Non DubitoA Mind in Many Languages
Search搜索搜尋
EN
日本語FrançaisDeutschEspañol한국어

When Switching Stops Being a Choice为什么越来越难换一家為什麼越來越難換一家 · 07

The Observer Has No Vote. What Else Do They Have?

不投票的人,也可能握着你的下一步

不投票的人,也可能握著你的下一步

Investment, compute, models and distribution connect through more than voting rights. Emerging concentration can be examined without declaring its future already settled.

投资、算力、模型与分发接在一起,控制不能只看席位。AI的集中仍在形成,能问的路也不必等到一切定型。

投資、算力、模型與分發接在一起,控制不能只看席位。AI的集中仍在形成,能問的路也不必等到一切定型。

Han Qin ·

An observer sits at the board table. There is no vote, but there are meetings to attend and papers to receive. Counting votes alone makes it easy to conclude that the founders retain every decision. Outside the meeting, however, where the model trains, how it reaches customers, and what changing partners would cost may already be governed by other contracts.

In the imagined Kingdom State, Stone Granary funds the newly founded Beacon, taking roughly a fifth of its equity and supplying half-price compute. Without that support the four researchers might not get started. Cooperation is not first rewritten as conspiracy. Money, equipment, distribution, and human ability can genuinely bring something into existence. The further question concerns how they arrange each party's next possibilities.

There are additional terms: Beacon uses Stone Granary's cloud for training and operation; a revenue share continues until an agreed return; before licensing to another cloud, Beacon must let Stone Granary match the terms within thirty days. A matching right does not by itself establish an unlimited veto over all cooperation. It does show a relationship a boardroom vote count cannot fully describe.

Separate the material. Which decisions can the partner actually influence? How much compute, favorable pricing, and revenue does Beacon depend on it for? What technical and commercial information does it receive that others cannot? Read these together without first renaming any one of them completed control. The useful question is how they operate, especially where the partners' interests might diverge.

Access to board papers need not entail directing a decision. If the papers contain plans and prices, they may affect the cloud partner's dealings with other clients; contents and uses require examination. Discounted compute provides support and creates a potential loss on switching. Knowing how much influence the combination gives requires more than recognizing familiar contractual labels or assuming a motive.

The account does not say Beacon tried to change clouds. Minority equity, revenue sharing, and matching rights cannot independently complete that missing story. Neither can absence of a vote remove the other materials. Ask which conditions Beacon could actually determine when seeking a different route. A broad control-or-no-control label too easily hides the concrete decision we need to follow.

A 2025 FTC staff report on cloud-provider and generative-AI partnerships examines related combinations: investment rights, revenue sharing, cloud-spending commitments, switching costs, and sensitive information. It identifies potential competition implications, not a violation by every partnership examined. Its value here is to identify documents and dependencies worth examining before either assurance or accusation substitutes for the particular terms and practical relations.

Read regulatory clearance headlines with the same attention. In the Microsoft–OpenAI inquiry, the UK's competition authority found in March 2025 no relevant increase from existing material influence to the de facto control being examined. That did not mean Microsoft had never influenced OpenAI, nor that the authority had completed every later competition assessment. The scope of the decision matters as much as its closure.

The 2024 Amazon–Anthropic decision stopped elsewhere: turnover and supply-share thresholds were not met, so the authority did not need to conclude whether material influence existed. In Google–Anthropic and Microsoft–Mistral, it found that the examined arrangements did not confer material influence. Similar-looking outcomes therefore rested on different reasons and left different questions unanswered. They cannot responsibly be compressed into one universal verdict about AI partnerships.

Each decision concerns its dated arrangement. It cannot certify every subsequent change, and the absence of substantive competition review is no ground for assuming a hidden violation either. Read what the institution actually decided and what it did not. A single headline verb should not be allowed to perform the entire inquiry, whether the reader welcomes the outcome or dislikes it.

AI is more than the chat window. Chips and compute sit below cloud, models, data, and distribution through assistants, marketplaces, and public-service systems. Control can occur at different layers or connect through partnerships. Capability in one does not automatically answer for admission rules in another. Following the layers reveals where a useful input becomes a condition somebody else must negotiate to continue their own work.

This is why price alone is insufficient. Buyers first encounter one or two selected products. Reviewers first encounter an assistant's summary. Public offices schedule through models. The resulting materials need not be false, but their selection and the routes for checking begin to share an entrance. The depth of influence changes when a service mediates what people first have available to judge.

Do not turn that into a claim that machines have replaced all judgment. Buyers can seek more information, reviewers can reread, and institutions can redesign processes. Those options require real time and capability, not merely verbal permission. Narrowing an entrance and eliminating human judgment are different conclusions. The former deserves scrutiny precisely because judgment needs materials and usable routes to exercise it.

Delegating a task does not delegate accountability to the model. Adopting marketplaces, offices, and companies must explain use, revision, and response to error. A model's widespread deployment does not automatically give its provider unified command over every decision. Each adopter's part needs tracing through the actual arrangement, rather than letting a common technical component absorb all the distinct institutional choices above it.

Has the whole AI field already become monopolized? The paper issues no such verdict. An OECD report in July 2026 records dynamism in foundation models while discussing concentration and longer-term entrenchment risks in hardware, data, and other layers. This mixed picture need not be forced into either everything is already safe or everything is already lost. Different layers can change in different ways.

Present competition cannot guarantee tomorrow; possible entrenchment cannot convict today. Risk research helps direct attention to relevant evidence without converting a possible future into an accomplished fact. Capabilities, prices, and access need examination at each stage. A developing sector can remain innovative while some important routes grow harder to bypass, making attentiveness more useful than a single permanent label applied to the whole.

Arrangements for effective challenge therefore need not wait for a proven violation. A position arranging other people's routes should offer explanation and constraints proportionate to its reach. Small positions can face lighter arrangements; deeper influence calls for stronger ones. Review is not merely punishment after wrongdoing. It is a condition that people need while legitimate activities are still being carried out.

There is a practical account too. While alternatives remain and systems are less entangled, preserving connections and building independent testing may be easier. A laboratory established three years earlier would give reviewers their own eyes three years earlier. That is a possibility to assess, not a guarantee that earlier action must prevent concentration or reveal every error before it causes harm.

Early arrangements can also be too heavy. Rules fixed too rigidly can block newcomers; too many vetoes can prevent anyone doing anything. Explain what is protected, what is lost, and whether a lighter option works, while permitting revision. Rejecting wait-until-entrenchment does not entail fixing every technical detail forever at the beginning. The constraint itself remains answerable for what it makes possible and prevents.

Now take a more distant thought experiment. Machines perform agricultural and water-service tasks; models enter food, water, and basic-payment arrangements. These capabilities are imagined premises, not technological forecasts. They let us place dependencies previously examined separately into one household's day, then ask what appears when a person encounters their combined effects rather than one contract at a time.

Its breakfast comes from fields worked by Fieldworks and is ordered through Crossways' Wayfinder. Water operation uses Beacon models and bills are paid through its assistant. Two older residents receive basic payments there. Their daughter works at a sorting station with model-generated shifts. Each arrangement can offer genuine convenience. Daily life has become easier in several respects; the example preserves that improvement.

One evening the daughter considers another assistant. Changing assistants would require the older residents to undergo roughly six weeks of counter-based reverification for their payments, with the counter two hours away; three years of memory cannot transfer; other assistants have lower-priority access for water bills. Changing still leaves only Wayfinder able to place Crossways orders. That final restriction is not a new switching cost. It is an existing dependence the proposed change does not remove.

She stays and submits a complaint that receives no reply by year-end. The story does not establish complete closure of every route. It shows burdens accumulating around one decision. Each service's usefulness cannot independently answer the combined problem; formal alternatives at each point cannot establish that this household can actually use them. The person needs a path across the arrangement, not a separate brochure for each component.

The responsible parties remain different. The state removes counters, Beacon writes interfaces, Crossways determines marketplace access, and water delegation and employment have other addresses. Sharing a model does not by itself prove unified control. We need both the combined difficulty and each position's ability to change its part. A common technical name must not become a place where all the decision makers disappear.

Food, water, and basic payments require protection, so public money should be examined through provision and continuation. Does it secure people's lives or become a guarantee that one provider cannot change? Even a backup built inside the original firm needs an account of who can use it and who can take over. The word backup does not itself establish independent continuity.

The day predicts no inevitable future. It makes a possible problem visible: dependence may grow not as a single locked gate but as many individually reasonable arrangements that together make another way of living hard to choose. While things are forming, the routes for questioning, revision, and continuation deserve to be built alongside them, rather than treated as work that can safely wait until everything else is complete.

董事会上有一位观察员。他没有表决票,可以听,也可以拿材料。如果只把桌上的票数数完,我们很容易说,决定还全在创办人手里。可散会以后,模型在哪训练、通过谁卖、换一家要付什么,可能早写在另几份合同里。

虚构的灯塔刚成立时,石仓出钱,占约两成股份,还给半价算力。没有这些,四位研究者也许起不来。我们不把合作写成阴谋。钱、设备、分发与人的能力相遇,确实可以造出原先没有的东西。问题在于,这份相遇后来怎样安排各自的下一步。

合同还有几项:训练与运行用石仓的云,收入分一部分给它,直到回报达到约定数;模型要给别家的云以前,先通知石仓,它可以在三十天内按相同条件先取。先取权不等于已经证明石仓能否决所有合作,却显然不是一张董事会表决票能够读完的关系。

把材料拆开,会看得清一些。哪些决定它能实际影响,是一问;灯塔靠它提供多少算力、优惠和营收,是另一问;它能看见哪些旁人看不到的技术与商业信息,又是一问。三种材料放在一起读,不必先把其中任何一种改名为已经控制。

例如,观察员拿到材料,未必因此指挥了决定。若里面有价格与计划,可能又影响它对别家客户的判断,内容和用途需要查。半价算力可以是支持,也会成为换云时必须面对的损失。要知道它们合起来有多大效力,得看利益相左时究竟怎么做。

灯塔是否试过换云,故事没有记。少数股、分成、优先权,都不能独自替这段未知写结局;不投票也不能把其他材料从桌上拿走。最有用的提问不是你控不控制,而是当它想走另一条路时,哪些条件真的由它自己决定?

现实研究也在看这种组合。FTC在2025年发布云服务商与生成式AI开发者合作的工作人员报告,讨论投资权利、收入分成、云支出承诺、转换成本及敏感信息。报告提出可能的竞争影响,不是给每对合作方作违法裁决;它把需要看的合同与实际依赖摆了出来。

读到监管新闻的放行二字,最好也往里走一步。英国CMA在2025年审微软与OpenAI时,判断既有实质影响没有上升成这次所问的事实控制,因此不构成相应并购情形。这不是一句微软从来没有影响,也不是已经把后续竞争影响全部审完。

亚马逊与Anthropic的2024年决定,停在另一处:营业额与供给份额的管辖门槛没有达到,CMA不需要就实质影响下结论。谷歌与Anthropic、微软与Mistral的决定,则认为所审合作没有赋予相应实质影响。结果看起来相近,理由和没有回答的问题并不相同。

这些都是相应日期、相应安排的决定,不能拿来替后来变化包办结论,也不能反过来从没有进入实体审查,猜出一定有问题。把机构实际判了什么、尚未判什么读清,才不会让一个新闻动词替我们完成整套判断。

再把视线拉远一点,AI并不只有聊天框。下面有芯片与算力,接着是云、模型、数据,再经过助手、市集或公共办事系统到人手里。控制可能发生在不同层,也可能借合作把几层接起来。一家公司在哪层有能力,不替其他层的准入规则自动作答。

这也是为什么价格之外,还要看取得信息的入口。买家先看见一两件已挑好的货,事务处先读到助手写的申报摘要,公共机构按模型排号。它们拿到的未必是错误材料,但材料怎样取舍、还能去哪里核,开始有共同的来处。

不要把这写成机器取代了所有判断。买家可以要求更多信息,审查者可以重读,公共机构可以重设程序;这些可能需要时间与能力,不能只留在口头。入口被收窄,与人已经不再能判断,不是一个结论。我们要保护的,是前一种变化不把后一种能力架空。

把判断委托出去,承担也没有一并委托给模型。采用它的市集、机构与公司,仍要说清怎样使用、谁能改、出错怎样接住。一个模型被许多地方调用,不自动使模型提供者统一取得全部决定权;各个采用者做的那一段,仍须沿实际安排查。

那么,现在是不是已经垄断了?论文并不给AI整个领域下这一判词。经合组织2026年的一份报告,一面记录基础模型近年的活力,一面讨论硬件、数据等环节的集中与长期固化风险。这种混合图景,不必被压成已经安全或已经没有机会两个极端。

今天有人竞争,不替明天保证;明天可能固化,也不替今天定罪。风险研究可以提早指明材料要往哪里找,不是把未来变成已经发生的事实。模型的能力、价格与谁能接近它,都还需要照每一阶段的实际情况看。

也因此,能问的安排不必等到认定越线以后才建立。一个位置有权安排别人的路,就应有与它作用相称的说明与制约,小的时候可以轻,影响深时更强。审查不是事后惩罚的另一种名字,而是人们共同生活里需要留着的条件。

还有一笔实践账:替代还在、系统还没完全缠住的时候,留接续、留接口、建独立检验能力,可能比后来容易。若检验室早建三年,审查者就早三年多一双自己的眼睛。这是值得比较的可能,不是保证提早做过就一定不会出事。

早设也可能太重。规定得太死,可能把新东西卡在门外;否决点太多,可能谁也做不了事。因此要说明想保什么、损失什么、有没有更轻的办法,并容许改变。反对等定型再问,不等于赞成一开始就把所有细节永远定死。

现在再向一个更远的设想走。机器做了农业、供水这些活,模型同时进入粮、水和基本给付的安排。这个前提属于想象,不是技术预测。它让我们把前面分开读的依赖,放到同一户人家的一天里。

早上的米菜由田畴的机器种,通衢的问路助手下单;水务运维用灯塔模型,水费也在助手里交。两位老人的基本给付从这里到账。女儿在分拣站工作,模型排班。每一件单独都可以说很方便,日常也真的变轻了。

晚上女儿想换助手,要重新盘算:老人给付去窗口核实约六周,最近窗口坐车两小时,三年记事带不走,别家的助手缴水费排在后面。换了,在通衢也仍只有问路能代买;最后这一项不是更换新添的损失,而是更换也没有消去的旧依赖。

她留下,并提交一条申诉,到年底没有回音。故事没有证明所有路都已经封死,却让人看见几份负担怎样叠到一个决定上。某项服务好用,不能单独回答这个合起来的问题;每处都有形式上的替代,也不能代证这一户真的用得起。

各处决定者还不同。窗口由州撤,接口由灯塔写,市集准入由通衢定,水务委托与女儿工作的安排也另有位置。共同用一个模型,不自行证明它们已经归一家指挥。要读合起来的困难,也要查各自能改动的部分,不能让所有人躲到同一个技术名字后面。

粮、水、给付需要保障,公共的钱便要继续问到供给与接续。是在替人保住生活,还是已变成某家机构不能改变的保证?备用系统即使建了,若仍只在原来那一家手里,也还得问它究竟备用给谁、出了事谁能接上。

这一天不是拿来证明未来必然如此。它把可能的问题提前变得可见:依赖并不总是一个被锁死的大门,也可能是许多原本合理的小安排,最后使一个人很难再换一种过法。既然还在形成,就更值得把那条能问、能改、能接续的路一起做出来。

董事會上有一位觀察員。他沒有表決票,可以聽,也可以拿材料。如果只把桌上的票數數完,我們很容易說,決定還全在創辦人手裡。可散會以後,模型在哪訓練、通過誰賣、換一家要付什麼,可能早寫在另幾份合同裡。

虛構的燈塔剛成立時,石倉出錢,佔約兩成股份,還給半價算力。沒有這些,四位研究者也許起不來。我們不把合作寫成陰謀。錢、設備、分發與人的能力相遇,確實可以造出原先沒有的東西。問題在於,這份相遇後來怎樣安排各自的下一步。

合同還有幾項:訓練與運行用石倉的雲,收入分一部分給它,直到回報達到約定數;模型要給別家的雲以前,先通知石倉,它可以在三十天內按相同條件先取。先取權不等於已經證明石倉能否決所有合作,卻顯然不是一張董事會表決票能夠讀完的關係。

把材料拆開,會看得清一些。哪些決定它能實際影響,是一問;燈塔靠它提供多少算力、優惠和營收,是另一問;它能看見哪些旁人看不到的技術與商業資訊,又是一問。三種材料放在一起讀,不必先把其中任何一種改名為已經控制。

例如,觀察員拿到材料,未必因此指揮了決定。若裡面有價格與計劃,可能又影響它對別家客戶的判斷,內容和用途需要查。半價算力可以是支持,也會成為換雲時必須面對的損失。要知道它們合起來有多大效力,得看利益相左時究竟怎麼做。

燈塔是否試過換雲,故事沒有記。少數股、分成、優先權,都不能獨自替這段未知寫結局;不投票也不能把其他材料從桌上拿走。最有用的提問不是你控不控制,而是當它想走另一條路時,哪些條件真的由它自己決定?

現實研究也在看這種組合。FTC在2025年發佈雲服務商與生成式AI開發者合作的工作人員報告,討論投資權利、收入分成、雲支出承諾、轉換成本及敏感資訊。報告提出可能的競爭影響,不是給每對合作方作違法裁決;它把需要看的合同與實際依賴擺了出來。

讀到監管新聞的放行二字,最好也往裡走一步。英國CMA在2025年審微軟與OpenAI時,判斷既有實質影響沒有上升成這次所問的事實控制,因此不構成相應並購情形。這不是一句微軟從來沒有影響,也不是已經把後續競爭影響全部審完。

亞馬遜與Anthropic的2024年決定,停在另一處:營業額與供給份額的管轄門檻沒有達到,CMA不需要就實質影響下結論。谷歌與Anthropic、微軟與Mistral的決定,則認為所審合作沒有賦予相應實質影響。結果看起來相近,理由和沒有回答的問題並不相同。

這些都是相應日期、相應安排的決定,不能拿來替後來變化包辦結論,也不能反過來從沒有進入實體審查,猜出一定有問題。把機構實際判了什麼、尚未判什麼讀清,才不會讓一個新聞動詞替我們完成整套判斷。

再把視線拉遠一點,AI並不只有聊天框。下面有芯片與算力,接著是雲、模型、資料,再經過助手、市集或公共辦事系統到人手裡。控制可能發生在不同層,也可能借合作把幾層接起來。一家公司在哪層有能力,不替其他層的準入規則自動作答。

這也是為什麼價格之外,還要看取得資訊的入口。買家先看見一兩件已挑好的貨,事務處先讀到助手寫的申報摘要,公共機構按模型排號。它們拿到的未必是錯誤材料,但材料怎樣取捨、還能去哪裡核,開始有共同的來處。

不要把這寫成機器取代了所有判斷。買家可以要求更多資訊,審查者可以重讀,公共機構可以重設程序;這些可能需要時間與能力,不能只留在口頭。入口被收窄,與人已經不再能判斷,不是一個結論。我們要保護的,是前一種變化不把後一種能力架空。

把判斷委託出去,承擔也沒有一並委託給模型。採用它的市集、機構與公司,仍要說清怎樣使用、誰能改、出錯怎樣接住。一個模型被許多地方調用,不自動使模型提供者統一取得全部決定權;各個採用者做的那一段,仍須沿實際安排查。

那麼,現在是不是已經壟斷了?論文並不給AI整個領域下這一判詞。經合組織2026年的一份報告,一面記錄基礎模型近年的活力,一面討論硬體、資料等環節的集中與長期固化風險。這種混合圖景,不必被壓成已經安全或已經沒有機會兩個極端。

今天有人競爭,不替明天保證;明天可能固化,也不替今天定罪。風險研究可以提早指明材料要往哪裡找,不是把未來變成已經發生的事實。模型的能力、價格與誰能接近它,都還需要照每一階段的實際情況看。

也因此,能問的安排不必等到認定越線以後才建立。一個位置有權安排別人的路,就應有與它作用相稱的說明與制約,小的時候可以輕,影響深時更強。審查不是事後懲罰的另一種名字,而是人們共同生活裡需要留著的條件。

還有一筆實踐帳:替代還在、系統還沒完全纏住的時候,留接續、留接口、建獨立檢驗能力,可能比後來容易。若檢驗室早建三年,審查者就早三年多一雙自己的眼睛。這是值得比較的可能,不是保證提早做過就一定不會出事。

早設也可能太重。規定得太死,可能把新東西卡在門外;否決點太多,可能誰也做不了事。因此要說明想保什麼、損失什麼、有沒有更輕的辦法,並容許改變。反對等定型再問,不等於贊成一開始就把所有細節永遠定死。

現在再向一個更遠的設想走。機器做了農業、供水這些活,模型同時進入糧、水和基本給付的安排。這個前提屬於想象,不是技術預測。它讓我們把前面分開讀的依賴,放到同一戶人家的一天裡。

早上的米菜由田疇的機器種,通衢的問路助手下單;水務運維用燈塔模型,水費也在助手裡交。兩位老人的基本給付從這裡到帳。女兒在分揀站工作,模型排班。每一件單獨都可以說很方便,日常也真的變輕了。

晚上女兒想換助手,要重新盤算:老人給付去窗口核實約六周,最近窗口坐車兩小時,三年記事帶不走,別家的助手繳水費排在後面。換了,在通衢也仍只有問路能代買;最後這一項不是更換新添的損失,而是更換也沒有消去的舊依賴。

她留下,並提交一條申訴,到年底沒有回音。故事沒有證明所有路都已經封死,卻讓人看見幾份負擔怎樣疊到一個決定上。某項服務好用,不能單獨回答這個合起來的問題;每處都有形式上的替代,也不能代證這一戶真的用得起。

各處決定者還不同。窗口由州撤,接口由燈塔寫,市集準入由通衢定,水務委託與女兒工作的安排也另有位置。共同用一個模型,不自行證明它們已經歸一家指揮。要讀合起來的困難,也要查各自能改動的部分,不能讓所有人躲到同一個技術名字後面。

糧、水、給付需要保障,公共的錢便要繼續問到供給與接續。是在替人保住生活,還是已變成某家機構不能改變的保證?備用系統即使建了,若仍只在原來那一家手裡,也還得問它究竟備用給誰、出了事誰能接上。

這一天不是拿來證明未來必然如此。它把可能的問題提前變得可見:依賴並不總是一個被鎖死的大門,也可能是許多原本合理的小安排,最後使一個人很難再換一種過法。既然還在形成,就更值得把那條能問、能改、能接續的路一起做出來。

Source & reading note原论文与阅读说明原論文與閱讀說明

Nine essays draw on the Monopoly prequel, seven papers and dialogue with earlier thinkers. Kingdom State, Crossways, Beacon and future water provision are imagined, not disguised allegations or forecasts. Linked cases distinguish investigations, allegations, jurisdiction, liability and remedies within their stated institutional and dated scope. Size, success, defaults and investment do not alone establish a violation; benefits do not answer for every exercise of control. Separation, interoperability and standby capacity are discussed as proposals with costs, not proven universal solutions. This is not legal or investment advice. English is independently written.

九篇据《垄断论》前置篇、七篇正文与传统对话篇改写。王国州、通衢、灯塔及未来供水等场景为设想,不是现实企业的化名或对未来的预测。真实案件与制度材料另列来源,区分调查、指控、管辖判断、责任判决与救济;按相应机构和材料日期说明,不以一项法律结论代答全部规范问题。规模、成功、默认设置或投资关系本身不直接证明越线;有益的一面也不替具体控制作答。关于拆分、互通与备用能力的建议不是已经验证的万能方案,不提供法律或投资建议。简繁中文同文,英文独立重写。

九篇據《壟斷論》前置篇、七篇正文與傳統對話篇改寫。王國州、通衢、燈塔及未來供水等場景為設想,不是現實企業的化名或對未來的預測。真實案件與制度材料另列來源,區分調查、指控、管轄判斷、責任判決與救濟;按相應機構和材料日期說明,不以一項法律結論代答全部規範問題。規模、成功、默認設置或投資關係本身不直接證明越線;有益的一面也不替具體控制作答。關於拆分、互通與備用能力的建議不是已經驗證的萬能方案,不提供法律或投資建議。簡繁中文同文,英文獨立重寫。

Read the source paper ↗阅读原论文 ↗閱讀原論文 ↗

FTC's original release on its AI partnerships and investments study, 2025FTC关于AI合作与投资研究的原始发布说明(2025年)FTC關於AI合作與投資研究的原始發佈說明(2025年) ↗

CMA Microsoft/OpenAI official decision summary, March 2025CMA微软/OpenAI决定官方摘要:控制层级变化(2025年3月)CMA微軟/OpenAI決定官方摘要:控制層級變化(2025年3月) ↗

CMA Amazon/Anthropic decision, September 2024CMA亚马逊/Anthropic决定原文(2024年9月)CMA亞馬遜/Anthropic決定原文(2024年9月) ↗

CMA Google/Anthropic decision, November 2024CMA谷歌/Anthropic决定原文(2024年11月)CMA谷歌/Anthropic決定原文(2024年11月) ↗

CMA Microsoft/Mistral decision, May 2024CMA微软/Mistral决定原文(2024年5月)CMA微軟/Mistral決定原文(2024年5月) ↗

OECD, Artificial Intelligence Markets: dynamism and structural risks, 2026OECD《人工智能市场》:活力与结构风险(2026年)OECD《人工智慧市場》:活力與結構風險(2026年) ↗