dates back.

對日常工作的影響
IT 不再只看雲服務報價,要問供電與交期;財務要把電費與碳排放列入長期採購,老闆要接受「買到晶片不等於上線」。IT must ask vendors about power and delivery, not just price; finance must put electricity and carbon into long-term procurement, and owners must accept that chips bought are not capacity live.
如何改善
本週列出公司三個最依賴算力的系統,逐一寫上服務商、所在地區與續約時點,交給財務做情境演算。List your three most compute-hungry systems with provider, region and renewal date, and hand the page to finance for scenario testing.

OpenAI 推 GPT-6.2 Astra 與 Sol 雙層定價,Astra 為首個觸及自我改進高門檻的對外模型OpenAI splits GPT-6.2 into Astra and Sol, with Astra its first released model rated High for self-improvement

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 發布 GPT-6.2 Astra 與 Sol,Astra 每百萬詞元輸入 8 美元、輸出 40 美元。OpenAI released GPT-6.2 Astra and Sol; Astra costs $8 input and $40 output per million tokens, a 20 per cent cut.
背後
同一代分成旗艦與日常兩層,回應企業「貴模型做難事、便宜模型做量」的實際用法,也趕在競爭對手上市當日搶注意力。Splitting one generation into a flagship and an everyday tier mirrors how firms work: expensive models for hard tasks, cheap ones for volume, and the launch landed on a rival’s listing day.
顧問觀點

我們認為,模型選擇應由任務成本決定,而非排行榜。Astra 自評觸及自我改進高門檻,企業須同時要求供應商交代失控時的監控與中止安排。We think model choice should follow task cost, not leaderboards. With Astra rated High for self-improvement, buyers should also demand the vendor’s monitoring and shutdown arrangements.

對日常工作的影響
IT 要為每個部門設定可用模型層級與月度上限;法務要在合約加入模型版本變更通知;市場部可按任務分流,省下的預算用於評測。IT should set permitted model tiers and monthly caps per department; legal should require notice of model changes; marketing can route tasks to cheaper tiers and spend the saving on evaluation.
如何改善
本週挑一個高頻任務,用旗艦與日常兩層各跑十次,記錄完成率與單次成本,作為下季採購依據。This week, run one high-frequency task ten times on each tier and record completion rate and cost per run as next quarter’s procurement evidence.

谷歌發布 Gemini 4 Neon,每百萬詞元 3 與 15 美元,主打編程代理Google releases Gemini 4 Neon at $3 and $15 per million tokens, pitched at coding agents

影響力✓ 已核實Verified已在發生Happening now資訊科技IT市場推廣Marketing
事件
谷歌十一月十八日推出 Gemini 4 Neon,每百萬詞元 3 美元輸入、15 美元輸出,支援百萬詞元上下文。Google released Gemini 4 Neon on 18 November at $3 input, $15 output and $0.15 cached input per million tokens, with a one-million-token context and four thinking levels.
背後
Neon 夾在 Carbon 與 Hydrogen 之間,谷歌稱預設檔位單任務成本較 Carbon 低三成二。Neon sits between Carbon and Hydrogen; Google says it costs 32 per cent less per task than Carbon by default, with cached input an order of magnitude cheaper.
顧問觀點

我們認為,中階模型正成為代理工作量主力。分水嶺不在跑分,而在緩存定價:重複讀取同一批文件的工作流,成本差距遠大於表面價差。We see mid-tier models becoming the workhorse for agent workloads. The dividing line is cached-input pricing: workflows that re-read the same documents save far more than headline rates suggest.

對日常工作的影響
IT 要把「上下文重用率」列入模型評估;市場與客服可把長文件問答搬到緩存計價較低的模型,直接壓低月費。IT should add context-reuse rate to model evaluation; marketing and service teams can move long-document Q&A onto cheaper cached input and cut monthly spend directly.
如何改善
本週統計三個最常重複讀取的文件集,量出每月被讀次數,據此比較各模型的緩存單價。Count how often your three most-reused document sets are read each month, then compare cached-input prices across models on that volume.

微軟 Ignite 發布七款 MAI 模型,Maia 200 芯片投產並稱每瓦詞元提升 1.7 倍Microsoft launches seven MAI models at Ignite as its Maia 200 chip enters production

影響力✓ 已核實Verified已在發生Happening now資訊科技IT企業決策者Business owners
事件
微軟 Ignite 發布七款 MAI 模型,涵蓋編程與安全;自研加速器 Maia 200 投產,每瓦詞元升 1.7 倍。At Ignite Microsoft released seven MAI models for coding and security, and said its Maia 200 accelerator is in production at 1.7 times more tokens per watt.
背後
微軟為每款模型標明對標對象與價格比例,把競爭焦點由能力榜轉向每元產出,同時把自研芯片定位為推理降本工具。Microsoft named a benchmark rival and a price ratio for each model, moving the contest from capability leaderboards to output per dollar, and positioned its own silicon as an inference cost tool.
顧問觀點

我們認為,真正的含義是議價:當同一雲端同時提供自研與第三方模型,客戶可以要求按任務競價,而不是整包續約。The real meaning is bargaining power: when one cloud offers both first-party and third-party models, customers can ask for per-task competition instead of a single renewal.

對日常工作的影響
採購要把合約由年度席位改為按任務計價;IT 要測試自研芯片承載的推理是否影響延遲;HR 可據此重算自動化崗位的單位成本。Procurement should move contracts from seats to per-task pricing; IT should test latency on first-party silicon; HR can re-price automation per role.
如何改善
本週在現有雲端帳戶開一個小測試環境,用同一任務比較自研與第三方模型的成本與延遲各一次。Open a small test environment in your existing cloud account and run one identical task on first-party and third-party models, measuring cost and latency once each.

騰訊開源混元 Hy4 flash:總參 2,240 億、推理僅激活 120 億,首日調用 50 萬億詞元Tencent open-sources Hunyuan Hy4 flash with 224bn parameters and 12bn active at inference

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT法務合規Legal & compliance
事件
騰訊混元十一月十八日開源 Hy4 flash,總參 2,240 億、推理僅激活 120 億,首日調用 50 萬億詞元。Tencent open-sourced Hunyuan Hy4 flash on 18 November: 224 billion parameters with 12 billion active at inference, multimodal input, a one-million-token context and 50 trillion tokens drawn on day one.
背後
稀疏激活把推理成本壓到只算約百分之五的參數,配合開源授權,讓企業可在自有環境部署而毋須長期租用 API。Sparse activation means inference pays for only a small share of the parameters, and the open licence lets firms deploy on their own premises rather than renting an API indefinitely.
顧問觀點

我們認為,對資料不能出境的企業,開源權重是唯一可行路徑。但自建要算上維運人力與安全更新,不能只比詞元單價。For firms whose data cannot leave their premises, open weights are the only workable route. But self-hosting must include staff and security updates, not just a token price comparison.

對日常工作的影響
IT 要評估自有伺服器能否承載長上下文推理;法務要審視開源授權條款;內地業務線可先試點,不必等總部雲端審批。IT must check whether in-house servers can handle long-context inference; legal must read the licence; mainland business units can pilot without waiting for head-office approval.
如何改善
本週在隔離環境部署一次該模型,用一份內部常見文件測試答案準確度與單次耗時。Deploy the model once in an isolated environment this week and test accuracy and time per response on a document your team handles daily.

新加坡金管局推「認識你的代理」共同協議,並設沙盒測試代理調撥代幣化存款MAS plans a common Know-Your-Agent protocol and a sandbox for agents moving tokenised deposits

影響力✓ 已核實Verified6–12 個月6–12 months法務合規Legal & compliance企業決策者Business owners
事件
新加坡金管局總裁謝德俊十一月十九日表示,將制定共同協議核實代理的身份、代表權與授權範圍,並設沙盒測試代理調撥代幣化存款。MAS managing director Chia Der Jiun said on 19 November that it will develop a common protocol verifying an agent’s identity, representation and mandate, and set up a sandbox for agents moving tokenised deposits.
背後
代理開始代客戶搜尋、比較、議價與付款,金融機構需要可查核的授權憑證,否則無法界定責任與限額。Agents now search, compare, negotiate and pay on a client’s behalf, so banks need checkable proof of mandate before they can set limits or assign liability.
顧問觀點

我們認為,這是把代理正式列為交易對手的監管起點。企業日後與銀行打交道,可能要為每個代理出具授權書,並接受審計。We read this as regulators formally treating agents as counterparties. Firms may soon need a written mandate per agent for their bank, plus an audit trail.

對日常工作的影響
財務與法務要為自動付款代理設定金額上限與覆核人;IT 要保留授權紀錄;客服要準備回應「為何機械人替我付款」。Finance and legal need amount caps and a checker for any paying agent; IT must retain mandate records; customer service needs an answer for “why did a bot pay for me”.
如何改善
本週盤點現有自動付款或自動下單流程,為每個流程補一份授權範圍與中止方式的書面紀錄。Inventory every automated payment or ordering flow this week and attach a written record of scope and stop mechanism to each.

工信部一次實施五項人形機器人標準,智能化能力分 L0 至 L5China’s industry ministry puts five humanoid-robot standards into force, grading intelligence from L0 to L5

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
工信部十一月十九日發布第 30 號公告,五項人形機器人標準即日實施,其中智能化能力按自主性分 L0 至 L5 六級。China’s industry ministry issued announcement No. 30 on 19 November, putting five humanoid-robot standards into force that day, with intelligence graded L0 to L5 by autonomy.
背後
行業長期缺乏統一評價基準,買家不敢用亦難比較;標準另涵蓋訓練場、訓練數據、一體化關節與靈巧手,補上選型依據。The sector lacked a common yardstick, leaving buyers unable to compare machines; the package also covers training grounds, training data, integrated joints and hands, giving a basis for selection.
顧問觀點

我們認為,分級標準的真正作用是採購語言。日後買機械人,可以要求供應商出示等級與能力輪廓,而不是只看示範影片。The practical effect is a procurement vocabulary. Buyers can demand a grade and a capability profile instead of judging from a demo video.

對日常工作的影響
採購要把等級寫進規格書;廠務要按標準重新設計訓練與驗收流程;HR 要據此界定操作員與維護員的技能要求。Procurement should put the grade into specifications; plant operations should redraw training and acceptance around it; HR can set operator and maintenance skills against it.
如何改善
本週向現有或擬採購的機械人供應商索取等級證明與能力輪廓,附在採購檔案內。Ask current or prospective robot suppliers for grade evidence and a capability profile this week, and file it with the procurement record.

國家數據局啟動第二批可信數據空間試點,明確以「與人工智能雙向賦能」為突破口China’s data administration opens a second round of trusted data space pilots tied to AI

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
國家數據局近日印發通知,啟動第二批可信數據空間試點,推進企業、行業、城市三類空間,以與人工智能雙向賦能為突破口。China’s National Data Administration has issued a notice opening a second round of trusted data space pilots across enterprise, industry and city categories, with two-way enablement with AI as the stated breakthrough.
背後
內地推動行業數據互通同時要守住安全底線,可信數據空間正是讓數據「可用不可見」的制度載體,本輪明確對接 AI 需求。Beijing wants industry data to flow while keeping security controls; trusted data spaces are the vehicle for usable-but-not-visible sharing, and this round explicitly links them to AI needs.
顧問觀點

我們認為,試點的真正門檻是規則互信,而非技術平台。企業若要參與,先要能回答數據誰有權用、用到哪裡、如何退出。The real barrier is mutual trust in the rules, not the platform. Firms wanting in must first answer who may use which data, for what, and how to exit.

對日常工作的影響
法務要準備數據授權與退出條款;IT 要盤點可共享的數據資產;業務部門要明確以哪些場景換取參與資格。Legal needs authorisation and exit clauses; IT must inventory shareable data assets; business units must name the use cases they would trade for participation.
如何改善
本週選一項可對外共享的數據資產,先寫出用途、權限與退出三項條件,作為參與試點的第一份文件。Pick one data asset you could share externally and write its permitted use, access rights and exit terms as the first document for any pilot application.

新加坡金融科技調查:八成六已部署 AI,僅百分之四視增收為最大效益Singapore fintech survey: 86 per cent have deployed AI, but only 4 per cent cite new revenue as the main benefit

影響力✓ 已核實Verified已在發生Happening now市場推廣Marketing企業決策者Business owners
事件
新加坡金融科技協會調查一百名當地金融科技主管,八成六已部署 AI、九成四明年加預算,僅百分之四稱最大效益來自增收。A survey of 100 Singapore fintech leaders by the Singapore FinTech Association, Airwallex and Kapronasia found 86 per cent have deployed AI and 94 per cent plan bigger budgets next year, but only 4 per cent cite new revenue.
背後
部署率高於當地金融機構六成五與全球金融科技六成二,說明工具已普及,但多數仍停留在降本增效,未走到定價與新產品。Deployment runs ahead of Singapore financial institutions at 65 per cent and global fintechs at 62, so tools are widespread, but most use remains cost and efficiency rather than pricing or new products.
顧問觀點

我們認為,這個落差不是技術問題,是衡量問題。當指標只記省下多少工時,AI 永遠不會被用來創造收入。The gap is a measurement problem, not a technology one. While metrics only record hours saved, AI will never be pointed at revenue.

對日常工作的影響
市場部要把 AI 專案綁定至少一個收入指標;財務要為試點設營收與成本雙軌記帳;老闆要在季度會追問收入歸因。Marketing should tie each AI project to at least one revenue metric; finance should book pilots on two tracks; owners should demand revenue attribution each quarter.
如何改善
本週為手上每個 AI 專案補一欄「預期收入影響」,寫不出數字的專案改列為效率項目。Add an expected-revenue column to every AI project you run this week, and reclassify as efficiency work anything that cannot be filled in.

本週可做的三件事

  1. 把過去一個月的模型帳單按任務拆開,算出每個任務的單次成本,作為下次換模型的基準線。Break last month’s model bill down by task and compute a cost per task, then use it as the baseline for the next model switch.
  2. 向資訊部門索取一份數據中心或雲端用電與續約時點清單,把電費風險寫進明年預算。Ask IT for a one-page list of data-centre or cloud power exposure and renewal dates, and put the electricity risk into next year’s budget.
  3. 挑一個由代理代為操作的工作流程,補上「誰授權、可做什麼、如何中止」三行書面紀錄。Pick one workflow an agent already performs and write three lines: who authorises it, what it may do, and how to stop it.

常見問題

今日(2026-11-20)「AiX 環球 AI 情報」有哪十條重點?

今日十條為:1、英偉達第三季營收 1,158 億美元,第四季指引 1,300 億且不計中國數據中心收入;2、國際能源署:數據中心投資 2030 年達 4 萬億美元,不到一百個項目抵得上英國峰值用電;3、OpenAI 推 GPT-6.2 Astra 與 Sol 雙層定價,Astra 為首個觸及自我改進高門檻的對外模型;4、谷歌發布 Gemini 4 Neon,每百萬詞元 3 與 15 美元,主打編程代理;5、微軟 Ignite 發布七款 MAI 模型,Maia 200 芯片投產並稱每瓦詞元提升 1.7 倍;6、騰訊開源混元 Hy4 flash:總參 2,240 億、推理僅激活 120 億,首日調用 50 萬億詞元;7、新加坡金管局推「認識你的代理」共同協議,並設沙盒測試代理調撥代幣化存款;8、工信部一次實施五項人形機器人標準,智能化能力分 L0 至 L5;9、國家數據局啟動第二批可信數據空間試點,明確以「與人工智能雙向賦能」為突破口;10、新加坡金融科技調查:八成六已部署 AI,僅百分之四視增收為最大效益。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?

今日十條圍繞同一問題:智能愈便宜,帳愈難算。英偉達單季營收 1,158 億美元、九成三來自數據中心;國際能源署指數據中心投資將由 5,650 億增至 4 萬億美元,不到一百項目已抵得上英國峰值用電。同日四家大廠再壓詞元價,新加坡金管局則要為代理發身份證。

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:證券時報(新浪財經轉載)、騰訊新聞、界面新聞(騰訊新聞轉載)、36氪(騰訊新聞轉載)、新浪財經、鳳凰網科技(IT之家)、每日經濟新聞(騰訊新聞轉載)、TechUpscale),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。

每週一封 · 可隨時取消 · 私隱聲明