The New Space Race: Seizing the Means of Intelligence Production¶
English Summary¶
Mary, writing from her experience working on Capitol Hill, argues that the AI race is fundamentally a resource competition disguised as a software competition, and DeepSeek's release represents this era's Sputnik moment—not because of technical achievement alone, but because it shattered the psychological assumption that America's AI lead was insurmountable. Unlike the space race which had a finish line, the AI race is a contest to control the infrastructure upon which all future races will be run, requiring sovereign control over the entire physical stack from rare earth mines to nuclear power plants. China spent two decades integrating supply chains while America spent two decades offshoring them, creating a structural asymmetry that explains the Trump administration's sudden resource nationalism and interest in Greenland's rare earth deposits.
The DeepSeek Sputnik Moment: Fear as Political Catalyst¶
The pattern of government action: Mary's core insight from Capitol Hill is that the U.S. government runs primarily on fear—politicians only overcome friction when both parties share a common enemy. The historical pattern repeats: Sputnik → NASA/DARPA/Interstate Highway System; 5G anxiety → Huawei surveillance panic; DeepSeek → AI infrastructure mobilization.
What made DeepSeek shocking: On January 20, 2025 (President Trump's second inauguration day), a Chinese AI startup released its R1 model with intentional timing. Within a week, DeepSeek overtook ChatGPT as the most downloaded free app on Apple's U.S. App Store. NVIDIA lost $600 billion in market value in a single day. The psychological shock wasn't just technical achievement—it was discovering the gap Americans assumed existed was far narrower than believed.
The reported economics: DeepSeek claimed development costs of $5.6 million total with training costs under $300,000, compared to OpenAI and companies backed by tens of billions in American capital. (Mary notes her AI lab friends now reassure her that DeepSeek is "a bit of a fork" and innovative ideas remain in America's domain, but the damage to perception was done.)
Why this differs from the space race: The space race had a destination (the Moon). AI has no finish line—it's a race to control the infrastructure upon which all future economic activity will run. As Mary frames it: "Humanity stretches its fingertips towards owning the means of intelligence production."
The Asymmetry: China's Infrastructure Approach vs. America's Product Approach¶
China's strategic integration: Beijing treats AI as strategic infrastructure to be deployed across the entire apparatus of state and economy. The integration is mandatory, centralized, total. There is no distinction between commercial application and national strategy—they are the same thing.
Concrete example—Agricultural AI deployment: - Diantian Farm (Shanghai outskirts): 70+ engineers "herd" AI-enabled robots via WeChat mini-programs - Weeding robots: 8 hours of operation on 1 hour charge, covering 33 hectares per day, using computer vision trained on millions of images to distinguish crops from weeds - Xinjiang unmanned farm: 3,000-mu facility with 75% automation, achieving cotton yields of 529 kg/mu (previously unimaginable) - National policy: January 2025 State Council No. 1 Central Document (highest-priority agricultural directive) identifies "new quality productive forces in agriculture" as top priority, mandating AI integration from seed to harvest - Scale: 5,000+ agricultural drones powered by BeiDou satellite system; agricultural robotics market projected to grow from $750M (2022) to $3B (2030)
Why China must succeed: China must feed 1.4 billion people with less than 10% of the world's arable land and even less freshwater. Food self-sufficiency declined from 93.6% (2000) to 65.8% (2020), projected to fall to 58.8% (2030). Average farmer age exceeds 50. AI is not productivity enhancement—it's existential survival strategy.
America's fragmented adoption: AI adoption in America is driven almost entirely by private enterprise. Americans treat AI as a product—something to subscribe to, optimize workflow, generate social media content. Integration is voluntary, commercial, piecemeal. Federal agencies still run on legacy systems. AI adoption happens in spite of the state, not because of it. The Farm Bill debates center on subsidy allocation, not technological transformation.
The Nine-Layer AI Supply Chain: Chokepoints and Concentration Risk¶
Mary structures the AI supply chain as a vertical stack where bottlenecks at lower layers propagate upward:
Layer 1: Raw Materials (Rare Earths, Copper, Silicon, Uranium)¶
China's rare earth dominance: - 61% of global rare earth mining - Over 90% of refining capacity - 94% of permanent magnet production (essential for EVs, wind turbines, hard drive motors)
Export control escalation: - April 2025: China imposed export controls on 7 medium/heavy rare earth elements (samarium, gadolinium, terbium, dysprosium, lutetium, scandium, yttrium) - October 2025: Beijing asserted jurisdiction over foreign-made products containing Chinese-origin rare earth materials
Copper supply crunch: - December 2025: $11,705 per tonne (record high), up 31% year-to-date - IEA warns of potential 30% supply shortfall by 2035 - AI data centers emerging as significant new demand: Bloomberg NEF estimates 500,000+ tonnes annually by 2030 - Structural constraint: New U.S. copper mines take 29 years on average to permit and build; ore grades fallen 40% since 1991
Nuclear renaissance for AI power: - Microsoft: 20-year PPA with Constellation Energy to restart Three Mile Island - Meta: 1.1 GW from Clinton plant - Amazon: 2 GW from Susquehanna - Google: Partnership with Kairos Power for SMRs - Driver: Deloitte projects U.S. data center power capacity will grow from 33 GW (2024) to 176 GW (2035)—more than 5x increase - Only carbon-free baseload source that can deliver 24/7 reliability AI workloads require - Enriched uranium prices surged to $190/SWU from $56 three years ago (Russia controls 40% of global enrichment capacity)
Layer 2 & 3: Semiconductor Equipment and Foundries¶
ASML's 100% monopoly on EUV lithography: - Only equipment capable of printing sub-7nm features required for cutting-edge AI chips - $150-200M per machine - 18 months to build - 250 crates to ship - Technical complexity: Requires 13.5nm light generated by vaporizing tin droplets with 50,000-watt CO2 laser, hitting each droplet twice, 50,000 times per second, reflected through mirrors polished to atomic smoothness - No other company has mastered this (Canon and Nikon abandoned EUV development decades ago) - Netherlands restricted EUV exports to China under U.S. pressure—arguably the most consequential export control in modern history - Without EUV, China cannot manufacture chips below ~7nm, creating a ceiling that increasingly constrains AI ambitions
TSMC's concentration risk: - 92% of the world's most advanced chips fabricated by TSMC - Every NVIDIA GPU, AMD data center processor, Apple chip, Amazon Graviton manufactured by TSMC - Taiwan sits 100 miles from mainland China (which claims the island as its territory) - Military conflict, blockade, or severe earthquake could trigger global technology crisis - CHIPS Act's $52 billion explicitly designed to reduce this concentration - Advanced packaging capability (Chip-on-Wafer-on-Substrate/CoWoS) equally critical—essential for stacking HBM memory directly on AI accelerators - Focus on TSMC Arizona factory and shift towards Intel for domestic production
Layer 4: Memory (High-Bandwidth Memory as Strategic Asset)¶
HBM's emergence as critical enabler: - Specialized DRAM architecture that stacks memory dies vertically and places adjacent to processor - Provides bandwidth necessary to feed data to AI accelerators at required rates - Capacity explosion: NVIDIA GB200 NVL72 rack contains 13.4 TB HBM vs 640 GB in previous DGX H100; upcoming GB300 increases to 21.7 TB; AMD MI400 Helios rack will contain 31.1 TB - AI servers use 34x more HBM content than previous generations
Market concentration: - SK Hynix: 62% market share - Samsung: 17% - Micron: 21% (only non-Korean supplier) - All three have production sold out through end of 2026
Micron's strategic importance: - Only U.S.-headquartered HBM supplier (geographic diversification from Korean concentration risk) - ~$200B in U.S. investments: 2 fabs in Idaho, up to 4 in New York, Virginia expansion, domestic HBM packaging capabilities - CHIPS Act funding: $6.165B in grants + $7.5B in loans - Stock performance reflects transformation: Micron +239% (2025), SanDisk +388%, Western Digital +219% - Memory undergoing structural shift from commodity to strategic asset - Micron CEO: "Memory is now essential to AI's cognitive functions, fundamentally altering its role from a system component to a strategic asset."
Layer 5: Processors (NVIDIA's CUDA Moat and Emerging Competition)¶
NVIDIA's dominance: - 94%+ share of discrete GPU market - Data center revenue: $51.2B in fiscal Q3 2026 - Moat rests not merely on hardware but on CUDA software ecosystem (two decades of developer investment)
Competitive evolution: - AMD MI350 series: Fastest-ramping product in AMD history - MI450 (launching Q3 2026 on TSMC 2nm): Targets direct competition with NVIDIA Blackwell and Rubin architectures - Multi-year partnership with OpenAI: 1 gigawatt of MI450 deployment in H2 2026 validates competitive threat - Hyperscaler custom silicon: Google TPUs, Amazon Inferentia/Trainium, Microsoft Azure Maia—all designed to reduce NVIDIA dependence
Geopolitical impact: - China represented 26% of NVIDIA revenue in FY2022 - Export restrictions reduced to ~13% in 2025 - Domestic alternatives like Huawei Ascend 910C scaling rapidly
Layer 7: Energy Infrastructure (Nuclear Renaissance)¶
Demand explosion: - U.S. electricity consumption growing 2.5% annually after 25 years of stagnation - Goldman Sachs projects 165% increase in data center power demand by 2030 - Interior Secretary Doug Burgum: "The U.S. must win the AI arms race, linking energy security to national security"
Market response: - Independent power producers with nuclear exposure (Constellation Energy, Vistra, Public Service Enterprise Group) surged - Westinghouse $80B contract for large-scale reactors signals policy direction
Greenland: The Logical Endpoint of Sovereignty Repricing¶
Why Greenland matters: - 8th largest rare earth reserves globally: 36-42 million metric tons of rare earth oxides (second only to China) - Largest rare earth reserves of any territory with zero active mines - Kvanefjeld deposit: 3rd largest land-based rare earth deposit on Earth - Contains 25 of 34 EU critical raw materials
The extraction challenge: - 80% covered in ice - Arctic mining 5-10x more expensive than elsewhere - Expert characterization: "Completely bonkers... might as well mine on the moon"
Yet development proceeds: - U.S. Export-Import Bank issued $120M letter of interest for Tanbreez - Prediction markets price 40% probability U.S. takes some form of control - Represents logical endpoint of sovereignty repricing trend: frozen landmass valued not for what it is but for what it contains
The Sovereignty Trade: Physical Reality Reasserts Itself¶
The structural repricing thesis: The world is returning to its natural state of naked competition for control of physical resources. The postwar rules-based order—in which sovereignty was subsidized by American hegemony and supply chains were optimized for efficiency rather than resilience—is being unwound. Physical possession is becoming the only law.
Price signals across the stack: - Uranium enrichment: $190/SWU (up from $56 three years ago) - Copper: Record highs with deficits projected through the decade - Memory stocks: Up 200-400% in a single year - Gold and silver: Hitting 45-year simultaneous records as central banks accumulate - Greenland: Subject of great power competition
Mary's unifying framework: These are not disconnected phenomena—they are manifestations of a single underlying trend: the repricing of sovereignty itself. The companies and commodities that form the substrate of intelligence production are not merely AI beneficiaries; they are the physical foundation upon which the future of human-machine intelligence will be built.
The Cybernetic Manifesto: China's Mandatory Integration vs. America's Voluntary Adoption¶
The prediction vs. reality: The Cybernetic Manifesto predicted humanity would become cyborgs fused with machines into seamless informational systems by the late 20th century. The authors were directionally correct but temporally incorrect—this fusion is happening now, faster than imagined.
The two paths diverging: - America: AI as product → voluntary, commercial, piecemeal adoption - China: AI as infrastructure → mandatory, centralized, total integration
The competitive question: Can America rebuild the industrial base it dismantled over decades of peace, secure the resources it neglected, and deploy AI at the scale and speed that strategic competition demands? China has a 20-year head start on supply chain integration, a political system capable of mandating adoption, and a population accustomed to state-directed technological transformation.
Mary's conclusion: Robots harvesting rice in Shanghai suburbs are a signal. The country that can grow food, generate power, manufacture chips, and train AI models without depending on rivals will dominate the 21st century. The cybernetic fusion is coming either way. The only question is who controls the mines, refineries, and foundries that make the code possible.
繁體中文總結¶
Mary 根據其在美國國會山工作的經驗指出，AI 競賽本質上是一場偽裝成軟體競爭的資源競爭，而 DeepSeek 的發布代表了這個時代的「史普尼克時刻」——不僅因為技術成就本身，更因為它粉碎了美國 AI 領先優勢不可撼動的心理假設。與有終點線的太空競賽不同，AI 競賽是爭奪未來所有競賽都將運行在其上的基礎設施控制權，需要從稀土礦到核電廠整個物理堆疊的主權控制。中國花了二十年整合供應鏈，而美國花了二十年外包供應鏈，這種結構性不對稱解釋了川普政府突然的資源民族主義以及對格陵蘭稀土儲量的興趣。
DeepSeek 史普尼克時刻：恐懼作為政治催化劑¶
政府行動模式： Mary 從國會山得出的核心洞察是，美國政府主要靠恐懼運作——政客只有在兩黨共享共同敵人時才能克服行動摩擦。歷史模式重複：史普尼克 → NASA/DARPA/州際公路系統；5G 焦慮 → 華為監控恐慌；DeepSeek → AI 基礎設施動員。
DeepSeek 的衝擊點： 2025 年 1 月 20 日（川普總統第二次就職日），中國 AI 新創公司刻意選在這天發布 R1 模型。一週內，DeepSeek 超越 ChatGPT 成為 Apple 美國 App Store 下載量最高的免費應用。NVIDIA 單日市值蒸發 6000 億美元。心理衝擊不僅是技術成就——而是發現美國人假設存在的差距遠比想像中小。
報導的經濟數據： DeepSeek 聲稱開發成本總計 560 萬美元，訓練成本低於 30 萬美元，相比之下 OpenAI 和其他公司獲得數百億美元資本支持。（Mary 指出她在 AI 實驗室的朋友現在向她保證 DeepSeek「有點像分叉」，創新想法仍在美國領域，但對認知的損害已造成。）
為何與太空競賽不同： 太空競賽有目的地（月球）。AI 沒有終點線——它是爭奪所有未來經濟活動都將運行在其上的基礎設施控制權的競賽。Mary 的框架：「人類正伸手觸及擁有智能生產手段的控制權。」
不對稱：中國的基礎設施方法 vs 美國的產品方法¶
中國的戰略整合： 北京將 AI 視為戰略基礎設施，在國家和經濟的整個機構中部署。整合是強制性的、集中化的、全面的。商業應用與國家戰略之間沒有區別——它們是同一件事。
具體案例——農業 AI 部署： - 店田農場（上海郊區）： 70+ 工程師透過微信小程序「放牧」AI 機器人 - 除草機器人： 充電 1 小時運行 8 小時，每天覆蓋 33 公頃，使用在數百萬張圖像上訓練的電腦視覺區分作物和雜草 - 新疆無人農場： 3000 畝設施，75% 自動化，實現 529 公斤/畝棉花產量（以前無法想像） - 國家政策： 2025 年 1 月國務院一號中央文件（最高優先級農業指令）將「農業新質生產力」確定為首要任務，要求從種子到收穫的 AI 整合 - 規模： 5000+ 架由北斗衛星系統驅動的農業無人機；農業機器人市場預計從 7.5 億美元（2022）增長到 30 億美元（2030）
中國為何必須成功： 中國必須用不到全球 10% 的可耕地和更少的淡水養活 14 億人。糧食自給率從 93.6%（2000）下降到 65.8%（2020），預計到 2030 年降至 58.8%。農民平均年齡超過 50 歲。AI 不是生產力提升——是生存戰略。
美國的碎片化採用： 美國的 AI 採用幾乎完全由私營企業驅動。美國人將 AI 視為產品——訂閱的東西、優化工作流程、生成社交媒體內容。整合是自願的、商業的、零碎的。聯邦機構仍在舊系統上運行。AI 採用是儘管有國家而發生，而非因為國家。農業法案辯論集中在補貼分配，而非技術轉型。
九層 AI 供應鏈：瓶頸與集中風險¶
Mary 將 AI 供應鏈結構化為垂直堆疊，底層瓶頸會向上傳播：
第 1 層：原材料（稀土、銅、矽、鈾）¶
中國的稀土主導地位： - 全球稀土開採 61% - 提煉產能超過 90% - 永磁體生產 94%（對電動車、風力渦輪機、硬碟馬達至關重要）
出口管制升級： - 2025 年 4 月：中國對 7 種中重稀土元素實施出口管制（釤、釓、鋱、鏑、鑥、鈧、釔） - 2025 年 10 月：北京宣稱對含有中國來源稀土材料的外國製產品擁有管轄權
銅供應緊縮： - 2025 年 12 月：每噸 11,705 美元（創紀錄高點），年增 31% - IEA 警告到 2035 年可能出現 30% 供應短缺 - AI 資料中心成為重要新需求：Bloomberg NEF 估計到 2030 年每年 50 萬噸以上 - 結構性限制：美國新銅礦平均需要 29 年許可和建設；礦石品位自 1991 年下降 40%
核電復興為 AI 供電： - 微軟：與 Constellation Energy 簽訂 20 年 PPA 重啟三哩島 - Meta：從 Clinton 廠 1.1 GW - Amazon：從 Susquehanna 2 GW - Google：與 Kairos Power 合作 SMR - 驅動力：Deloitte 預測美國資料中心電力容量將從 33 GW（2024）增長到 176 GW（2035）——超過 5 倍 - AI 工作負載需要的唯一無碳基載電源能提供 24/7 可靠性 - 濃縮鈾價格從三年前 56 美元飆升至 190 美元/SWU（俄羅斯控制全球濃縮產能 40%）
第 2 & 3 層：半導體設備與晶圓廠¶
ASML 對 EUV 光刻的 100% 壟斷： - 唯一能印刷尖端 AI 晶片所需亞 7nm 特徵的設備 - 每台 1.5-2 億美元 - 需 18 個月製造 - 需 250 個板條箱運輸 - 技術複雜性：需要 13.5nm 光，由 50,000 瓦 CO2 激光蒸發錫滴產生，每秒擊打每個液滴兩次，50,000 次，通過拋光至原子級平滑的鏡子反射 - 沒有其他公司掌握此技術（Canon 和 Nikon 幾十年前放棄 EUV 開發） - 荷蘭在美國壓力下限制 EUV 出口至中國——可以說是現代史上最重要的出口管制 - 沒有 EUV，中國無法製造低於約 7nm 的晶片，創造越來越限制其 AI 野心的上限
TSMC 的集中風險： - 全球 92% 最先進晶片由 TSMC 製造 - 每個 NVIDIA GPU、AMD 資料中心處理器、Apple 晶片、Amazon Graviton 都由 TSMC 製造 - 台灣距離中國大陸 100 英里（中國聲稱該島為其領土） - 軍事衝突、封鎖或嚴重地震可能引發全球技術危機 - CHIPS 法案的 520 億美元明確旨在減少這種集中 - 先進封裝能力（晶片晶圓基板/CoWoS）同樣關鍵——對於將 HBM 記憶體直接堆疊在 AI 加速器上至關重要 - 重點關注 TSMC 亞利桑那工廠和轉向 Intel 進行國內生產
第 4 層：記憶體（高頻寬記憶體作為戰略資產）¶
HBM 的崛起作為關鍵推動力： - 專用 DRAM 架構，垂直堆疊記憶體晶片並放置在處理器旁 - 提供以所需速率向 AI 加速器饋送數據所需的頻寬 - 容量爆炸：NVIDIA GB200 NVL72 機架包含 13.4 TB HBM vs 前代 DGX H100 的 640 GB；即將推出的 GB300 增加到 21.7 TB；AMD MI400 Helios 機架將包含 31.1 TB - AI 伺服器使用的 HBM 內容是前代的 34 倍
市場集中度： - SK Hynix：62% 市佔率 - Samsung：17% - Micron：21%（唯一非韓國供應商） - 所有三家的產能都銷售到 2026 年底
Micron 的戰略重要性： - 唯一美國總部的 HBM 供應商（韓國集中風險的地理多元化） - 約 2000 億美元美國投資：愛達荷州 2 座晶圓廠、紐約最多 4 座、維吉尼亞擴張、國內 HBM 封裝能力 - CHIPS 法案資金：61.65 億美元贈款 + 75 億美元貸款 - 股票表現反映轉型：Micron +239%（2025）、SanDisk +388%、Western Digital +219% - 記憶體從商品轉變為戰略資產 - Micron CEO：「記憶體現在對 AI 的認知功能至關重要，從根本上改變了它從系統組件到戰略資產的角色。」
第 5 層：處理器（NVIDIA 的 CUDA 護城河與新興競爭）¶
NVIDIA 的主導地位： - 獨立 GPU 市場 94%+ 市佔率 - 資料中心收入：2026 財年第三季 512 億美元 - 護城河不僅在硬體，更在 CUDA 軟體生態系統（二十年開發者投資）
競爭演化： - AMD MI350 系列：AMD 歷史上增長最快的產品 - MI450（2026 年第三季在 TSMC 2nm 上推出）：直接與 NVIDIA Blackwell 和 Rubin 架構競爭 - 與 OpenAI 多年合作：2026 下半年部署 1 gigawatt MI450 驗證競爭威脅 - 超大規模自定義矽：Google TPU、Amazon Inferentia/Trainium、Microsoft Azure Maia——都旨在減少對 NVIDIA 的依賴
地緣政治影響： - 中國占 NVIDIA 2022 財年收入的 26% - 出口限制將其減少到 2025 年約 13% - 國內替代品如華為 Ascend 910C 快速擴展
第 7 層：能源基礎設施（核電復興）¶
需求爆炸： - 美國電力消耗在 25 年停滯後以每年 2.5% 增長 - Goldman Sachs 預測到 2030 年資料中心電力需求增加 165% - 內政部長 Doug Burgum：「美國必須贏得 AI 軍備競賽，將能源安全與國家安全聯繫起來」
市場反應： - 具有核曝險的獨立發電商（Constellation Energy、Vistra、Public Service Enterprise Group）飆升 - Westinghouse 800 億美元大型反應爐合約標誌政策方向
格陵蘭：主權重新定價的邏輯終點¶
為何格陵蘭重要： - 全球第 8 大稀土儲量：3600-4200 萬公噸稀土氧化物（僅次於中國） - 零活躍礦場的任何領土中最大的稀土儲量 - Kvanefjeld 礦床：地球上第 3 大陸基稀土礦床 - 包含 34 種歐盟關鍵原材料中的 25 種
開採挑戰： - 80% 被冰覆蓋 - 北極採礦成本比其他地方高 5-10 倍 - 專家評價：「完全瘋狂……不如在月球上採礦」
然而開發仍在進行： - 美國進出口銀行為 Tanbreez 發出 1.2 億美元意向書 - 預測市場對美國採取某種形式控制的可能性定價為 40% - 代表主權重新定價趨勢的邏輯終點：凍土地塊的價值不在於它是什麼，而在於它包含什麼
主權交易：物理現實重新主張自己¶
結構性重新定價論點： 世界正在回歸其自然狀態——對物理資源控制權的赤裸競爭。戰後基於規則的秩序——主權由美國霸權補貼、供應鏈為效率而非韌性優化——正在被瓦解。物理佔有正在成為唯一法則。
整個堆疊的價格信號： - 鈾濃縮：190 美元/SWU（三年前為 56 美元） - 銅：創紀錄高點，預計整個十年赤字 - 記憶體股票：單年上漲 200-400% - 黃金和白銀：央行累積時創 45 年同時紀錄 - 格陵蘭：大國競爭的主體
Mary 的統一框架： 這些不是分散的現象——它們是單一潛在趨勢的表現：主權本身的重新定價。構成智能生產基礎的公司和商品不僅是 AI 受益者；它們是人機智能未來將建立在其上的物理基礎。
控制論宣言：中國的強制整合 vs 美國的自願採用¶
預測 vs 現實： 控制論宣言預測人類將在 20 世紀末成為與機器融合成無縫信息系統的賽博格。作者方向正確但時間錯誤——這種融合現在正在發生，比想像的更快。
兩條分歧的道路： - 美國： AI 作為產品 → 自願、商業、零碎採用 - 中國： AI 作為基礎設施 → 強制、集中、全面整合
競爭問題： 美國能否重建在數十年和平中拆除的工業基礎，確保被忽視的資源，並以戰略競爭要求的規模和速度部署 AI？中國在供應鏈整合上有 20 年領先優勢、能夠強制採用的政治體系，以及習慣於國家指導技術轉型的人口。
Mary 的結論： 上海郊區收穫稻米的機器人是一個信號。能夠在不依賴對手的情況下種植食物、發電、製造晶片和訓練 AI 模型的國家將主宰 21 世紀。控制論融合無論如何都會到來。唯一的問題是誰控制使代碼成為可能的礦山、煉油廠和鑄造廠。
Key Quotes¶
"In space there is no place to hide. From space, masters of the earth would have the power to control the world."
"My biggest takeaway after working on Capitol Hill was that our government runs primarily on fear. Politicians only overcome the friction of taking action when both parties share a common enemy."
"The space race was about space to the extent that it was a national defense priority; it was also about demonstrating technological supremacy and proving that American systems could out-innovate Soviet systems."
"What made DeepSeek a Sputnik moment was not merely technical achievement, it was the psychological shock of discovering the gap that Americans convinced themselves existed was much closer than assumed."
"The space race had a finish line. AI does not. It is not a race towards a destination, it is a race to control the infrastructure upon which all future races will be run. Humanity stretches its fingertips towards owning the means of intelligence production."
"Every industry, every ounce of economic activity, every instrument of state will run on AI infrastructure. Whichever nation controls the means of intelligence production controls the terms on which everyone else competes. The rest will pay tribute or be cut off entirely."
"In America, artificial intelligence adoption has been driven almost entirely by private enterprise. The integration is voluntary, commercial, piecemeal. Federal agencies still run on legacy systems. AI adoption in America is happening in spite of the state, not because of it."
"Beijing treats AI as strategic infrastructure; a capability to be deployed across the entire apparatus of state and economy. The integration is mandatory, centralized, total. There is no distinction between commercial application and national strategy. They are the same thing."
"AI is not a productivity enhancement for China. It is an imperative survival strategy for the nation."
"You cannot build advanced semiconductors without rare earth elements. You cannot process rare earth elements without massive energy inputs. You cannot generate that energy without uranium, natural gas, or grid infrastructure. You cannot train frontier AI models without advanced chips. Every layer of the AI stack rests on physical foundations that America does not fully control."
"China controls 60% of rare earth mining and 90% of rare earth processing. China dominates the refining of cobalt, lithium, and graphite, the materials that power the batteries that power the data centers that power the AI. China has spent two decades securing supply chains while America spent two decades offshoring them."
"The Trump administration's resource nationalism is not random belligerence. It is a recognition that sovereignty in the AI age requires control over the physical inputs to intelligence production."
"If there is a single company that embodies the fragility of the AI supply chain, it is ASML Holding. The Dutch firm maintains a 100% monopoly on extreme ultraviolet lithography machines."
"Without EUV, China cannot manufacture chips below approximately 7nm, a ceiling that will increasingly constrain its AI ambitions."
"Memory is now essential to AI's cognitive functions, fundamentally altering its role from a system component to a strategic asset." — Micron CEO
"The world is returning to its natural state of naked competition for control of physical resources. The postwar rules-based order, in which sovereignty was subsidized by American hegemony and supply chains were optimized for efficiency rather than resilience, is being unwound. Physical possession is becoming the only law."
"The companies and commodities that form the substrate of intelligence production are not merely AI beneficiaries, they are the physical foundation upon which the future of human-machine intelligence will be built."
"Robots harvesting rice in the Shanghai suburbs are a signal. The country that can grow food, generate power, manufacture chips, and train AI models without depending on rivals will dominate the 21st century."
"The cybernetic fusion is coming either way. The only question is who controls the mines, refineries, and foundries that make the code possible."
Investment Implications & Chokepoint Analysis¶
Mary's framework identifies specific chokepoints worth monitoring for "sovereignty trade" exposure:
Layer 1 (Raw Materials): - Rare earths: China 90% processing control → U.S. domestic mining/processing plays - Copper: 30% supply shortfall by 2035 → copper producers with new mine development - Uranium: $190/SWU (up from $56) → enrichment capacity, SMR developers - Nuclear power PPAs: Microsoft/Meta/Amazon/Google all securing capacity → independent power producers with nuclear exposure (Constellation, Vistra, PSEG)
Layer 2-3 (Equipment/Foundries): - ASML: 100% EUV monopoly → single point of failure, no substitutes - TSMC: 92% advanced chip manufacturing → geographic concentration risk, CHIPS Act beneficiaries (Intel domestic production)
Layer 4 (Memory): - HBM emergence as strategic asset: SK Hynix 62%, Samsung 17%, Micron 21% - Micron: Only U.S.-headquartered HBM supplier → sovereign exposure, $200B U.S. investment, CHIPS Act funding - Memory stock performance: Micron +239%, SanDisk +388%, WD +219% (2025)
Layer 5 (Processors): - NVIDIA: 94% GPU share, CUDA moat → dominance but geopolitical exposure (China 26% → 13%) - AMD: MI350/MI450 fastest-ramping, OpenAI partnership → emerging competition - Hyperscaler custom silicon: Google TPU, Amazon Inferentia/Trainium, Microsoft Azure Maia → NVIDIA de-risking
Greenland exposure: - U.S. Export-Import Bank: $120M letter of interest for Tanbreez - Prediction markets: 40% probability U.S. takes control - 36-42M metric tons rare earth oxides, 8th largest reserves globally
Thematic basket: - Physical sovereignty: Gold, silver, copper, uranium - Energy infrastructure: Nuclear power producers, SMR developers - Memory: Micron (U.S. HBM sovereign play) - Rare earths: Domestic mining/processing (vs. China 90% control)
Personal Reflection¶
This essay is exceptional for several reasons. First, Mary's Capitol Hill experience gives her a credible lens on how fear drives U.S. policy—the DeepSeek → Sputnik parallel isn't just metaphor, it's pattern recognition of how both parties mobilize around perceived threats. Second, the nine-layer supply chain breakdown is the clearest articulation I've seen of why the AI race is fundamentally a resource race disguised as a software race. Third, the China agricultural AI examples (70 engineers herding robots via WeChat, 529 kg/mu cotton yields) are concrete evidence of mandatory state-directed integration that differs fundamentally from America's voluntary commercial approach.
The most valuable insight: bottlenecks at lower layers propagate upward. If China controls 90% of rare earth processing and asserts jurisdiction over foreign-made products containing Chinese-origin materials (October 2025 escalation), it doesn't matter if America has better AI models—the physical stack is compromised. Similarly, ASML's 100% EUV monopoly and TSMC's 92% advanced chip manufacturing create single points of failure that no amount of CUDA optimization can overcome.
The Greenland analysis elevates the discussion beyond "Trump wants to buy Greenland" memes to the logical endpoint of sovereignty repricing: a frozen landmass valued not for habitability but for 36-42 million metric tons of rare earth oxides and 25 of 34 EU critical raw materials. When experts call Arctic mining "completely bonkers...might as well mine on the moon" yet the U.S. Export-Import Bank issues $120M letters of interest anyway, the desperation becomes clear.
What's missing: Mary doesn't deeply engage with the counterargument that DeepSeek's reported \(5.6M development / <\)300K training costs might be understated or that "innovative ideas remain in America's domain" (her AI lab friends' reassurance). If DeepSeek is truly "a bit of a fork," the Sputnik parallel weakens—but the perception damage and policy mobilization are real regardless.
The sovereignty trade thesis is compelling: uranium $190/SWU (up from $56), copper at records with decade-long deficits, memory stocks +200-400%, gold/silver 45-year simultaneous highs. These aren't disconnected—they're the market repricing physical control as the foundation of AI dominance. The country that can grow food (agricultural AI), generate power (nuclear/grid), manufacture chips (EUV/foundries/HBM), and train models (compute) without depending on rivals will dominate. China has a 20-year supply chain head start and mandatory integration capacity. America is rebuilding industrial base after decades of offshoring.
The cybernetic fusion is inevitable. The only question is who controls the mines, refineries, and foundries that make the code possible.