01
OpenAI 推 GPT-6 Sol 與 Luna，API 價格較上代減半OpenAI ships GPT-6 Sol and Luna with API prices halved
★★★★☆已在發生Happening now企業決策者Business owners資訊科技IT前線員工Frontline staff
- 事件
- OpenAI 在 GPT-6 系列下新增 Sol 與 Luna 兩個型號，官方稱快取與推理效率改善，令 API 定價較 GPT-5.6 推廣價低 50%，主打大規模工作負載。OpenAI has added two models to its GPT-6 line, Sol and Luna. The company says better caching and inference efficiency let it cut API pricing by 50% against GPT-5.6 promotional rates, aimed at large-scale workloads.
- 背後
- 前沿模型的性能差距正在收窄，價格成為新的分水嶺。當每百萬 token 的價差可以達數十倍，企業選型的第一條問題已經由「哪個最聰明」變成「哪個單位成本下交付得最多」。推理成本下降亦意味著過去因太貴而做不成的自動化流程，現在在數學上成立了。Performance gaps between frontier models are narrowing, and price is becoming the new dividing line. When the spread per million tokens can reach several tens of times, the first question in enterprise selection has moved from which model is smartest to which delivers most per unit of cost. Cheaper inference also means workflows that were arithmetically impossible a year ago are now viable.
- 顧問觀點
- 我們認為價格下調不是好消息本身，而是門檻重設。真正的分野在於：同一間公司有沒有能力把省下來的成本，換成多跑的流程。省價而不改流程，等於只賺了一次折扣。We read a price cut not as the news itself but as a reset of the threshold. The real difference is whether a company can convert the savings into processes it now runs. A discount taken without changing any workflow is a one-off rebate, nothing more.
- 對日常工作的影響
- 原本因為推理費用而被否決的自動化項目——例如全量客戶郵件分類、每日全店商品描述重寫——現在值得重新計一次數。財務與業務的對話會由「太貴」變成「你打算怎麼用」。Automation projects previously rejected on inference cost — classifying every customer email, rewriting every product description daily — deserve a fresh calculation. The conversation between finance and the business shifts from "too expensive" to "what will you do with it".
- 如何改善
- 本週挑一個曾經因成本被否決的流程，用新價格重算一次年度成本，並寫明如果成本減半、你會不會批准。這一頁紙就是下季度的自動化路線圖。This week, pick one process that was killed on cost. Recalculate its annual cost at the new prices, and write down whether you would approve it if the cost halved. That single page is next quarter's automation roadmap.
02
Anthropic 發布 Claude Opus 5.5，運行成本較前代低四成Anthropic releases Claude Opus 5.5 at roughly 40% lower running cost
★★★★☆已在發生Happening now資訊科技IT企業決策者Business owners
- 事件
- Anthropic 推出 Claude 5.5 家族首個型號 Opus 5.5，官方稱在多數任務上追平 Claude Fable 5.1，但運行成本較前代 Opus 5 低約 40%。Anthropic has released Opus 5.5, the first model in its Claude 5.5 family. The company says it matches Claude Fable 5.1 on most tasks while costing about 40% less to run than its predecessor, Opus 5.
- 背後
- 與 OpenAI 同日降價，反映一件更根本的事：頭部實驗室已不再靠單點性能領先，而是靠「同等能力、更低單價」搶企業預算。企業市場的採購邏輯向來如此——穩定性與成本比跑分重要。Coming on the same day as OpenAI's cut, this points to something more fundamental: the leading labs are no longer winning on a single benchmark, but on equal capability at a lower unit price. Corporate procurement has always worked this way — stability and cost outrank leaderboard scores.
- 顧問觀點
- 我們認為企業應該停止追逐「最強模型」，改為建立模型替換的能力。當代際成本一年可以跌四成，把業務邏輯綁死在單一模型上，就是每年自動加價一次。We think enterprises should stop chasing the strongest model and start building the ability to swap models. When generation-on-generation cost can fall 40% in a year, hard-wiring business logic to one provider is a self-imposed annual price increase.
- 對日常工作的影響
- 技術團隊若把提示詞、工具定義與流程邏輯寫進應用代碼，每次換模型都要重做整合。反之，把模型調用集中在一層路由，換模型只是改一行設定。If prompts, tool definitions and process logic are written into application code, every model change means re-integration work. If model calls sit behind a single routing layer instead, a model swap is a one-line change.
- 如何改善
- 本週做一次盤點：公司現時有幾個地方直接寫死了模型名稱？把它們列出來，標明哪些可以收歸一層統一路由。這是未來兩年最省錢的一項技術整理。This week, count how many places in your systems hard-code a model name. List them, and mark which ones could sit behind one shared routing layer. It is the cheapest technical tidy-up available over the next two years.
03
Google 推出 Gemini 3.8 Live，語音對話加入即時思考Google ships Gemini 3.8 Live with thinking-while-speaking voice
★★★☆☆6–12 個月6–12 months前線員工Frontline staff企業決策者Business owners資訊科技IT
- 事件
- Google DeepMind 於 9 月 15 日發布 Gemini 3.8 Live 與 Gemini 3.8 Live Extended Thinking，兩者為原生語音對語音模型，可在說話同時推理並在背景調用工具，支援 97 種語言中途切換，已於 Gemini API 與 AI Studio 開放。Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September. Both are native speech-to-speech models that reason while speaking and call tools in the background, switching between 97 languages mid-conversation, and are available in the Gemini API and AI Studio.
- 背後
- 過去兩年語音 AI 的主要問題不是聽不準，而是「接不上話」——延遲太高、不能邊聽邊想。Extended Thinking 加入即時語音，等於把「現場對話」與「深度推理」兩件過去互斥的事接在一起。客服、現場銷售、巡檢記錄這類工種，是最先被改寫的一批。For two years the problem with voice AI was not accuracy but turn-taking: latency was too high and the model could not think while listening. Adding extended reasoning to live audio joins two things that used to be mutually exclusive. Customer service, field sales and inspection logging are the first job families to be rewritten.
- 顧問觀點
- 我們認為語音是這一輪最被低估的入口。文字對話需要員工主動打字，語音可以嵌進正在做事的雙手裡。對前線工序而言，語音的滲透速度會快過桌面軟件。We think voice is the most underrated interface of this cycle. Text chat requires a worker to stop and type; voice fits inside a pair of hands that are already busy. On the frontline, voice will spread faster than any desktop tool.
- 對日常工作的影響
- 零售、地產、保險、工程巡檢等以「邊做邊講」為主要工作形態的行業，記錄與交接可以即時完成，不必回到辦公室補打字。管理層亦會第一次拿到前線的真實對話數據。In retail, property, insurance and engineering inspection — industries where talking while working is the normal mode — records and handovers can be produced on the spot instead of being typed up back at the office. Management also gets something it has never had: real frontline conversation data.
- 如何改善
- 本週找一位前線同事，用一部手機做一次實測：讓他邊做邊講十分鐘，事後由 AI 生成一份交接記錄，再叫他評價準確度。這一次實測的結論，比任何產品介紹都有用。This week, take one frontline colleague and one phone. Have them talk through ten minutes of real work, then let AI produce a handover note and ask them to score its accuracy. That single test tells you more than any vendor presentation.
04
五家實驗室十天內連發模型，Agent 開發者被迫付「評估稅」Five labs ship in ten days, and agent builders pay the evaluation tax
★★★★★已在發生Happening now企業決策者Business owners資訊科技IT前線員工Frontline staff
- 事件
- 9 月 1 日至 10 日，Anthropic、Google、Meta、OpenAI、DeepSeek 先後推出前沿模型，輸入價格相差達 67 倍；有分析指評估單一 Agent 跨八項基準的中位成本約 800 美元，跑完一輪完整排行榜約 4 萬美元。Gartner 預測 40% 的 AI Agent 項目會在 2027 年前被取消，主因是成本超支而非技術失敗。Between 1 and 10 September, Anthropic, Google, Meta, OpenAI and DeepSeek all shipped frontier models, with input pricing spanning a 67-fold range. One analysis puts the median cost of evaluating a single agent across eight benchmarks at about US$800, and a full leaderboard run at roughly US$40,000. Gartner forecasts that 40% of AI agent projects will be cancelled by 2027, driven by cost overruns rather than technical failure.
- 背後
- 這條新聞表面是產品發佈，實際是一張帳單。每推一個新模型，企業就要重新做基準測試、改整合層、做安全覆核、重新談採購條件——這些成本實驗室從不承擔。更關鍵的是，Agent 每次任務所耗的 token 是普通對話的五到三十倍，而根據麥肯錫 2026 年 7 月的調查，六成 Agent 開支花在反覆修正答案的循環上，93% 企業已超出 AI 預算，約五分之一機構已因營運成本而限制 AI 使用。This reads as a product cycle but functions as an invoice. Every new model forces fresh benchmarking, rework of the integration layer, a safety review and a renegotiation of procurement terms — costs the labs never carry. The sharper point is usage: an agentic task consumes five to thirty times the tokens of an ordinary chat prompt. McKinsey's July 2026 survey found 60% of agentic spend already goes to refining answers in check-revise-regenerate loops, 93% of enterprises exceeding their AI budgets, and roughly one in five already constraining AI use because of running costs.
- 顧問觀點
- 我們認為企業目前最缺的不是模型，是「算得清」。多數 AI 項目失敗不是因為做不到，而是因為沒有人在立項時寫下單位成本與退出條件。當你每隔兩週就要重新評估一次模型，你需要的不是更快的評估，而是更少的評估對象。What enterprises lack right now is not models but arithmetic. Most AI projects fail not because the work was impossible but because nobody wrote down the unit cost and the exit condition when the project was approved. If you must re-evaluate models every fortnight, what you need is not faster evaluation — it is fewer candidates.
- 對日常工作的影響
- 技術與財務部門會第一次因為 AI 而吵架：技術想追新模型，財務看到的是無上限的試錯支出。沒有成本紀律的公司，會在 2027 年前看到一批示範做得很漂亮、但沒有上線的項目。Expect the first genuine argument between engineering and finance over AI: engineering wants the newest model, finance sees open-ended trial spend. Companies without cost discipline will end 2027 with an impressive portfolio of demonstrations and almost nothing in production.
- 如何改善
- 本週為每一個進行中的 AI 項目補上三行字：每月單位成本上限、由誰負責看住、以及什麼情況下要停。三行字，足夠擋住大部分超支。This week, add three lines to every AI project currently running: the monthly unit-cost ceiling, the person accountable for watching it, and the condition under which the project stops. Three lines will prevent most overruns.
05
三大實驗室同日推出網絡安全專用模型與企業治理方案Three labs release cyber-security models and enterprise governance on the same day
★★★★☆已在發生Happening now資訊科技IT法務合規Legal & compliance企業決策者Business owners
- 事件
- 9 月 20 日，Google 透過 Fairwind 計劃向超過 650 個安全夥伴開放 Gemini 3.8 Flash Cyber；Anthropic 為 Claude Fable 5.1 解鎖網絡安全用途，並推出 Enterprise Frontier Safeguards 治理產品；OpenAI 則將 Astra 定位為達到其準備框架的 Critical 門檻。On 20 September, Google opened Gemini 3.8 Flash Cyber to more than 650 security partners through its Fairwind programme; Anthropic unlocked Claude Fable 5.1 for cyber-security work and launched Enterprise Frontier Safeguards; and OpenAI positioned its Astra model as meeting the Critical threshold in its Preparedness Framework.
- 背後
- 三年來企業在做的都是「保護 AI」——防止模型被攻擊、防止資料外洩。這一週三家同步轉向，變成「用 AI 保護其他一切」。協調發佈並非巧合，反映工具已成熟到可以在企業規模上作攻擊性使用。For three years the enterprise agenda was protecting AI: stopping models being attacked, stopping data leaking. This week all three labs pivoted to using AI to protect everything else. The coordination was not coincidence; it signals that the tooling is mature enough for offensive work at enterprise scale.
- 顧問觀點
- 我們認為真正的瓶頸不在模型，而在政策。多數企業拿到授權之後會卡在同一個地方：內部沒有那份寫明「可以用 AI 對自家系統做什麼」的政策，於是能力買了、不敢用。這是未來十二個月最普遍的停滯點。We think the bottleneck is not the model but the policy. Most enterprises will stall at the same point after acquiring the licence: there is no internal document stating what AI may and may not do against their own systems, so the capability is bought and left unused. That is where the next twelve months get stuck.
- 對日常工作的影響
- 資訊安全主管會被要求交出 AI 安全使用政策，而這份文件以往不在任何人的職責範圍內。同時，供應商會開始把 AI 安全能力寫進報價單——這筆預算將在明年出現在你的帳上。Security leads will be asked to produce an AI security-use policy, a document that until now belonged to nobody's job description. Vendors will also start quoting AI security capability as a line item, and that cost will land on your budget next year.
- 如何改善
- 本週先寫一頁「AI 可以做／不可以做」清單，涵蓋三個最敏感的動作：對自家系統掃描、對外寄送內容、自動修改配置。不需要完整，只需要有第一版。This week, draft a one-page list of what AI may and may not do, covering the three most sensitive actions: scanning your own systems, sending content externally, and changing configurations automatically. It does not need to be complete — only to exist.
06
歐盟 AI Act 首輪執法開跑，招聘與篩選工具成頭號對象EU AI Act enforcement begins, and hiring tools are first in line
★★★★★已在發生Happening now人力資源HR法務合規Legal & compliance企業決策者Business owners
- 事件
- 歐盟 AI Office 與法國 CNIL、德國 BfDI、西班牙 AESIA 等監管機構自 9 月起展開首輪合規檢查，9 月 10 日確認自動履歷篩選與人力資源決策工具在檢查範圍內；重點打擊的三類用途為履歷篩選、零售銀行信貸評分、私人醫療分流。第 50 條透明度義務已於 8 月 2 日可執行，違規罰款最高 1,500 萬歐元或全球營業額 3%。The EU AI Office, working with France's CNIL, Germany's BfDI and Spain's AESIA, has begun its first wave of compliance inspections. On 10 September it confirmed automated résumé screening and HR decision tools are in scope, with initial requests focused on résumé screening, retail credit scoring and private healthcare triage. Article 50 transparency duties became enforceable on 2 August, with fines up to €15 million or 3% of global turnover.
- 背後
- 這裡有一個多數人讀錯的地方：高風險系統的完整合規責任延後到 2027 年 12 月，但透明度義務沒有延後。所以企業現在的風險不是「未做齊文件」，而是「用了 AI 做決定卻沒有講」。監管機構用的第一件工具是資料索取通知（RFI），不是罰單——但這一步已經需要交出技術文件。There is a widely misread detail here: full high-risk conformity duties were deferred to December 2027, but the transparency obligation was not. So the live risk is not missing paperwork — it is using AI to make a decision without saying so. Regulators' first instrument is a request for information, not a fine, but that step already requires handing over technical documentation.
- 顧問觀點
- 我們認為香港與內地企業會以兩種方式被捲入：一是直接服務歐盟客戶，二是替歐盟客戶做外包。前者的風險寫在合約裡，後者的風險寫在問卷裡。而問卷來得比法規快。Hong Kong and mainland Chinese firms will be drawn in two ways: serving EU clients directly, or subcontracting for them. For the first group the risk sits in the contract; for the second it sits in a questionnaire. And the questionnaire arrives faster than the regulation does.
- 對日常工作的影響
- 人力資源部會最先被問到：你們用什麼工具篩履歷？有沒有告知應徵者？這條問題以往由法務回答，現在由 HR 回答。同時，任何向歐盟交付 AI 功能的產品團隊，問卷與文件準備會變成銷售流程的一部分。HR will be the first to be asked: what screens your résumés, and do candidates know? That question used to go to legal; it now lands on HR. And for any product team shipping AI features to EU clients, questionnaire and documentation work becomes part of the sales cycle.
- 如何改善
- 本週做一件事：把公司現有會「自動篩掉人」或「自動評分」的流程列出來——包括 Excel 巨集與評分表，不限於 AI 工具。這份清單就是未來問卷的底稿。This week, list every process in the company that automatically filters people out or scores them — including Excel macros and scoring sheets, not just AI tools. That list is the first draft of your questionnaire answers.
07
Cloudflare 改寫爬蟲預設：訓練與 Agent 流量預設封鎖，搜尋保持開放Cloudflare rewrites crawler defaults: training and agent traffic blocked, search left open
★★★★☆已在發生Happening now市場推廣Marketing資訊科技IT企業決策者Business owners
- 事件
- Cloudflare 於 9 月 15 日把 AI 爬蟲控制由單一開關拆成三類：搜尋、AI 訓練、AI Agent。新接入的廣告變現網域，預設允許搜尋、封鎖訓練與 Agent 流量；現有免費方案且從未設定過的客戶同樣被套用新預設。On 15 September Cloudflare split its AI crawler control from a single switch into three categories: search, AI training and AI agents. For newly onboarded ad-monetised domains the default is to allow search while blocking training and agent traffic, and the same default applies to existing free-plan customers who had never set a preference.
- 背後
- 這條表面上與「網站設定」有關，實質是整個 AI 搜尋生態的分水嶺。更需要注意一個技術陷阱：Googlebot、Applebot、Bingbot 屬於混合用途爬蟲，按「最嚴格者優先」原則，若站方封鎖訓練類別，有機會連搜尋抓取一併被封。同期有分析指，8 月 8 日 ChatGPT Search 開始把查詢限定於特定網域並優先官方品牌站之後，部分網站的 AI 引用明顯下滑。This reads as a hosting setting; it is actually a watershed for the whole AI search ecosystem. There is also a technical trap: Googlebot, Applebot and Bingbot are mixed-use crawlers, and under the most-restrictive-rule principle, blocking the training category can take search crawling down with it. In parallel, analysis suggests that after ChatGPT Search began scoping queries to specific domains and favouring official brand sites on 8 August, some sites saw a marked drop in AI citations.
- 顧問觀點
- 我們認為未來兩年「被 AI 引用」會取代「被 Google 排名」，成為內容工作的主要 KPI。而這件事有多脆弱，這一週展示得很清楚：一個你從未打開過的預設值，就可以決定你的內容會不會出現在答案裡。We think that within two years being cited by AI will displace ranking on Google as the main content KPI. How fragile that position is became clear this week: a default setting you never opened can decide whether your content appears in an answer at all.
- 對日常工作的影響
- 任何依賴自然流量獲取客戶的公司——包括我們自己——都需要開始定期檢查兩件事：爬蟲政策是否誤封、以及 AI 答案引擎有沒有引用自己。這項工作以往不存在，現在需要有人負責。Any company that acquires customers through organic traffic — ours included — now needs a routine check on two things: whether crawler policy is blocking too much, and whether AI answer engines cite them. This job did not exist before; now someone has to own it.
- 如何改善
- 本週用五分鐘做兩件事：打開主機商的 AI 爬蟲設定，確認搜尋類別沒有被誤封；然後在一個 AI 搜尋工具裡搜自己的公司名，看回答引用了哪些來源。第二件事往後每月做一次。Spend five minutes this week on two checks: open your host's AI crawler settings and confirm the search category is not blocked by accident, then search your own company name in an AI tool and see which sources it cites. Repeat the second one monthly.
08
阿里發布「企業上下文」與 899 元桌面機械人，數字員工由雲端走向桌面Alibaba ships Enterprise Context and a RMB 899 desktop robot as digital staff move onto the desk
★★★★★已在發生Happening now企業決策者Business owners資訊科技IT前線員工Frontline staff人力資源HR
- 事件
- 9 月 22 日雲棲大會，千問辦公發布面向企業的「企業上下文」（Enterprise Context）功能，將企業資料壓縮與結構化後供 Agent 判斷，並同步推出安全中心；自研桌面機械人 Eva 首次亮相，定價 899 元，內置大模型，已打通釘釘。同期字節跳動把豆包、飛書、火山引擎整合為一體化企業智能平台，百度推出「百度搭子」。At its Apsara Conference on 22 September, Alibaba's Qwen Office launched Enterprise Context, which compresses and structures company data for agents to reason over, alongside a new security centre. Its in-house desktop robot Eva made its debut at RMB 899 with a built-in model and DingTalk integration. In the same period ByteDance merged Doubao, Feishu and Volcano Engine into a single enterprise AI platform, and Baidu launched its own office assistant.
- 背後
- 中國大廠的 AI 辦公競爭已由「功能疊加」轉向「生態綁定」。其中一個技術判斷值得留意：Agent 用得好不好，關鍵是上下文數據——AI 需要知道你做什麼工作、最近有哪些事未完成。這意味著真正的護城河不是模型，而是企業自己沉澱的流程資料。同期亦有一個現實數字：大量企業 AI 項目仍停留在概念驗證階段，瓶頸不在模型能力，而在缺少深入業務現場、對落地結果負責的執行層。Competition among China's platform giants has moved from stacking features to binding ecosystems. One technical judgement is worth noting: agent performance turns on context data — the AI needs to know what work you do and what is still outstanding. That makes the real moat the process data a company accumulates, not the model. A related reality check: many enterprise AI projects remain stuck at proof of concept, and the bottleneck is not model capability but the lack of an execution layer that goes into the business and owns the outcome.
- 顧問觀點
- 我們認為中國市場這一輪的做法比歐美更貼近企業現實——直接把 Agent 放進既有的協作工具（釘釘、飛書）裡，而不是要員工另開一個介面。對中小企而言，最實際的訊號是：不需要先建 AI 平台，先整理好自己那份「工作上下文」。We think this round in the China market is closer to enterprise reality than the Western approach: agents are placed inside the collaboration tools staff already use — DingTalk, Feishu — rather than behind another interface. For an SME the practical signal is that you do not need to build an AI platform first; you need to organise your own working context first.
- 對日常工作的影響
- 行政與營運主管會發現，AI 項目成敗愈來愈取決於內部資料的整理程度，而不是預算多少。同時 899 元級別的桌面硬件，會令「一人一個 AI 助手」由概念變成可以計預算的採購項目。Operations and admin leads will find that AI outcomes depend increasingly on how well internal data is organised, and less on budget size. Meanwhile, desktop hardware at the RMB 899 level turns one-AI-assistant-per-person from a concept into a line item you can actually budget for.
- 如何改善
- 本週做一件低成本的事：選一個崗位，把他「每次做決定時要看哪些資料」寫成一頁清單。這一頁就是未來餵給 Agent 的企業上下文雛形，也是 AI 落地最容易被跳過、但最關鍵的一步。Do one cheap thing this week: pick a single role and write down, on one page, every piece of information that person consults when making a decision. That page is the prototype of the enterprise context you will feed an agent later, and it is the step most often skipped.
09
行業調查：七成組織授予 Agent 超出人類員工的權限，「KYA」標準正式立項Survey: 70% of organisations grant agents more access than human staff, as KYA standard is filed
★★★★★已在發生Happening now企業決策者Business owners資訊科技IT法務合規Legal & compliance人力資源HR
- 事件
- 9 月 22 日深圳一場產業智能體大會披露多項調查數字：70% 組織授予智能體超出同等人類員工的權限；80% 組織已報告智能體執行超出預期範圍的行為；62% 企業將數據權限與安全合規列為跨系統執行的首要障礙。9 月 10 日，螞蟻數科與中國電子技術標準化研究院等推動的《智能體身份管理要求》國家標準化指導性技術文件正式立項。An industry agent conference in Shenzhen on 22 September disclosed a set of survey figures: 70% of organisations grant agents more access than an equivalent human employee; 80% have already seen an agent act beyond its intended scope; and 62% name data permissions and compliance as the primary obstacle to cross-system execution. On 10 September a national standardisation guidance document on agent identity management, advanced by Ant Group's technology arm with the China Electronics Standardization Institute, was formally filed.
- 背後
- 數字背後是一個很簡單的管理漏洞：企業為 Agent 開權限時，用的是「方便」而不是「職責」的邏輯。人類員工入職要簽權限表、要分級授權，Agent 卻往往一次拿到全部。KYA（Know Your Agent）被提出為繼 KYC、KYP 之後的新底層要求，反映監管與業界已把 Agent 當作需要身份、需要問責的「準員工」看待。Behind the numbers is a simple management gap: agents are granted access on the logic of convenience rather than responsibility. A human employee signs an access schedule and gets tiered permissions; an agent often receives everything at once. Know Your Agent is being advanced as the next foundational requirement after KYC and KYP, which tells you regulators and industry now treat agents as quasi-employees who need an identity and someone to answer for them.
- 顧問觀點
- 我們認為這一條是十條之中最應該貼在老闆檯頭的一條。當企業開始統計每項任務中「人與 AI 的分工比例」，管理就不再是感覺，而是可以量度的數字。給 Agent 開權限的那一刻，其實是企業在為它寫職位說明書——多數公司還沒有寫。Of the ten items today, this is the one worth pinning to the owner's desk. Once a company starts measuring the human-to-AI split on each task, management stops being a matter of impression and becomes a number. The moment you grant an agent access, you are effectively writing its job description — and most companies have not written one.
- 對日常工作的影響
- 資訊科技與合規部門要為每一隻 Agent 建立身份檔案：它叫什麼、屬於哪個部門、可以碰哪些資料、誰是它的負責人。這套機制以往只用於員工與外部系統，現在要延伸到非人類執行者。IT and compliance will need an identity file for every agent: what it is called, which department it belongs to, what data it may touch, and who is accountable for it. That mechanism used to apply only to people and external systems; it now extends to non-human actors.
- 如何改善
- 本週做一張表：列出公司現時所有在跑的 Agent（包括自動化流程與機械人帳號），每一行寫上「擁有者」與「可存取的資料範圍」。凡是填不出擁有者的，先停用。Build one table this week: list every agent currently running — automated flows and bot accounts included — and give each row an owner and an access scope. Anything you cannot assign an owner to should be switched off first.
10
Nvidia 支持的 Firmus 擬籌百億美元買晶片，並以馬來西亞數據中心接住 OpenAINvidia-backed Firmus seeks US$10bn for chips and anchors OpenAI in Malaysia
★★★☆☆3 年3 years企業決策者Business owners資訊科技IT
- 事件
- 據報 Nvidia 有份投資的 Firmus Technologies 正洽談約 100 億美元融資（約 75 億債務、25 億股權），用於購入 Nvidia 晶片並在印尼巴淡島興建 360 兆瓦 AI 工廠，同時籌備澳洲上市，預計集資約 50 億美元。公司已鎖定 OpenAI 為兩座馬來西亞數據中心的錨定租戶。另一邊，Nvidia 為 SB Energy 在俄亥俄州、租予 OpenAI 的數據中心園區提供最高 1,050 億美元擔保。Firmus Technologies, in which Nvidia holds a stake, is reported to be negotiating roughly US$10bn in financing — about US$7.5bn of debt and US$2.5bn of equity — to buy Nvidia chips and build a 360MW AI facility on Batam island in Indonesia, alongside an Australian listing expected to raise around US$5bn. The company has secured OpenAI as anchor tenant for two planned Malaysian data centres. Separately, Nvidia is providing up to US$105bn of support for SB Energy's Ohio data-centre campus leased to OpenAI.
- 背後
- 兩件事同時出現，揭示 AI 基建的資金結構正在變化：晶片商不只是供應商，而是透過入股、收入分成與信用支持，直接參與其客戶的融資。這種「循環融資」加快部署速度，但亦令風險在晶片商與基建商之間高度相連。另一位於亞美尼亞、由 Nvidia 持股的 300 兆瓦項目，同樣需要美國出口批文才可引入先進晶片——說明算力地理正被政策而非市場決定。Together these two stories show how AI infrastructure is being financed: the chip vendor is no longer just a supplier but a participant in its customers' funding, through equity stakes, revenue sharing and credit support. That circularity accelerates deployment while tightly coupling risk between chipmaker and developer. A separate 300MW project in Armenia, also part-owned by Nvidia, required US export approval to import advanced chips — a reminder that the geography of compute is set by policy, not markets.
- 顧問觀點
- 我們認為企業不需要預測泡沫，只需要分清兩件事：那些與 AI 算力掛鉤的長期承諾（例如多年雲端合約），與可按年檢視的工具支出。前者一旦簽下就難以退出，後者才是你現階段應該集中投入的地方。Enterprises do not need to call the bubble. They need to separate two things: long-term commitments tied to AI compute, such as multi-year cloud contracts, and tooling spend that can be reviewed annually. The first is hard to exit once signed; the second is where your investment belongs today.
- 對日常工作的影響
- 東南亞將在未來兩年內增加大量算力供給，對區內企業的直接意義是：合規區內的模型服務會變得更便宜、延遲更低。同時簽署多年期 AI 基建合約的風險，會在 2028 年後才真正顯現。South-East Asia will gain a large amount of compute capacity over the next two years. For regional firms the direct effect is cheaper, lower-latency model services inside the compliance perimeter. The risk of signing multi-year AI infrastructure contracts, by contrast, will not show up until after 2028.
- 如何改善
- 本週確認一件事：公司現有的雲端與 AI 服務合約，年期是多久、有沒有中途調整條款。若有一份三年以上、且未寫明可按使用量下調的合約，值得現在就與供應商談一次。Confirm one thing this week: what is the term of your existing cloud and AI service contracts, and is there a mid-term adjustment clause? If you hold a three-year-plus agreement with no volume-based reduction clause, it is worth a conversation with the vendor now.
