今日十條 · 2026 年 9 月 29 日資本買能力、代理做實事：AI 由示範走向執行的一天

我們認為重點不在 82 億美元，而在 AMD 承認看不清下一個工作負載。當硬體廠要靠買模型團隊才能規劃產品，代表先做模型、再配算力的順序已經倒轉。未來算力會圍繞特定模型形狀定價，而非通用規格。We think the point is not the $8.2 billion but AMD admitting it cannot see the next workload clearly. When a hardware vendor has to buy a model team to plan its products, the sequence of build-model-then-buy-compute has reversed. Smaller firms need not follow, but should read it: compute will be priced around specific model shapes, not generic specs.

我們認為最值得留意的是 Meta 未公布定價、服務條款與資料處理安排，連 Llama 在新架構中的位置也未說明。企業採用代理，真正簽的是資料條款而非模型能力；條款出來前，試點只應放非敏感工序。What stands out is what Meta did not publish: pricing, service terms, data handling, and where Llama sits in the new stack. Enterprises adopting agents are really signing data terms, not model capability. Until those terms exist, pilots should stay on non-sensitive work.

我們認為要看清授權：功能較強的 27B 採 CC BY-NC 4.0 不得商用，可商用的 35B-A3B 是 Apache 2.0，但公司自報長流程基準分數只有一半左右。開放權重不等於可免費落地，法務須先讀模型卡。Read the licences carefully. The stronger 27B carries CC BY-NC 4.0 and cannot be used commercially; the commercially usable 35B-A3B is Apache 2.0 but, by the vendor's own numbers, scores roughly half as much on long-horizon benchmarks. Open weights are not free deployment; legal must read the model card first.

我們認為這份更新的實質意義，是把「你用哪家的雲與模型」由技術選型升格為治理議題。當監管機構點名非歐盟依賴，企業要準備的不再是模型說明書，而是斷供情境下的替代方案與退出測試紀錄。The practical effect is that which cloud and which model you use has been upgraded from a technical choice to a governance matter. When supervisors name non-EU dependence, what you must produce is no longer a model datasheet but an exit test and a fallback plan for supply being cut.

我們認為最值得抄的是先拆職務、再配課程的次序。多數企業倒過來做，先買課程再想誰要學。新加坡把財富管理與營運後台列為首批路徑，正因為這兩類職位的工作內容正被改寫，而非被取代。The transferable lesson is the order: break down the job first, then choose the course. Most firms do the reverse. Singapore put wealth management and back-office operations first precisely because those roles are being rewritten rather than removed.

我們認為這條消息對香港與內地中小企業的意義是間接但實在的：資金開始追逐「能供電的場地」，意味數據中心租金與電價的定價權會進一步向能源方傾斜。企業續租機房或自建機房前，應把電力條款看得比樓面面積更重。For SMEs in Hong Kong and the mainland the relevance is indirect but real: capital is chasing sites that can supply power, which shifts pricing power further toward the energy side. Before renewing or building a server room, weigh power terms more heavily than floor area.

我們認為這類「需求拆解型」模型值得中小企業留意，因為它解決的正是企業最不擅長的一步：把模糊的技術痛點寫成可採購的規格。但數據基座由政府背景平台持有，企業提問等於披露策略方向，敏感議題不宜放進去。This class of requirement-decomposition model deserves SME attention because it addresses the step firms handle worst: turning a vague technical pain into a purchasable specification. But the data layer sits with a government-backed platform, so queries disclose strategic direction; sensitive questions do not belong in it.

我們認為這條消息要讀兩層。第一層是它仍屬媒體引述消息人士的報道，官方未確認，任何採購決策都不應以此為據；第二層是它顯示推理算力已成真實瓶頸。在內地部署代理，應同時準備國產與進口兩套推理方案。Read this at two levels. First, it remains a report citing unnamed sources with no official confirmation, so no procurement decision should rest on it. Second, it shows inference compute has become a genuine bottleneck. Firms deploying agents on the mainland should keep both domestic and imported inference options ready.

我們認為兩份研究其實可以同時成立：招聘放緩與失業率未升，代表企業是在少請人而非在裁人。風險落在應屆生的入職機會，不在在職員工的飯碗。用裁員恐慌去驅動重組，是把錯的風險回應在錯的人身上。Both studies can hold at once: slower hiring with flat unemployment means firms are hiring less, not cutting more. The practical implication is that the risk sits with graduates' entry opportunities, not with incumbents' jobs. Restructuring out of layoff panic answers the wrong risk on the wrong people.

我們認為這類工具真正改變的是初級財務人員的工作內容：從逐筆核對變成審閱代理提出的根因與信心分數。企業若只買工具而不改職責與考核，結果會是代理在跑、人仍在做同樣的重複核對。What these tools really change is the work of junior finance staff: from line-by-line checking to reviewing the root cause and confidence score an agent proposes. Buy the tool without changing duties and appraisals, and you get agents running while people repeat the same checking, saving nothing.