白宮拿走了AI的鑰匙，台積電的籌碼同時變大也變燙

發佈日期: 作者:
白宮拿走了AI的鑰匙，台積電的籌碼同時變大也變燙

2026年6月12日，Anthropic的Claude Mythos 5上線72小時後，美國商務部強制要求下架。理由是其網路入侵能力超過出口管制門檻。六月底，商務部再發函解封，但這次只向一百多個「可信夥伴」機構開放。OpenAI的GPT-5.6同日以「受限發布」方式上線。範式已定：高能力模型先停、再審、再按政治優先序放行。

審核權從企業移走了

這週的另一條線是白宮正式介入前沿AI模型的存取分配——直接決定誰能使用OpenAI、Anthropic、Google、Meta等公司的最先進模型。這不是新的出口管制條文，而是更上游的控制：在模型還沒到用戶手上之前，政府先過一遍。矽谷那套「決定誰能用自家模型」的話語權，被安靜地移走了。

同週，SpaceX正洽談成為美國國防部的雲端算力供應商。SpaceX已分別與Google及Anthropic簽署算力協議；若DoD合約成立，同一個算力節點將同時服務商業AI與軍事AI應用。Meta則評估向Anthropic出租算力，協議規模達一百億美元、為期兩年，Anthropic按月付款。三條線交叉不是巧合：AI存取國家化、AI算力軍事化、AI模型商業聯盟，在同一個月份同步成形。

算力的地圖長什麼樣

把這些協議畫出來，會看到一個清楚的拓樸：晶片製造→算力基礎設施→模型→存取審批。每一層都有人卡位。NVIDIA的H100在台積電的廠房裡生產，算力租給SpaceX和Meta，模型跑在Anthropic和OpenAI的資料中心，最後由白宮決定哪些機構能連線。

台積電坐在這條鏈的最底層。全球超過九成的先進製程晶片在這裡製造，NVIDIA對台積電的依賴程度接近百分之百。這個位置的特殊之處在於：白宮管控越深，對算力基礎設施的需求就越集中，對底層晶片的依賴就越剛性。台積電的地緣政治價值，隨著管控框架收緊而上升。

但這個上升不是免費的。布魯金斯學會的分析早就指出，美國財政部長已公開警告台灣晶片產能集中是全球經濟的「單點失效」風險。價值越高，就越可能被要求搬移、被要求複製、被當成談判籌碼。台積電在鏈條裡的位置是不可替代的，但「不可替代」在地緣政治裡有時候是風險，不是保護。

台灣在「可信夥伴」名單上嗎

Mythos 5解封的「可信夥伴」框架，是未來AI能力地緣政治分層的模板。台灣目前不是直接的優先開放對象，但已被納入可信AI供應鏈的討論範圍。這個位置有點尷尬：提供底層硬體，卻不一定能第一批用到跑在這些硬體上的模型。

更結構的問題是AI主權。NYT報導Anthropic發布Mythos模型時，指出缺乏算力基礎設施的國家將高度依賴OpenAI、Google、Anthropic等少數公司，在產品設計和安全政策上幾乎沒有影響力。台積電在供應鏈的製造端擁有無可取代的位置，卻幾乎沒有參與AI標準制定的份量。掌控全球九成先進製程，卻不在決定AI怎麼跑的那張桌子上。

白宮的這一步，把AI從產業競爭拉進了主權政治的範疇。管控框架一旦制度化，能否進入「可信夥伴」清單，就不只是技術問題，也不只是商業問題。

台積電2026年Q1營收三百五十七億美元,年增三十五個百分點。這個數字現在同時是地緣政治的籌碼,也是地緣政治的靶心。

— 邱柏宇

The White House Took the Keys to AI

On June 12, 2026, Claude Mythos 5 was pulled offline by the U.S. Commerce Department 72 hours after launch. The stated reason: its cyber-intrusion capabilities exceeded export control thresholds. Two weeks later, Commerce sent Anthropic a letter lifting the block — but only for over 100 designated “trusted partner” institutions. OpenAI’s GPT-5.6 went live the same day under a similar “restricted release” framework. The pattern is now set: high-capability models get stopped, reviewed, and re-released in order of political priority.

Who Moved the Access Keys

The larger shift happened quietly the same week. The White House formally took over the distribution of access rights to frontier AI models — directly deciding who gets to use the most advanced systems from OpenAI, Anthropic, Google, and Meta. This isn’t a new export control statute. It’s something upstream: the government reviews before users ever see the model. The tech industry’s claim that “deciding who uses proprietary models” has been transferred to a different office.

Also that week: SpaceX is in talks to become the U.S. Department of Defense’s cloud computing supplier. SpaceX has already signed compute agreements with both Google and Anthropic separately. If the DoD contract closes, a single compute node will serve both commercial AI and military AI applications. Meanwhile, Meta is evaluating a deal to rent compute capacity to Anthropic — $10 billion over two years, paid monthly. Three developments converged in one month: AI access nationalized, AI compute militarized, and AI model commercial alliances formalized.

What the Compute Map Looks Like

Draw this out and the topology becomes visible: chip fabrication → compute infrastructure → model → access approval. Every layer has someone positioned at the gate. NVIDIA’s H100s are manufactured at TSMC’s fabs. Compute gets rented through SpaceX and Meta. Models run on Anthropic and OpenAI infrastructure. The White House decides which institutions can connect.

TSMC sits at the bottom of this chain. Over 90% of advanced-node chips globally come from its fabs. NVIDIA’s dependence on TSMC is near total. The logic that follows is uncomfortable: the deeper the White House control framework goes, the more concentrated the demand for compute infrastructure becomes — and the more rigid the dependence on foundational chips. TSMC’s geopolitical value rises as the control architecture tightens.

That rise is not free. Brookings Institution analysis has noted a U.S. Treasury Secretary publicly warning that Taiwan’s concentrated chip production represents a “single point of failure” for the global economy. Higher strategic value invites pressure to relocate, replicate, or be used as leverage. Being irreplaceable in a supply chain is sometimes a risk, not a shield.

Is Taiwan on the Trusted Partner List

The “trusted partner” framework from the Mythos 5 reinstatement is the template for how AI capability will be distributed geopolitically going forward. Taiwan is not a direct priority recipient, but has been included in discussions around trusted AI supply chains. The position is awkward: the entity providing the foundational hardware may not be among the first to access the models running on it.

The structural problem runs deeper. When Anthropic released the Mythos model line, reporting noted that countries without their own compute infrastructure would be almost entirely dependent on a handful of companies — with near-zero influence over product design or safety policy. TSMC holds an irreplaceable position in the manufacturing layer. It has almost no seat at the table where AI standards are set. Controlling over 90% of advanced-node fabrication, without being in the room where the model governance decisions happen.

TSMC posted Q1 2026 revenue of $35.7 billion, up 35% year-over-year. That number is now simultaneously a geopolitical asset and a geopolitical target.

— 邱柏宇

延伸閱讀

短網址分享
https://justfly.idv.tw/s/FLAiRvx