文 | 高见Pro
9月28日，AMD宣布以约82亿美元的全股票交易收购World Labs，即李飞飞创办的空间智能公司。这是AMD历史上第二大收购，仅次于当年约500亿美元收官的赛灵思。
比价码更值得关注的，是买卖双方的身份：买方苏姿丰，执掌市值约1万亿美元的算力巨头；卖方李飞飞，ImageNet的缔造者、全球AI科学家的旗帜性人物。
按黄仁勋在GTC大会上的说法，AI产业是一块"五层蛋糕"，能源、芯片、基础设施、模型、应用。这笔交易的意味在于：芯片层的人，买下了模型层的门票。
掏钱的和收钱的，都是华人。
黄仁勋讲"五层蛋糕"，本意是讲英伟达的全栈布局。但它同时提供了一把意外的尺子。沿这五层逐层看下去，会发现华人掌控了四层的半壁江山。只是每一层"掌控"的含义不同，五层之间还隔着一道看不见的墙。
前不久，还有两则媒体透露的消息备受瞩目。
旧金山，Anthropic聘请前白宫AI高级政策顾问黄思灏出任“前沿计算战略负责人”，统筹这家估值已冲击2万亿美元公司的芯片、电力与数据中心版图；
杭州，DeepSeek下一代模型或将推向8万亿参数，公司估值约5000亿元人民币，或正筹备科创板IPO。
清点“AI五层”中每一层里坐着的都是谁，会得到一张更精确、也更有意思的全球华人AI分布地图。苏姿丰与李飞飞的握手，黄思灏的新任职、梁文锋的动态，是目前这张地图上最新且醒目的坐标之一。P
能源层：面孔最稀薄、筹码最硬
华人掌控度：★★☆☆☆
先往最底下看。
能源，AI工厂的电力、冷却、供电架构。这层乍看与华人无关：叙事由美国人主导。
OpenAI公布佐治亚州“山茶花项目”，200亿美元自建数据中心；OpenAI与英伟达计划部署至少10GW的AI系统，相当于十座核电机组全力运转；亚马逊、微软、Alphabet、Meta、甲骨文五大云厂2026年资本开支指引合计约7350亿至7600亿美元，约占美国GDP的2.5%。
黄仁勋自己说，到2030年，全球AI基础设施投入将达3万亿至4万亿美元，“能源是最令人担忧的瓶颈”。新一代推理AI的能源需求，可能是现存需求的千倍级。
决策桌边确实坐满了美国人。但黄仁勋在2025年底的访谈里提及一个事实：中国的能源产能约为美国的两倍，且补贴力度大，约五折，这让中国在AI基础设施建设上具备显著优势。
更具体的注脚：2025年8月，英伟达更新800V直流供电架构合作伙伴名录，中国企业英诺赛科（氮化镓功率芯片）和麦格米特（电源）入选，成为AI工厂下一代供电体系的中国供应商。
这条供应链，比看上去的长。
郑崇华1971年创办的台达电子，如今是英伟达AI机柜电源与散热的头部供应商；大陆一侧，曹仁贤的阳光电源把光伏逆变器做到全球出货第一，曾毓群的宁德时代把储能电池做到全球第一。发电、变流、储电、供电，能源链条上每一环的隐形冠军，几乎都坐着一家华人企业。
硅谷决定在哪里建“电厂”，但把电力接进AI机柜的“电工”，绕不开华人的工厂。
芯片层：五张华人面孔，十万亿市值
华人掌控度：★★★★☆
芯片，最广为流传的一层，也是最坚硬的一层。
截至2026年9月，全球半导体市值前列中，五家AI算力核心公司的掌舵人全是华人：
英伟达市值5万亿美元，全球第一，黄仁勋出生于中国台湾；台积电市值2.1万亿美元，魏哲家出生地为中国台湾南投；博通市值1.8万亿美元，陈福阳出生于马来西亚槟城；AMD市值约1万亿美元，苏姿丰出生于中国台湾台南；英特尔市值约6262亿美元，CEO陈立武出生于马来西亚。
上述五家公司市值合计超过10万亿美元，超过日本等多数经济体的GDP总量。GPT、Gemini等模型公司的芯片布局，至今需经过这五个人执掌的公司设计或制造。
名单还能拉长：Marvell由出生上海的戴伟立与印尼华裔周秀文兄弟联合创立；泛林集团创始人是广东出生的林杰屏；AI服务器之王Supermicro的创始人叫梁见后；手机芯片巨头联发科由蔡力行执掌。1993年黄仁勋创办英伟达时，是硅谷唯一的华人芯片CEO；三十多年后，“硅谷唯一”变成了“半导体层几乎全员”。
但这层必须写出它的边界。
芯片层的华人掌控集中在中下游的设计与制造，比如中国台湾地区与东南亚华裔。上游的三道闸门不在华人手里：EDA软件（新思、楷登）、EUV光刻（ASML）、HBM存储（SK海力士、三星）。
中国大陆的芯片从业者面对的是另一重现实。2026年，10万张国产芯片形成新格局，昇腾、摩尔线程、海光的推理集群，训练端还得在多家芯片巨头之间权衡。
值得注意的是：陈云霁、陈天石兄弟从中科院计算所的实验室孵化出寒武纪，2016年成立时少有人看懂。2025年8月股价一度超越贵州茅台，被称作“寒王”，成为国产AI芯片在资本市场的图腾。
基础设施层：被低估的“华人包工层”
华人掌控度：★★★★☆
2026年5月台北，一场被业内称为"兆元宴"的饭局上，黄仁勋的身旁坐着几乎清一色的华人面孔：台积电魏哲家、鸿海刘扬伟、广达林百里、纬创林宪铭、和硕童子贤、联发科蔡力行。
这场饭局的背景，是黄仁勋口中那个"AI工厂"的厂房、机柜、网络与光互联。这一层在舆论场存在感最低，却是华人密度和掌控度反差最大的一层。
黄仁勋自己把它叫“达链”。他的说法是：“芯片从这里来，封装从这里来，系统从这里造，AI超级计算机也在这里诞生。”这里的“这里”，指台湾地区：英伟达在全球有350余家ODM和零部件工厂，其中150家集中在台湾，每年为英伟达提供超过100万个部件。
产能格局更直观。TrendForce数据显示，广达、纬创、鸿海三家台系厂商联手拿下全球AI服务器近90%的ODM代工份额。
其中鸿海系一家占全球AI服务器代工约42%，是英伟达GB200/GB300整机柜的主力ODM；微软50%-60%的服务器代工订单、谷歌TPU服务器超六成的份额，都在它手里。
镜头转向大陆。这里的隐藏冠军是光模块，没有它，万卡集群就是一堆信息孤岛：
LightCounting 2026年5月榜单：全球光模块前十中，中国企业占七席，合计份额超过60%；中际旭创以近三成份额居全球第一，新易盛以15%-18%居第二；800G及以上高端数通光模块，中国厂商市占率突破70%。
中际旭创2026年上半年营收417.78亿元、净利136.51亿元；新易盛上半年营收209.1亿元，海外收入占比超90%。北美四大云厂每一座新建AI数据中心，大概率都要从中国企业采购光互联核心器件。
名单还能继续拉长：AI机柜的液冷散热，头部供应商是台湾的奇鋐与双鸿；高速铜缆与电连接，绕不开王来春的立讯精密。AI工厂不直接烧华人的钱，但它拧的每颗螺丝、插的每根线缆，几乎都能追溯到一家华人企业。
全球AI工厂的水泥、钢筋、管道和电线，几乎全是华人造的。
但这层有一个残酷的注脚：台系代工的毛利率普遍只有5%-8%，光模块稍高。华人密度最高的一层，恰恰是利润被压得最薄的一层。
模型层：密度最高，一分为二
华人掌控度：★★★☆☆
这里的故事最复杂，华人密度极高，但掌门权被地缘政治劈成了两半。
先看硅谷的密度。据相关统计，美国头部AI公司的核心研发团队里，华人占比高得反常：Meta超级智能实验室首批核心团队中华人约64%，扩编后有超20人本科毕业于中国大陆高校；xAI的创始团队中5位华人；OpenAI的GPT-4o核心团队中6位华人，研究副总裁翁荔本科毕业于北大；Cognition AI三位联合创始人全部是华裔。
他们脚下的地基，一半也是华人打的。
何恺明的ResNet是被引最高的AI论文之一，没有它就没有今天大模型的深度；李飞飞的ImageNet奠定了整个视觉识别范式；贾扬清的Caffe与ONNX成为全球AI工程的事实标准。
宏观口径更惊人。全球顶级AI研究者中，38%本科受教育于中国高校，是全球第一大单一来源。从某角度看，美国AI的优势，相当程度是中国教育体系供给、美国机构雇佣的。
MacroPolo追踪全球AI 人才流动研究后的原话：“America's AI advantage isn't homegrown — it's imported.”（美国的AI 优势并非源自本土培养的人才，而是靠引进海外人才，尤其来自中国，建立起来的）
但模型层硅谷的华人密度极高，掌舵者却几乎不是华人。
OpenAI的门面是Altman，Anthropic是Amodei兄弟，Google AI在Pichai手里，xAI属于马斯克。华人掌握了“how”，即怎么把模型训出来；而“what”和“why”，做什么、为什么做、讲什么故事，多数时候握在别人手里。黄思灏这类算力战略新角色，算是第一道被敲开的裂缝。
国内是另一支队伍。他们自己当掌门。
截至9月，梁文锋的DeepSeek，V4开源、估值约4000亿、8万亿参数在途；杨植麟的月之暗面，K3上线即冲进全球榜单第三，估值约500亿美元。姚顺雨从OpenAI到执掌腾讯混元，代表成建制人才回流。加上智谱、阿里通义、字节豆包.....
.2026年前九个月，八家国产大模型公司至少完成24次重要模型发布，节奏不输海外。
这份掌门人的名单也可以开得更长：闫俊杰的MiniMax用同一个多模态底座同时供养国内对话与海外娱乐两条产品线，姜大昕的阶跃星辰押注多模态跃迁，王小川把搜狗时代的产品方法论带进了百川智能，更早下场的还有李开复......
模型层的华人掌门，从不缺样本。但收入是这层的分水岭。
OpenAI预计2026年营收约360亿美元，年化收入8月已突破400亿；Anthropic年化收入7月约650亿美元，年底预计破千亿。这两家硅谷实验室的年化收入之和，超过中国全部大模型公司收入的总和。
尽管后者坐拥全球最大的统一应用市场。国内公司用免费和低价换规模，把“效率”打成了护城河，但效率暂时还没有兑换成收入量级。
应用层：中国产品真正“占领”的一层
华人掌控度：★★★★★
应用，价值兑现的地方。这一层，事实最一边倒。
AI产品榜（AICPB）2026年6月全球AI应用月活榜，前十名是：ChatGPT、字节豆包、阿里千问、谷歌Gemini、阿里夸克、DeepSeek、腾讯元宝、百度网盘、即梦AI、字节海外版Dola，除了ChatGPT和Gemini，前十名里的中国产品占了八席。
豆包+千问+DeepSeek三款合计月活超7亿，是全球AI应用第二梯队的主力。榜单拉长还有：MiniMax出海产品Talkie、Kimi、可灵AI。加上TikTok的AI化改造，中国产品的应用层出海是一条水面下的暗线。
更硬的是物理AI，黄仁勋口中下一个50万亿美元的市场。
Counterpoint数据：2026年上半年全球人形机器人出货超2.2万台，同比增长近300%；前五名全部是中国厂商，即智元、宇树、银河通用、优必选、乐聚，合计占86%。
这波物理AI里还有两个值得记住的华人。
一个是汪滔：大疆早已拿走全球消费级无人机约七成份额，证明华人公司可以在“软件定义的物理机器”上同时握住规模和定价权，那正是物理AI的上一次预演。
另一个是彭志辉，网名“稚晖君”，从华为“天才少年”计划走出来的B站科技顶流，如今是智元机器人的联合创始人：拿下全球人形机器人43%出货份额的那家公司，站着一个在社交媒体上长大的工程师偶像。
硬币的背面是单位价值。ChatGPT以9.95亿月活撑起360亿美元年收入；豆包3.24亿月活，仍以免费和低价获客为主。
也就是说，目前应用层的中国掌控是“用户级”的，还不是“利润级”的。规模全球第一，ARPU（单用户价值）仍是洼地。
中国公司擅长把应用做成基础设施，免费、规模、生态，美国公司擅长把应用做成订阅，付费、利润、复购。两种模式的对决，才刚到中盘。
沙漏、裂层与未定之局
从AI五层逐层清点，层与层的关系中还透露着一些关键信息。
华人的掌控度是一个沙漏”，底层芯片和次底层基础设施是境外华人的寡头领地，掌控度最高；顶层的应用，是大陆华人的规模领地，掌控度次之；中间的模型层，华人执行密度全球第一，但收入差距大，是全栈中最“薄”的一段，即沙漏的腰。
能源层则完全倒挂：目前金融与决策权在美国机构手里，物理产能和供应链在中国大陆手里。这是五层中唯一一层“谁都没有完整掌控”的，也是未来十年胜负最未定的一层。
当竞争的度量衡从“峰值算力”变成“每瓦特token产出”，电力就是AI时代的新石油，而全球一半的“油井”在中国。
五层对应四种“掌控”，兑换率天差地别。
芯片层掌控的是定价权。5万亿市值是定价权的资本化；基础设施层掌控的是产能权。九成代工份额、七成光模块，但毛利率5%-8%；产能权在供应链上的兑换率极低。
模型层国外华人握的是执行权、国内华人握的是定义权。定义权更值钱，所以国内掌门人的估值涨得更快；应用层握的是用户权，规模最大、ARPU最低。
一个残酷的规律：华人密度越高的环节，利润兑换率越低。代工层华人近乎垄断，毛利最薄；芯片层华人执掌半壁，但真正的定价者目前只有黄仁勋一人。
五层地图上最重要的地质构造，是那道从芯片层一直裂到模型层的墙。
黄仁勋的“达链”，150家台湾工厂、台系代工九成份额、五华人CEO的十万亿市值，嵌在美元叙事里；大陆的“国产链”，10万张昇腾推理集群、近8成的全球人形机器人份额、八席全球AI应用前十，在中国市场和人民币叙事里。
台南人黄仁勋和汕头人梁文锋，说的都是同一种母语，处在同一张价值链的两端，却几乎不可能出现在同一张供应链清单上。
历史上从未有过这样的结构：一个族群同时深度嵌入两个相互对冲的技术体系，并且各自都是对方最大的“备胎”。
更深入地看，每层的地位几乎都是上一个时代的“遗产”。
从某种程度上，芯片层是台湾地区三十年代工与半导体积累的存量；基础设施层是iPhone时代练出的制造能力在AI时代的复用；模型层是高考、竞赛、留学管道三十年积累的人才存量；应用层是移动互联网十年“APP工厂”方法论的自然变现。
换言之，华人今天在四层的领先，可视为旧时代的复利。
而真正的新变量只有两个：
一是能源层，中国装机两倍的先手对美国核电容量的存量，或将在2030年前见分晓；
二是模型层的定义权之争，当开源模型的地板以十分之一价格逼近闭源天花板，也就是DeepSeek、GLM、Kimi正在做的事，模型层“沙漏之腰”的宽度可能被撑开，那是唯一可能让墙两边重新坐回一张牌桌的力学。
芯片与制造的地盘是用三十年换来的，动摇它可能同样需要三十年；模型层的收入差距和应用层的ARPU洼地是明牌缺口；能源层是唯一尚未分配完毕的层。
牌桌是全世界的，牌是华人洗的，但底注谁下、赌注归谁，还没有定。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Documents
On 28 September, AMD announced the acquisition of World Labs, a space intelligence company founded by Li Fei Fei, with a full-stock deal of about $8.2 billion. This was the second-largest acquisition in AMD history, after the Serings of about $50 billion received in that year.
What is more interesting in comparison to the price is the identity of both the buyer and the buyer: the buyer Sozi, who has a market value of about $1 trillion, and the seller, Li Fei Fei, the founder of ImageNet, and the flag character of the global AI scientist.
According to Wong In-hoon at the GTC conference, the AI industry is a five-storey cake, energy, chips, infrastructure, models, applications. The deal is that people on the chip floor bought tickets to the model floor.
Those who pay and those who collect are Chinese.
Wong In-hoon said "Five Layers of Cake" and was meant to talk about the entire inn layout of Inweida. But it also provides an unexpected ruler. Looking down these five layers, the Chinese are in control of the four layers of the Half-Far River Mountain. It's just that every layer of "control" means something different, and there's an invisible wall between the five.
In the recent past, two other media reports were well received.
In San Francisco, Anthropic engaged former White House AI Senior Policy Adviser, Wong Si-chul, as “head of the front-line calculation strategy”, to consolidate the chip, power and data centre maps that the valuation had hit at $2 trillion;
In Hangzhou, the next generation of DeepSeek models will be pushed to 8 trillion parameters, with companies valuing about 500 billion yuan yuan or preparing an IPO.
Take an inventory of who sits on every floor of the "AI-5" and get a more accurate and interesting map of the global Chinese AI distribution. The handshake between Sozefeng and Li Fei Fei, the new assignment of Wong Si Swe, and the dynamics of Liang Wenbei are among the most recent and visible coordinates on the current map. P
Energy layer: the rarest face, the hardest chips
Chinese control: ★★☆☆☆
Look down first.
Energy, power, cooling, power structures at the AI plant. At first glance, it's not Chinese: the narrative is American-dominated.
Open AI announced the Georgia Mountain Tea Project, a $20 billion self-constructed data centre; Open AI and In Weida plan to deploy an AI system of at least 10 GW, equivalent to 10 nuclear power units operating at full capacity; and Amazon, Microsoft, Alphabet, Meta, Oracle Clouds' 2026 capital expenditure guidelines totalling some $735 billion to $760 billion, representing about 2.5 per cent of United States GDP.
Wong In-hoon himself said that global AI infrastructure investment would reach $3 trillion to $4 trillion by 2030, “energy being the most worrying bottleneck”. A new generation of reasoning on AI ' s energy needs could be a thousand times the existing demand.
The decision-making table is filled with Americans. However, Huang In-hoon, in an interview at the end of 2025, referred to the fact that China’s energy capacity is about twice that of the United States, and that it is heavily subsidized at about 50%, which gives China a significant advantage in the construction of AI infrastructure.
More specifically, in August 2025, Young Weida updated the 800V direct-flow power supply partnership list, and Chinese enterprises, Innosekor (Nitrogen Power Chip) and Magmet (electricity) were selected as Chinese suppliers of the next generation of electricity supply at the AI plant.
This supply chain is longer than it looks.
Zheng Zhonghua, founded in 1971, is now the head supplier of power from the Yin Wei Da ' ai cabinet; on the mainland, Cao In-hyun's solar power has made the light-volatilizers first in the world, and the former Ningde era has made the storage batteries first in the world. The invisible champions of power generation, currents, storage, electricity supply, and every ring of the energy chain are almost all sitting in a Chinese enterprise.
Silicon Valley decided where to build the “power plant”, but the “electrician” who attached the power to the AI cabinet could not bypass the Chinese factory.
Chip layer: five Chinese faces at a value of 10 trillion
Chinese control: ★★★★☆
Chips, the most widely circulated, are also the strongest.
As of September 2026, the five AI Nucleus companies were all Chinese:
Young Weidar's market value is $5 trillion, and Huang In-hoon is the world's first born in Taiwan; Taiwan's market value is $2.1 trillion for electricity accumulated in Taiwan, and Wei Cheol's family was born in South China; Chase's market value is $1.8 trillion and Chen Fookyang was born in Penang, Malaysia; AMD's market value is about $1 trillion, and Susef is born in Tainan, Taiwan, China; Intel's market value is about $626.2 billion and CEO Chen Li-Wu was born in Malaysia.
The combined market value of the five companies exceeds $10 trillion, exceeding the total GDP of most economies, such as Japan. The chip layout of model companies such as GPT, Gemini and others has to date been designed or manufactured by the companies in charge of these five individuals.
The list can also be extended: Marvell was created jointly by Dai Waili, born in Shanghai, and the Chinese brothers Zhou Xiu Wen; the founder of the Pan-Lin Group was Lin Jie-shik, born in Guangdong; the founder of the King Supermicro of the AI server was called Liang; and the UCDC was run by Cai Li. When Wong In-hoon founded In Weeda in 1993, it was the only Chinese chip in Silicon Valley, the CEO; more than 30 years later, Silicon Valley's only was transformed into a semiconductor layer.
But this layer must write its boundaries.
The Chinese on the chip layer control the design and construction concentrated in mid-downstream, such as Taiwan, China, and South-East Asian Chinese. Three gates upstream are not in the hands of the Chinese: EDA software (new ideas, role models), EUV light carving (ASML), HBM storage (SK Hercules, Samsung).
In mainland China, chip practitioners face another reality. In 2026, 100,000 nationally produced chips formed a new pattern, with a combination of stomping, mooring, sealight reasoning, and a trade-off between the training end and the giant chips.
It is worth noting that Chen Yunjun and Chen Tian Shik have hatched from the laboratories of the Central School of Calculus, which was rarely understood at its inception in 2016. In August 2025, stock prices once exceeded Guizhou's pagoda and were called "The Cold King" and became the totem in the capital market for domestically produced AI chips.
Infrastructure level: undervalued Chinese-contracted
Chinese control: ★★★★☆
In May 2026, in Taipei, Huang In-hoon sat next to an almost one-size-fits-all Chinese face: the House of Wei Cheol, Hong Hai Yang Wei, Guang Dalinbri, Lian Jiu Lin's constitution, and Suk-hyun, Uyoko Zai Li.
The background to this meal is the factory, the cabinet, the network and the light of the "Ai" factory in Hoang In-kwan. This level is the lowest in public opinion, but the highest in Chinese density and control.
Wong In-hoon himself called it "Closure". And he said, "Chips come from here, seals from here, systems are made from here, and AI supercomputers are born here." Here, the term “here” refers to the Taiwan region: Welda has more than 350 ODM and spare parts factories around the world, 150 of which are concentrated in Taiwan and provide over 1 million parts a year.
The energy patterns are more visible. TrendForce data show that the three major plants in China, Latitudinal, and Honghai jointly take nearly 90 per cent of ODM's share of the global AI server.
Of these, the Hong Hai System, which accounts for about 42 per cent of the global AI server workforce, is the main ODM for the GB200/GB300 whole cabinet in Weidar; it holds 50-60 per cent of Microsoft server orders and over 60 per cent of Google TPU servers.
shot turned to the mainland. The hidden champion here is the light module, without which the Vanka cluster is a collection of information islands:
LightCounting May 2026: China has seven seats in the top ten of the global light module, with a combined share of more than 60 per cent; China has the world ' s highest share of nearly 30 per cent in the Middle East and the second highest of 15-18 per cent in the New Mercure; and China ' s manufacturing city has a 70 per cent breakthrough in the high-end light module of 800G and above.
In the first half of 2026, the middle was set at $41,778 million, with a net gain of $13,651 million; in the first half of the first half of the new harvest, 20,911 million was received, and overseas revenues exceeded 90 per cent. Each of the four major cloud plants in North America will probably have to purchase light-connected cores from Chinese enterprises for each of the new AI data centres.
The list can also continue to grow: the liquid cooling heat in the Ai cabinet, with Taiwan's suppliers Qi Yi and Zhong, and high-speed copper cables connected to electricity to bypass Wang Laipun's well-informed advice. The AI factory does not burn Chinese money directly, but it screws up every screw, plugged in every cable, and can trace it almost to a Chinese enterprise.
The cement, steel bars, pipes and wires at the global AI plant are almost entirely Chinese-made.
There is, however, a cruel note: in general, only 5 to 8 per cent of the total number of employees working in the line are under 5 or 8 per cent and the light module is slightly higher. The highest concentration of Chinese is precisely the lowest profit.
Model layer: highest density divided into two
Chinese control: ★★★☆☆
The story here is the most complex, with a very high Chinese density, but the power is divided in half by geopolitics.
Look at the density of Silicon Valley first. According to statistics, the Chinese share of the core research and development team of head AI in the United States is abnormally high: about 64 per cent of the Chinese in the first core team of the Meta SuperIntelligence Laboratory, with more than 20 undergraduate graduates from mainland China; five Chinese in the founding team of XAI; six Chinese in the GPT-4o core team of Open AI, with the Vice-President of Research, Ong Yipenko, graduated from North China; and three co-founders of Cognion AI, all Chinese.
Half of them are Chinese.
Ho Keeming ' s ResNet is one of the most quoted AI papers without which the depth of today ' s LLM is not available; Lee ' s ImageNet lays out the whole visual recognition paradigm; Jayanq ' s Caffe and ONNX are the de facto standards for global AI projects.
The macro calibre is even more amazing. Of the top AI researchers globally, 38 per cent are undergraduates in Chinese universities, the world ' s largest single source. In one way or another, the advantages of AI in the United States are to a considerable extent the supply of the Chinese education system and the employment of American institutions.
MacroPolo tracks the story of AI after the study: “America's AI influence isn't hometown — it's built.” (American AI's advantage is not from indigenous talent, but rather from bringing in talent from abroad, especially from China).
However, the model Silicon Valley has a very high concentration of Chinese, and the rudder is almost non-Chinese.
Open AI's front is Altman, Anthropic is the Amodei brother, Google AI is in Pichai's hands, and XAI belongs to Mask. The Chinese have mastered “how”, that is, how to train models; and “what” and “whoy”, what to do, why to do, what to tell, most often in the hands of others. New players such as Huang Si-chul are the first cracks to be broken.
The country is another team. They're their own masters.
As of September, the DeepSeek, V4 open source, valued at about 400 billion and 8 trillion parameters, was on the way; on the dark side of Yang's moon, K3 went into third on the global list, with an estimate of about $50 billion. Yao Sun-yun, from Open AI to Tencent, represents a return of organic talent. Plus Zhipu, Ali Qwen, byteDoubao...
In the first nine months of 2026, eight national LLM companies completed at least 24 important models publish, with no loss of pace overseas.
The list of leaders can also be opened longer: Mini Max of Junjie uses the same multi-mode base to support both domestic dialogue and overseas entertainment lines, Kang's step-by-step stars are embroiled in a multi-mode leap, Wang Xiaochuan brings the product-s search method to the Mawakawa Intelligence, and sooner it ends up with Lee Zhuan...
The Chinese master of the model, never missing sample. But income is the watershed of this layer.
Open AI expects to collect about $36 billion in 2026, and annualized revenues have exceeded $40 billion in August; and anthropic annualized revenues of about $65 billion in July, which is expected to break hundreds of billions by the end of the year. The sum of the annualized revenues of the two Silicon Valley laboratories exceeds the sum of the revenues of all LLM companies in China.
Despite the fact that the latter dominates the largest global integrated application market. Domestic firms have converted “efficiency” into a moat for free and at low prices, but efficiency has not been converted to a revenue scale.
Application level: the real “occupation” of Chinese products
Chinese control: ★★★★★
Apply, value where it comes from. On this level, the truth is upside down.
The AI product list (AICPB) was the global AI application monthly list in June 2026, with the top 10 being ChatGPT, byteZ64Z, Ali, Google Gemini, Alikuk, DeepSeek, Tencent Dollar, Baidu Netboard, Jimeng AI, byte Overseas Dola, with the exception of ChatGPT and Gemini, with the top 10 Chinese products holding eight seats.
Doubao+000+DeepSeek three sections, which together account for over 700 million people per month, are the backbone of the second tier of the global AI application. The list is also long: MiniMax seagoing products Talkie, Kimi, Kling AI. Together with the AI conversion of TikTok, the application of Chinese products is a dark line under water.
Even harder is the physical AI, the next $50 trillion market of Huang Jin-hoon.
Counterpoint data: In the first half of 2026, global human robots produced more than 22,000 units, representing an increase of nearly 300 per cent over the same period; the top five were all Chinese manufacturers, namely Ji-won, Woo-ki, Galaxy General, Presumptuous Select, and Lottery, accounting for 86 per cent.
There are two Chinese in this wave of physics.
One is Wang-ho: Daeji has already taken about 70% of the global share of consumer-grade drones, proving that Chinese companies can hold both size and pricing power on a “soft-defined physical machine” that was the last preview of Physical AI.
The other is Peng Zhihui, a Bilibili technology leader from China, who is now a co-founder of Jiwon robots: the company that took 43 per cent of the global human robot's share and stood up for an engineer who grew up in social media.
The back of the coin is the unit value. ChatGPT earns $36 billion in annual income for 995 million months; Doubao324 million months are still dominated by free and low-cost clients.
In other words, the Chinese control of the current application level is “user level” and not “profit level”. Globally, the ARPU (single-user value) remains low.
Chinese companies are good at making applications into infrastructure, free of charge, of scale and of ecology, and American companies are good at making applications into subscriptions, paid, profit, buy-backs. The two models are in the middle.
Sandglasses, cracks and undecided situations
A layer-by-layer count from A5 also reveals some key information in the relationships between layers.
The control of the Chinese is a sandglass, and the bottom chips and sub-structures are the oligopoly of the overseas Chinese, with the highest control; the top-level applications are the size of the mainland Chinese, with the second most control; and the middle model layer, with the Chinese performing the highest density globally, with a large income gap, is the most “silent” part of the house, namely, the back of the sandglass.
The energy layer is completely upside down: financial and decision-making power is now in the hands of United States institutions, and physical capacity and supply chains are in the hands of mainland China. This is the only one of the five floors that “no one is in complete control” and the most uncertain level of victory in the next decade.
When the measure of competition changes from “peak calculus” to “per watttoken output”, electricity is the new oil of the AI era, while half of the world's “oil wells” are in China.
The five layers correspond to four “controls” and exchange rates vary from day to day.
The chip layer controls pricing rights. The market value of 5 trillion is the capitalization of pricing rights; it is the power of production that controls infrastructure. The 90-generation share, 70 per cent light module, but the Māori ratio is 5-8 per cent; the rate of exchange of power rights in the supply chain is extremely low.
The Chinese in the model are in executive power and the Chinese in China are in definition. As definitional rights are more valuable, domestic managers ' valuations have increased faster; the application layer holds user rights, with the largest size and lowest ARPU.
A cruel pattern: the higher the Chinese density, the lower the profit rate. The Chinese are almost monopolistic and the Māori are the thinest; the Chinese on the chip hold half the wall, but the real pricer is now only Huang In-hoon.
The most important geological formation on the five-story map is the wall from the chip layer to the model layer.
Huang In-hoon's “Closure Chain”, a 90 per cent share of Taiwan's 150 factories, Taiwan's successors, the billion-dollar market of the Chinese CEO, embedded in the dollar narrative; the mainland's “national chain”, a 100,000-piece cluster of reasoning, nearly 8% of the global human robotic share, and the top 10 of the eight global AI applications, in the Chinese market and the renminbi narrative.
Hoang In-hoon of Tainan and Liang Wen Sing, all speaking the same mother tongue, are at both ends of the same value chain, but it is almost impossible to appear on the same supply chain list.
Never before in history has there been a structure in which a community is embedded in two competing technological systems at the same time, each of which is the largest “stand-by” of the other.
In greater depth, the status of each layer was almost “hermit” of the previous era.
To a certain extent, the chip layer is the stock accumulated by workers and semiconductors in the Taiwan area in the 1930s; the infrastructure layer is the re-use of manufacturing capacity acquired in the iPhone era in the AI era; the model layer is the stock of talent accumulated in the 30 years of advanced examinations, competitions, study study pipelines; and the application layer is the natural realization of the 10-year mobile Internet “AP plant” methodology.
In other words, the Chinese leadership today on four levels can be seen as the dividends of the old era.
There are only two real variables:
The first is the energy layer, with twice as many Chinese first-hand loads of nuclear power in the United States, or will be known by 2030;
The other is the definitional rights of the model layer, which, when the floor of the open source model approaches the closed source ceiling at one tenth of the price, that is, what DeepSeek, GLM, Kimi are doing, the width of the model layer “the back of the sandglass” is likely to be stretched, the only mechanics that may allow both sides of the wall to return to a table.
Chips are traded for 30 years, and it may take another 30 years to shake them; the income gap at the modelling level and the ARPU depression at the application level are the visible gaps; and the energy layer is the only layer that has not yet been distributed.
The table is all over the world, and the Chinese wash it, but who will bet and who will bet?
Auto-fetched and translated by AI Atlas. All rights belong to the original author.