登入 預約諮詢
AiX 環球 AI 情報 Global AI Intelligence
第 4 期 2026 年 9 月 28 日(星期一) 07:30 出刊 10 條情報 · 22 原文來源連結 RSS 過往期數

今日十條 · 2026 年 9 月 28 日價格向下、責任向上:AI 成本曲線與治理紅線同日轉向

今日主編判斷

今日十條的主線是「價格向下、責任向上」。OpenAI 把 GPT-6 兩款模型價格腰斬,Anthropic 以快取降價六成回應;同時 Oracle 就 2.45 吉瓦園區發出不可抗力通知,Akamai 以股權換取 116 億美元算力。治理端,英國議會傳召四家前沿實驗室,新加坡提議 AI 保障框架公約,上海開放生成式模型可控面客。我們認為,企業要算的是每項任務的總成本與責任歸屬。Today's through-line is prices down, accountability up. OpenAI halved the price of two GPT-6 models and Anthropic answered with a 60% cut on cache reads; in the same week Oracle filed a force majeure notice on a 2.45-gigawatt campus, and Akamai traded equity for an $11.6 billion CPU compute commitment. On governance, UK MPs called four frontier labs to a hearing, Singapore proposed a UN framework convention on AI safeguards, and Shanghai opened a controlled pilot for generative models facing customers. We believe the number that matters this year is not benchmark scores but the total cost per task and who carries the liability when it fails.

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 推出 GPT-6 Sol 與 Luna,API 價格較上一代促銷價再減半OpenAI releases GPT-6 Sol and Luna, halving API prices against its own promotional rates

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
OpenAI 於 9 月 22 日推出 GPT-6 Sol 與 Luna,輸入價每百萬 Token 為 2 美元與 0.10 美元,輸出價 10 美元與 0.50 美元,較 GPT-5.6 促銷價低一半。OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September. Prices are $2 and $0.10 per million input tokens and $10 and $0.50 per million output tokens, 50% below GPT-5.6 promotional rates.
背後
競爭焦點已由模型智力上限轉向代理工作流的每任務成本。OpenAI 稱降幅來自快取與推理效率改善,屬新的標準定價,而非限時促銷。Competition has shifted from the intelligence ceiling to the cost per task inside agentic workflows. OpenAI attributes the cut to caching and inference efficiency and presents it as new standard pricing, not a promotion.
顧問觀點

我們認為價格腰斬的真正意義,是把試錯成本從專案預算裡拿掉。以往團隊因重試太貴而放棄代理項目,如今成本紅線明顯下移;但計費仍有斷點,長上下文一超標,單位成本會跳升,省下來的錢可能在這裡還回去。We believe the real significance is that it removes trial-and-error cost from the project budget. Teams used to abandon agent projects because retries were too expensive, and that line has moved down. But billing still has cliffs: exceed the long-context threshold and unit costs jump, giving back the savings.

對日常工作的影響
資訊科技主管應重估長時段代理任務的每任務成本;財務與採購須把 Token 用量納入部門預算;市場部的高頻文案與資料整理工序最先受惠。IT leads should re-estimate cost per task for long-running agent jobs; finance and procurement must put token consumption into departmental budgets. Marketing workflows for high-volume copy and data cleanup benefit first.
如何改善
本週選一條重複工序,例如資料抽取或分類,用低價模型跑一日,記錄每任務實際成本與出錯率,再決定是否擴至其他工序。This week, pick one repetitive workflow such as data extraction or classification, run it for a day on the cheap tier, and log real cost per task and error rate before extending it.

Anthropic 發佈 Claude Opus 5.5,典型工作負載成本較 Opus 5 降四成Anthropic ships Claude Opus 5.5 with typical workload costs 40% below Opus 5

影響力✓ 已核實Verified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
Anthropic 於 9 月 22 日發佈 Claude Opus 5.5。每百萬 Token 輸入 4 美元、輸出 20 美元,較 Opus 5 低兩成;快取讀取降至 0.20 美元,降幅六成。Anthropic released Claude Opus 5.5 on 22 September at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, with cache reads cut to $0.20, a 60% reduction.
背後
官方稱輸出速度較 Opus 5 快逾三成,典型工作負載成本低四成,並首次為 Opus 系列加入網絡安全、生物研究與蒸餾三類防護;定價戰已由單價轉向長時段任務的總成本。Anthropic says output is more than 30% faster and typical workload cost is 40% lower, and it adds cybersecurity, biology and distillation safeguards to the Opus line for the first time. The pricing fight has moved from unit rates to the total cost of a long session.
顧問觀點

我們認為最值得留意的是快取讀取降六成。代理流程真正的大額帳單來自反覆讀取同一份上下文,而非輸出。若企業沒有把系統提示與工具定義固定下來,這六成降幅根本拿不到手。We think the number to watch is the 60% cut on cache reads. The real bill in an agent pipeline comes from re-reading the same context, not from output. If a company keeps rewriting its system prompt and tool definitions, that 60% never reaches its invoice.

對日常工作的影響
資訊科技主管須檢查代理流程的快取命中率與系統提示是否頻繁改動;法務可留意其水印設計已對應歐盟 AI 法案的透明度義務。IT leads should check cache hit rates and how often the system prompt changes. Legal teams may note that its watermarking is designed against the EU AI Act transparency duties.
如何改善
本週讓工程同事量一次現行代理任務的快取命中率。若低於一半,先把系統提示與工具定義固定,再談換模型。This week, have engineers measure the cache hit rate on one live agent task. If it is below half, freeze the system prompt and tool definitions before switching models.

Anthropic 與 Akamai 簽七年 116 億美元合約,以認股權換取 CPU 算力Anthropic signs a seven-year $11.6bn Akamai deal, trading equity for CPU compute

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
Akamai 於 9 月 24 日宣布與 Anthropic 簽訂七年、116 億美元的雲端服務合約,並向對方發出最多約佔已發行普通股 5% 的認股權,行使價每股 111.33 美元。Akamai announced a seven-year, $11.6 billion cloud services agreement with Anthropic on 24 September, issuing a warrant for up to about 5% of its common stock outstanding at an exercise price of $111.33 a share.
背後
生成式 AI 的算力瓶頸正由訓練用的 GPU 轉向代理執行所需的 CPU:資料檢索、代碼執行與網頁操作都靠 CPU。合約可再擴大 90 億美元,每增 30 億美元採購再歸屬約 1% 股權。The AI compute bottleneck is shifting from training GPUs to the CPUs that agents need for data retrieval, code execution and web navigation. The commitment can expand by a further $9 billion, with each extra $3 billion vesting about another 1% of equity.
顧問觀點

我們認為這類採購換股權的結構值得中小企業警惕。它把買方與供應商的財務風險綁在一起:當其中一方放慢投資,另一方的產能與估值同時受壓。便利是即時的,風險是延後的。We think small and mid-sized firms should treat equity-for-purchase structures with caution. They tie the financial risk of buyer and supplier together: when one side slows its spending, the other loses capacity and valuation at the same time. The convenience is immediate; the risk is deferred.

對日常工作的影響
企業決策者與資訊科技主管審視多年期雲端或算力合約時,須要求列明用量未達標的退出條款;財務部門要留意這類合約把資本開支前置。Owners and IT leads reviewing multi-year cloud or compute contracts should insist on exit terms when volumes fall short. Finance should note that such deals front-load capital expenditure.
如何改善
本週翻出公司最長的一份雲端或算力合約,標出用量未達標與調價兩項條款,記錄談判時可用的槓桿。This week, pull your longest cloud or compute contract, highlight the shortfall and repricing clauses, and note the leverage you would have in a renegotiation.

Oracle 就新墨西哥 2.45 吉瓦園區發出不可抗力通知,爭取延後付款Oracle files a force majeure notice on its 2.45GW New Mexico campus to defer payments

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
Oracle 向新墨西哥 Project Jupiter 園區的開發商 Blue Owl 發出不可抗力通知,稱電力取得存在延誤。園區設計用電 2.45 吉瓦,獲約 20 家銀行 180 億美元貸款,目標 2028 年啟用。Oracle issued a force majeure notice to developer Blue Owl over its Project Jupiter campus in New Mexico, citing potential delays in securing power. The site is designed for 2.45 gigawatts, carries an $18 billion bank loan, and targets a 2028 start.
背後
數據中心已由算力問題變成電力、水與許可問題。工地遇上天然氣管道延誤與許可訴訟,加上社區反對升溫,工期風險直接反映在融資條件上。Oracle 稱園區仍在計劃時間表內。Data centres have become a power, water and permitting problem rather than a compute problem. The site has faced a delayed gas pipeline and permit litigation, alongside rising local opposition, so schedule risk now shows up directly in financing terms. Oracle says the campus remains on its planned schedule.
顧問觀點

我們認為這是今年最重要的產業訊號之一:AI 供給的瓶頸已由晶片移到電網。任何依賴外部算力承諾的企業,都應把交付延誤寫進風險登記冊,而不是假設算力會按時到手。We consider this one of the year's most important industry signals: the AI supply bottleneck has moved from chips to the grid. Any company relying on an external compute commitment should log delivery delay as a risk rather than assume capacity arrives on time.

對日常工作的影響
企業決策者與財務主管規劃未來兩年 AI 產能時,須為算力交付加上緩衝期;法務應覆核租約與雲合約中的不可抗力及服務起始日條款。Owners and finance leads planning AI capacity for the next two years should build buffer into delivery dates. Legal should review force majeure and service-commencement clauses in leases and cloud contracts.
如何改善
本週在風險登記冊新增一項外部算力交付延誤,指定負責人,並寫下觸發後的三步應變。This week, add one line to your risk register for delayed external compute delivery, name an owner, and write the three steps you would take if it triggers.

英國議會傳召四家前沿實驗室,10 月 13 日就強制事前測試作證UK MPs call four frontier labs to give evidence on mandatory pre-release testing

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資訊科技IT
事件
英國下議院商業、創新、科學及貿易委員會於 9 月 22 日去信 OpenAI、Anthropic、Google DeepMind 與 Meta,邀請代表出席 10 月 13 日的聽證,並要求 9 月 29 日前回覆。On 22 September the UK Commons Business, Innovation, Science and Trade Committee wrote to OpenAI, Anthropic, Google DeepMind and Meta, inviting representatives to a hearing on 13 October and asking for replies by 29 September.
背後
英國至今沒有專門的 AI 法,事前測試依賴企業自願提交模型。委員會的問題集中在誰決定一個模型強到需要獨立測試,以及監管機構能否阻止部署或撤回模型。The UK still has no dedicated AI act, and pre-release testing relies on companies volunteering their models. The committee's questions centre on who decides a model is powerful enough to need independent testing, and whether a regulator could block deployment or withdraw a model.
顧問觀點

我們認為值得注意的不是聽證本身,而是自願二字已難維持。當最大市場之一開始討論強制事前測試與事故通報,香港及區內企業採購前沿模型時,很可能要對方出示第三方評估紀錄。We think the point is not the hearing but that voluntarism is getting hard to sustain. As one of the largest markets starts debating mandatory pre-release testing and incident reporting, buyers of frontier models in Hong Kong and the region will increasingly have to ask for third-party evaluation records.

對日常工作的影響
法務與合規主管應開始要求供應商提供模型評估與事故通報紀錄;資訊科技採購可把能否提供獨立評估證明列入評分項。Legal and compliance leads should start asking vendors for model evaluations and incident records. IT procurement can add independent evaluation evidence as a scored criterion.
如何改善
本週向主要模型供應商索取一份可公開引用的安全評估摘要,放入採購檔案備用。This week, ask your main model vendor for a publicly citable safety evaluation summary and file it with your procurement records.

新加坡在聯合國提議訂立《人工智能保障框架公約》Singapore proposes a UN framework convention on AI safeguards

影響力✓ 已核實Verified3 年3 years法務合規Legal & compliance企業決策者Business owners市場推廣Marketing
事件
新加坡外交部長維文於 9 月 26 日在聯合國大會提出《人工智能保障框架公約》構想,主張訂立共同規則與測試標準,並研究設立負責驗證的國際機構。Singapore's Foreign Minister Vivian Balakrishnan proposed a UN Framework Convention on AI Safeguards in his UN General Assembly address on 26 September, calling for common rules and testing standards and possibly a new verification institution.
背後
各國對 AI 風險的分類已有共識:自主系統失控、被用於製造大殺傷力武器、以及社會經濟動盪。分歧在互信不足,新加坡因此選擇由科學合作與跨境事故通報入手,避開主權爭議。There is already agreement on the categories of risk: loss of control over autonomous systems, misuse for weapons of mass destruction, and social and economic upheaval. The gap is trust, so Singapore starts from scientific cooperation and cross-border incident reporting rather than sovereignty disputes.
顧問觀點

我們認為公約短期不會落地,但它會改變採購語言。跨境服務合約將逐步出現測試方法可比較與嚴重事故跨境通報條款,香港作為區域樞紐會較早感受到這股壓力。We do not expect a convention soon, but it will change the language of procurement. Cross-border service contracts will gradually carry clauses on comparable testing methods and cross-border reporting of serious incidents, and Hong Kong, as a regional hub, will feel that pressure early.

對日常工作的影響
法務主管應留意服務合約新增的跨境通報義務;企業決策者要為可能的合規成本預留預算;市場部則須檢視 AI 生成內容的標示安排。Legal leads should watch for new cross-border notification duties in service contracts. Owners should budget for possible compliance cost. Marketing should review how AI-generated content is labelled.
如何改善
本週在一份對外服務合約樣本中加註一條,要求供應商在發生嚴重 AI 事故時於約定時限內通知本公司。This week, add one clause to a client-facing contract template requiring the supplier to notify you within an agreed window after a serious AI incident.

上海金融監管局推十六項措施,准金融機構在可控環境試行面客大模型Shanghai's financial regulator issues 16 measures, opening a controlled pilot for customer-facing models

影響力✓ 已核實Verified6–12 個月6–12 months法務合規Legal & compliance資訊科技IT企業決策者Business owners
事件
上海金融監管局印發《推動上海銀行業保險業人工智能應用的若干措施》,從三大方面提出十六項舉措,並提出探索生成式模型在金融領域應用的試點機制。The Shanghai Financial Regulatory Bureau issued measures to advance AI adoption across banking and insurance, setting out 16 steps under three headings and proposing a pilot mechanism for generative models in finance.
背後
內地金融 AI 的治理路線是先立規、後放行。措施一方面鼓勵智能營銷、智能核賠與智能風控落地,另一方面要求面向公眾的生成式模型備案或登記,並建立高風險應用場景的准入控制機制。The mainland approach to financial AI is rules first, release second. The measures encourage smart marketing, claims assessment and risk control, while requiring public-facing generative models to be filed or registered and high-risk scenarios to pass an admission gate.
顧問觀點

我們認為可控環境下直接面客是這次最實質的一步。過去金融機構因合規路徑不明而不敢把模型放到前台,現在有了試點通道。這對香港的金融科技供應商與顧問服務是明確的需求訊號。We see the controlled customer-facing pilot as the most substantive step. Institutions previously held models back from the front office because the compliance path was unclear; now there is a channel. That is a clear demand signal for Hong Kong fintech vendors and advisory services.

對日常工作的影響
金融機構的合規與風險主管須先建立高風險場景清單與准入機制;產品與市場部門可為面客場景準備試點方案;資訊科技要配合模型備案與可解釋性要求。Compliance and risk leads must first build a high-risk scenario list and an admission gate. Product and marketing can prepare pilot proposals for customer-facing use. IT must support filing and explainability requirements.
如何改善
本週由合規同事列出擬面客的 AI 場景清單,逐項標註是否屬高風險,缺漏的先補上人工覆核步驟。This week, have compliance list the AI scenarios you intend to put in front of customers, mark which are high risk, and add human review where it is missing.

阿里發佈真武 V900 芯片,並定下 2032 年逾 20GW 數據中心目標Alibaba unveils the Zhenwu V900 chip and a 20GW-plus data centre target for 2032

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
阿里平頭哥於 9 月 22 日雲棲大會發佈訓推一體芯片真武 V900,稱性能為上代 M890 的三倍,2027 年第一季量產;集團同時提出 2032 年阿里雲全球數據中心規模超過 20GW。Alibaba's chip unit T-Head unveiled the Zhenwu V900 training-and-inference chip at the Apsara Conference on 22 September, claiming three times the performance of the previous M890 with mass production in the first quarter of 2027, alongside a 2032 target of more than 20GW of global data centre capacity.
背後
國產算力走的是芯片、模型、雲協同的路線:以模型需求牽引芯片設計,再由雲平台對外輸出。新一代超節點可擴展至 50 萬卡,目標是降低單位算力的部署成本與能耗。Domestic compute is being built as a chip-model-cloud loop: model demand shapes chip design, and the cloud platform sells the result. The new supernode scales to 500,000 chips, aiming to cut deployment cost and energy per unit of compute.
顧問觀點

我們認為 20GW 這個數字比芯片參數更值得記住。它是把 AI 當成電力事業來規劃,意味未來數年國產算力的供應與價格會與政策節奏綁定,企業的選型與議價都要看政策窗口。We think 20GW is more memorable than any chip specification. It treats AI as a utility business, which means the supply and price of domestic compute will track policy timing over the next few years, and both selection and negotiation depend on that window.

對日常工作的影響
資訊科技主管評估國產算力時,須把遷移成本與工具鏈成熟度一併計入;企業決策者則要把供應鏈地域風險納入採購判斷。IT leads evaluating domestic compute must weigh migration cost and toolchain maturity, not just price. Owners should fold supply-chain geography into procurement decisions.
如何改善
本週選一個非關鍵工作負載,在國產芯片或國產模型上做一次小規模對照測試,記錄相容性問題。This week, take one non-critical workload, run a small side-by-side test on domestic silicon or a domestic model, and log the compatibility problems you hit.

馬來西亞為數據中心投資換尺:由引資規模轉向本地價值Malaysia changes the yardstick for data centre investment: from capital raised to local value

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners市場推廣Marketing資訊科技IT
事件
馬來西亞 2026 年已收到逾 40 宗數據中心建設或擴建申請。該國 2021 年至 2026 年上半年數據中心相關投資達 3,857 億令吉,當局稱日後以能源、用水與本地就業貢獻評估項目。Malaysia has received more than 40 applications to build or expand data centres in 2026. Data-centre-related investment reached RM385.7 billion between 2021 and the first half of 2026, and officials say projects will now be judged on energy, water and local job contribution.
背後
用電是現實約束:當局預測 2035 年全國數據中心電力需求將超過 5,000 MW。審批邏輯因此由引資轉為換能力,並以商業配對把工程、冷卻、電訊與保安工作留給本地企業。Power is the binding constraint: officials project national data centre electricity demand above 5,000 MW by 2035. Approval logic is shifting from attracting capital to buying capability, with business matching steering engineering, cooling, telecoms and security work to local firms.
顧問觀點

我們認為這是區內首個明確把數據中心要換來什麼寫成政策的市場。對香港與深圳的中小企而言,這是一條確實可參與的供應鏈:配對計劃的本地企業名額一年內由 17 家增至 51 家。We see this as the first market in the region to write into policy what a data centre must give back. For SMEs in Hong Kong and Shenzhen it is a supply chain they can genuinely enter: local firms in the matching programme rose from 17 to 51 in a year.

對日常工作的影響
市場與業務主管可留意馬來西亞數據中心供應鏈的採購機會;企業決策者若考慮在當地設點,須預算更嚴格的能源與用水審批時間。Marketing and business development leads should watch procurement openings in Malaysia's data centre supply chain. Owners considering a local presence must budget longer energy and water approval timelines.
如何改善
本週找出一項公司現有服務,評估能否切入數據中心的工程、冷卻、網絡保安或培訓環節,寫成一頁機會簡述。This week, take one service you already sell and assess whether it fits data centre engineering, cooling, cybersecurity or training, then write a one-page opportunity note.

史丹福研究:41 國數據顯示,採用 AI 企業的初級職位佔比下降Stanford study across 41 countries finds AI adopters cut the junior share of their workforce

影響力✓ 已核實Verified已在發生Happening now人力資源HR企業決策者Business owners前線員工Frontline staff
事件
史丹福數碼經濟實驗室 9 月 20 日發表研究,分析 41 國、12.5 億則招聘廣告與 1.54 億筆就業記錄,發現採用生成式 AI 的企業,初級職位佔比相對同類企業下降。Stanford Digital Economy Lab published a study on 20 September analysing 1.25 billion job postings and 154 million employment records across 41 countries, finding that generative-AI adopters reduced the junior share of their workforce relative to comparable firms.
背後
研究指出下降主要來自資深職位的擴張,而非初級職位被大規模裁減,整體就業甚至略有增長。企業把資深人手調往 AI 相關工作,同時放慢入門層的補充。The decline comes mainly from growth in senior employment rather than mass cuts at the junior end, with modest overall employment growth. Firms move experienced staff onto AI-related work while slowing intake at the entry level.
顧問觀點

我們認為這對人力資源的意義比 AI 取代人更迫切:問題不在裁員,而在入門階梯變窄。企業若同時放慢招聘與停止培訓,五年後將沒有人可以接手需要判斷的工作。For HR this matters more than the replacement narrative: the problem is not layoffs but a narrowing ladder. A firm that slows hiring and stops training at the same time will find no one able to take on judgement work five years out.

對日常工作的影響
人力資源主管須重新設計初級崗位內容,由執行轉為覆核與例外處理;企業決策者要為培訓編列明確預算;前線主管需學會把代理輸出當作下屬成果來審核。HR leads must redesign junior roles from execution towards review and exception handling. Owners should ring-fence a training budget. Frontline managers need to learn to review agent output the way they review a subordinate's work.
如何改善
本週把一份初級崗位職責表重寫一次:刪去純執行項目,加入覆核代理輸出與例外處理兩項職責。This week, rewrite one junior role description: remove purely executional tasks and add reviewing agent output and handling exceptions.

本週可做的三件事

  1. 為每一條在跑的自動化流程記錄三項成本:Token 費用、人工覆核時間、出錯後的返工成本。Record three costs for every automation already running: token spend, human review time, and the cost of rework after a failure.
  2. 挑一份外部算力或雲端合約,確認交付延誤時的補救條款與通知義務,記下不足之處。Take one external compute or cloud contract and confirm the remedies and notification duties when delivery slips, noting where it falls short.
  3. 把一份初級崗位職責表重寫一次:刪去純執行項目,加入覆核代理輸出與例外處理兩項職責。Rewrite one junior role description: remove purely executional tasks and add reviewing agent output and handling exceptions.

常見問題

今日(2026-09-28)「AiX 環球 AI 情報」有哪十條重點?

今日十條為:1、OpenAI 推出 GPT-6 Sol 與 Luna,API 價格較上一代促銷價再減半;2、Anthropic 發佈 Claude Opus 5.5,典型工作負載成本較 Opus 5 降四成;3、Anthropic 與 Akamai 簽七年 116 億美元合約,以認股權換取 CPU 算力;4、Oracle 就新墨西哥 2.45 吉瓦園區發出不可抗力通知,爭取延後付款;5、英國議會傳召四家前沿實驗室,10 月 13 日就強制事前測試作證;6、新加坡在聯合國提議訂立《人工智能保障框架公約》;7、上海金融監管局推十六項措施,准金融機構在可控環境試行面客大模型;8、阿里發佈真武 V900 芯片,並定下 2032 年逾 20GW 數據中心目標;9、馬來西亞為數據中心投資換尺:由引資規模轉向本地價值;10、史丹福研究:41 國數據顯示,採用 AI 企業的初級職位佔比下降。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?

今日十條的主線是「價格向下、責任向上」。OpenAI 把 GPT-6 兩款模型價格腰斬,Anthropic 以快取降價六成回應;同時 Oracle 就 2.45 吉瓦園區發出不可抗力通知,Akamai 以股權換取 116 億美元算力。治理端,英國議會傳召四家前沿實驗室,新加坡提議 AI 保障框架公約,上海開放生成式模型可控面客。我們認為,企業要算的是每項任務的總成本與責任歸屬。

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:OpenAI、OpenAI Developer Community、GIGAZINE、Anthropic、IT Brief Asia、AI Tech Daily、Akamai、DataCenterNews),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