今日十條 · 2026 年 10 月 29 日默認值與准入名單同日改寫：谷歌把最強模型只交給防守方，也把客戶請求默認留存五十五日

我們認為「先給防守方」將成為前沿模型的標準發布程序。企業評估模型時，供應商名單上的可用日期已不代表可採購日期，須另問准入路徑與附帶審查條件。We expect a defenders-first gate to become the standard release procedure for frontier models. When evaluating vendors, an availability date no longer implies a purchase date; buyers must ask separately about access routes and the reviews attached to them.

我們認為這一類系統的商業價值不在娛樂，而在訓練環境：機械人與自駕要先在可重複的虛擬場地試錯。企業應把它當成試驗場成本，而非內容工具。We see the commercial value not in entertainment but in training environments: robots and vehicles need repeatable virtual grounds for trial and error. Enterprises should budget it as a testing cost, not a content tool.

我們認為記憶體已由零件變成需要預訂的產能。企業規劃伺服器與終端換機時，應把交期與配額視為主要風險，而非價格折扣。We see memory turning from a component into capacity that must be reserved. When planning server and device refreshes, lead times and allocations should be treated as the primary risk rather than the discount on offer.

我們認為算力採購已與股權融資交織，買方議價空間收窄。企業應留意供應商之間的資本關係，因為它會影響配額先給誰、價格維持多久。We see compute procurement becoming entangled with equity financing, narrowing the buyer’s room to negotiate. Enterprises should track capital links between suppliers, since they shape who gets allocated first and how long prices hold.

我們認為短期可複製的不是造藥，而是文件與批次審閱。這類工序有明確對錯、可量度省時，是受監管行業最現實的人工智能落地點。We judge the replicable near-term gain to be document and batch review rather than drug design. Such work has clear right answers and measurable time savings, making it the most realistic entry point for regulated industries.

我們認為事故分類一旦統一，採購合約就會引用它。企業宜預先按同一口徑記錄自家近失誤，否則日後被要求補交時無從追溯。Once incident categories are common, procurement contracts will cite them. Enterprises should log their own near misses in that vocabulary now, or they will have nothing to produce when asked later.

我們認為這是本輪最容易被低估的結構風險：省下的起薪，會在三年後以中層斷層的形式回到帳上。企業應把入門招聘當作能力投資而非成本項。We see this as the most underrated structural risk of the cycle: saved starting salaries return as a middle-management gap within three years. Entry hiring should be budgeted as capability investment, not a cost line.

我們認為默認值就是治理邊界。企業若只在控制台手動關閉，任何新專案都會自動回到默認，因此政策必須寫在程式碼與部署清單裡，而非寫在備忘錄。We hold that a default is a governance boundary. If a firm only switches logging off in a console, every new project reverts to the default, so the policy must live in code and deployment checklists rather than a memo.

我們認為主權人工智能已由口號變成採購條件。供應商國別與託管地會直接影響能否投標，港資與區內企業在跨境專案中應預先交代數據駐留安排。We see sovereign AI shifting from slogan to procurement condition. Supplier nationality and hosting location now decide eligibility, so firms bidding across borders should state data-residency arrangements up front.

我們認為小型模型把「可接受風險」往前推了一步：資料不必離開設備。但授權門檻與全部基準由廠商自報，企業應先自測再決定，不宜以公關數字替代評估。Small models move the acceptable-risk line by keeping data on the device. Yet the licence threshold applies and every benchmark is vendor-reported, so firms should test before deciding rather than substituting publicity numbers for evaluation.