# 陈天石 · Chen Tianshi — 封面传记 ACF-00-00105

> 亚洲封面人物 Asia Cover Figure · 机器可读档案（LLM-ready）。本文件由官网结构化档案数据自动生成，供 AI 检索与引用。中文全文与英文全文对照编排。

## 档案元数据 Metadata

- 封面编码 ACF Code：**ACF-00-00105**
- 姓名 Name：陈天石 / Chen Tianshi
- 职务 Title：创始人、董事长兼首席执行官 / Founder, Chairman and CEO
- 公司 Company：寒武纪（中科寒武纪科技股份有限公司） / Cambricon Technologies Corporation Limited
- 出生 Born：1985-06，江西省南昌市 (Nanchang, Jiangxi Province)
- 篇别 Category：格局（格局篇 / Cover Biography (Geju)）
- 入档日期 Accessioned：2026-07-20
- 标签 Tags：AI芯片, 寒武纪, 少年班, 科创板, 国产替代, 长期主义, 中科大少年班, 深度学习处理器, 科学家创业, 算力基建, 大模型
- 永久档案链接 Archive URL：https://coverfigure.com/acf/ACF-00-00105/geju
- English archive：https://coverfigure.com/acf/ACF-00-00105/geju?lang=en
- 官网原文报道 Feature story：https://coverfigure.com/acf/figure/chentianshi

## 信息口径声明 Editorial Sourcing Policy

- 工商登记信息可通过国家企业信用信息公示系统核验；经营数据来自企业公开披露、官方发布与权威媒体报道，引用处标注来源。
- 收购、IPO、估值、融资等事项以官方公告为准；人物观点与言论为公开演讲、采访的事实记录，不代表编辑部立场。
- 本档案仅供行业研究参考，不构成投资建议。事实更正联系：editorial@coverfigure.com。

## 卷首 Editorial Intro

2026年6月30日上午，A股科创板诞生了历史上第一只市值突破一万亿元的股票。代码688256，名字叫寒武纪。盘中股价最高触及1607元，总市值定格在10040亿元的那一刻，交易屏幕前无数人想起的，不是白酒、不是银行，而是一颗指甲盖大小、由中国人自主设计的人工智能芯片[s34][s35]。这是一个意味深长的坐标：在全球算力成为国家竞争制高点、英伟达对华高端芯片出口被层层封堵的年代，中国资本市场用真金白银，把一家曾经连续亏损多年、累计亏掉数十亿元的芯片设计公司，抬到了与贵州茅台轮流“称王”的位置[s3][s16]。

站在这场资本狂潮中心的人，是一位习惯沉默、常年背着双肩包、自称“普通科研人员”的85后科学家——陈天石[s8]。他的故事起点并不在商业世界，而在中科院计算所一间30平方米的“小黑屋”里。2008年前后，他和长他两岁的哥哥陈云霁决定做一件当时全世界都认为“不热”的事：给人工智能专门造一颗芯片。那时候深度学习尚未爆发，AI本身是冷门，为AI设计专用芯片更是冷门中的冷门，兄弟俩投向顶级学术会议的论文接连被拒，只能在本职工作之余“坐冷板凳”[s1][s47]。几年后，DianNao、DaDianNao、PuDianNao、ShiDianNao和Cambricon指令集系列论文横扫体系结构学术界，他们成为全球公认的深度学习处理器先驱[s44][s45][s46]。

从论文到产品，从实验室到市值万亿，陈天石走了整整十八年。这十八年里，他创办的寒武纪先因华为麒麟970集成自家IP而一夜成名，又因华为自研达芬奇架构而收入断崖；上市首日市值破千亿，随后股价一路跌到46.59元的冰点；CTO梁军离职对簿公堂、美国实体清单骤然落下、子公司行歌裁员收缩——几乎每一种创业公司可能遭遇的重击，寒武纪都挨过一遍[s9][s26][s14][s41]。但陈天石始终只做一件事：闷头造芯片。他把芯片研发比作推土机，“沿大路而行，不抄小道，不搞奇袭”[s1]。

这是一个关于“冷板凳”如何坐上“热王座”的故事，也是一个关于科学家创业的中国式样本：它没有硅谷式的横空出世，只有在无人区里长年累月的坚持；它的财富爆发看似一夜之间，底下却是两代中国计算机人从龙芯到智能芯片、二十年自主可控的绵长脉络。理解陈天石，就是理解中国AI算力产业如何从一篇论文、一间小黑屋，生长为一条价值万亿的产业链。

On the morning of June 30, 2026, the STAR Market of the Shanghai Stock Exchange produced the first stock in its history to surpass a market capitalization of 1 trillion yuan. Its ticker was 688256; its name was Cambricon. At the moment the share price touched an intraday peak of 1,607 yuan and total market value settled at 1.004 trillion yuan, what came to mind for the countless traders watching their screens was not baijiu or banking, but a chip the size of a fingernail—an artificial-intelligence processor designed, from the ground up, by Chinese engineers.[s34][s35] It is a coordinate heavy with meaning: in an age when compute has become the commanding height of great-power competition and Nvidia's high-end chips are walled off from China by layer upon layer of export controls, China's capital market has voted with real money to lift a chip-design company that once lost money for years and burned through several billion yuan in cumulative deficits to a level where it takes turns with Kweichow Moutai wearing the crown of the A-share market.[s3][s16]

Standing at the center of this capital frenzy is a taciturn 1980s-born scientist who habitually carries a backpack and describes himself as "an ordinary researcher"—Chen Tianshi.[s8] His story begins not in the business world but in a thirty-square-meter " little black room" at the Institute of Computing Technology. Around 2008, he and his brother Chen Yunji, two years his senior, decided to do something the whole world then considered "cold": to build a chip dedicated to artificial intelligence. Deep learning had not yet exploded; AI itself was a backwater, and designing specialized silicon for it was the coldest corner of a cold field. Their submissions to the top academic conferences were rejected one after another, and the brothers could only "sit on the cold bench" in the margins of their day jobs.[s1][s47] A few years later, the DianNao, DaDianNao, PuDianNao and ShiDianNao papers, together with the Cambricon instruction-set series, swept the computer-architecture academic world, and the two men became globally recognized pioneers of the deep-learning processor.[s44][s45][s46]

From papers to products, from a laboratory to a trillion-yuan market value, Chen Tianshi's journey took fully eighteen years. In those years the company he founded first became famous overnight when Huawei integrated its IP into the Kirin 970, then saw its revenue collapse when Huawei went its own way with the Da Vinci architecture; its market value broke 100 billion yuan on its trading debut, only for the share price to slide all the way to an all-time low of 46.59 yuan; its CTO Liang Jun departed and ended up in court, the U.S. Entity List came down without warning, and its Xingge automotive subsidiary slashed jobs and retrenched—almost every heavy blow a startup can suffer, Cambricon absorbed.[s9][s26][s14][s41] Yet Chen kept doing only one thing: quietly building chips. He compares chip research to driving a bulldozer: "Stay on the main road; take no shortcuts, spring no surprises."[s1]

This is the story of how a "cold bench" came to occupy a "hot throne," and it is also a distinctly Chinese specimen of the scientist-turned-founder: there is no Silicon-Valley-style arrival out of nowhere, only years upon years of persistence in a no-man's-land; its explosion of wealth looks overnight, but beneath it lies an unbroken thread running through two generations of Chinese computer scientists—from Loongson to intelligent silicon, twenty years of striving for autonomous and controllable technology. To understand Chen Tianshi is to understand how China's AI-compute industry grew from a single paper and a single little black room into an industrial chain worth a trillion yuan.

## 人物速览 Lead

陈天石，寒武纪创始人、董事长兼首席执行官，中国人工智能芯片产业最具代表性的科学家创业者。1985年6月生于江西南昌，16岁考入中国科学技术大学少年班，2010年获中科大计算机博士学位后进入中科院计算所，与兄长陈云霁一道在全球率先叩开深度学习专用处理器的大门：DianNao系列论文接连拿下ASPLOS、MICRO最佳论文，Cambricon指令集论文创下ISCA审稿评分纪录[s2][s44][s45]。2016年他离岗创办寒武纪，公司以终端IP起家，借华为麒麟970一战成名，又在失去华为、连年亏损、被列入美国实体清单的至暗时刻坚持云端转型[s1][s14]。2020年寒武纪登陆科创板，成为“AI芯片第一股”[s11]。2024年第四季度公司首次单季盈利，2025年全年营收64.97亿元、归母净利润20.59亿元，上市五年首次年度盈利[s16]；2026年6月30日，寒武纪市值盘中突破1万亿元，成为科创板首只万亿市值股票，陈天石也被视作大模型时代最成功的“卖铲人”[s34][s51]。

Chen Tianshi is the founder, chairman and chief executive of Cambricon, and the most representative scientist-entrepreneur of China's artificial-intelligence chip industry. Born in June 1985 in Nanchang, Jiangxi Province, he was admitted at sixteen to the Special Class for the Gifted Young at the University of Science and Technology of China (USTC), took his doctorate in computer science there in 2010, and then joined the Institute of Computing Technology (ICT) of the Chinese Academy of Sciences (CAS). Together with his elder brother Chen Yunji, he was among the first people in the world to push open the door to processors purpose-built for deep learning: the DianNao series of papers carried off best-paper awards at ASPLOS and MICRO in succession, and the Cambricon instruction-set paper set an all-time record for review scores at ISCA.[s2][s44][s45] In 2016 he left his research post to found Cambricon. The company started out licensing intellectual property for terminal devices, shot to fame overnight through Huawei's Kirin 970, and then, in its darkest hours—having lost Huawei as a customer, bled losses for years on end and been placed on the U.S. Entity List—held its course and pushed through a strategic pivot to cloud chips.[s1][s14] In 2020 Cambricon listed on the STAR Market as China's first publicly traded pure-play AI chip stock.[s11] In the fourth quarter of 2024 it posted its first-ever quarterly profit; for full-year 2025 it reported revenue of 6.497 billion yuan and net profit attributable to shareholders of 2.059 billion yuan, its first annual profit in the five years since listing.[s16] On June 30, 2026, Cambricon's market capitalization broke through 1 trillion yuan in intraday trading, making it the first 1-trillion-yuan stock in the history of the STAR Market—and Chen came to be regarded as the most successful "shovel seller" of the large-model era.[s34][s51]

## 正文 Archive Chapters

### 1. 少年班兄弟：一个抓阄选中计算机的神童 / Brothers of the Gifted Young Class: A Prodigy Who Chose Computer Science by Drawing Lots

1985年6月，陈天石出生在江西南昌。在他之前两年，哥哥陈云霁已经是这个家庭里出了名的读书种子。兄弟俩成长于南昌一个普通家庭，父母对教育的看重，让两个儿子先后走上了同一条通往中国科学技术大学少年班的路[s52][s53]。江南都市报后来在报道中回忆，这对从南昌走出的80后兄弟，从小就是邻里口中“别人家的孩子”，但两人的性格与成才路径却截然不同：哥哥陈云霁从小目标明确、一路拔尖；弟弟陈天石则要散漫、随性得多[s53]。

1997年，14岁的陈云霁从南昌十中考入中科大少年班，是当年江西省年龄最小的大学生之一；四年后的2001年，16岁的陈天石从南昌二中考入同一所大学的少年班，编入“01少”班级[s2][s52]。创办于1978年的中科大少年班，是中国超常教育最著名的试验场，在那里，早慧只是入场券，周围人人都是各自家乡的“第一”。中国日报后来在长篇报道中将兄弟二人并称为为人工智能装上中国芯的“双子星”[s4]。

但陈天石的少年班岁月，远没有“神童”二字看上去那么光鲜。他后来多次自嘲，在天才云集的少年班里自己算不上出挑，甚至以“学渣”自况：大学期间一度沉迷电子游戏，成绩平平，对未来也没有什么清晰规划[s52][s53]。河南商报在一篇报道中记述了这个充满反差的细节——一个日后身家数百亿的芯片富豪，少年时主要的烦恼竟是游戏打不通关[s52]。2005年前后，面临读研方向选择的他，竟用抓阄这种近乎玩笑的方式替自己做决定：把几个专业写成纸团，随手一抓，展开是“计算机”三个字。于是他留在中科大继续深造，2010年获计算机博士学位[s1][s52]。多年以后，这段经历被媒体反复提起：一个靠抓阄选中计算机的人，最终为中国造出了最知名的人工智能芯片。

玩笑背后是真实的积淀。少年班素有“先理后工、宽口径培养”的传统，前两年不分专业、强化数理基础，这让陈天石先在2005年拿下数学学士学位，再转入计算机方向攻博，数学训练后来成为他理解神经网络与体系结构的底层功夫[s4][s5]。博士阶段，他师从我国并行计算领域的泰斗、中国科学院院士陈国良与教授姚新，主攻人工智能与神经网络算法[s4][s5]。央广网“天下英才”栏目记述，这位年轻博士对“让机器更聪明”这件事有着近乎本能的兴趣，谈起神经网络算法时眼睛发亮，早早认定智能计算是未来十年最值得投入的方向[s5]。

哥哥陈云霁的路则更“硬”。他在少年班期间就进入中科院计算所参与科研，师从“龙芯之父”胡伟武研究员；2002年正式进入计算所，25岁便出任八核龙芯3号的主架构师，是国产通用处理器事业中最年轻的技术领军人之一[s2][s4]。一个偏算法与智能，一个偏体系结构与芯片，兄弟俩的专业分野在学生时代已经埋下伏笔。陈天石性格内敛、话不多，习惯用论文和产品说话；陈云霁则更为外向果敢，长于在芯片体系结构的世界里排兵布阵[s4][s53]。

长兄如师。陈云霁在计算所参与龙芯研发的那些年，陈天石常去北京探望，兄弟俩的话题总绕不开“中国芯”。哥哥做通用CPU的经历，让陈天石早早看清了一个现实：芯片是信息产业的底座，而这个底座长期握在别人手里[s2][s4]。中国经济网在一篇人物报道中写道，陈天石那代计算机学子心中都憋着一股劲——龙芯证明了中国人能造CPU，那么在即将到来的智能时代，中国人能不能定义一颗属于AI的新芯片？[s6]

多年以后回望，少年班给陈天石的最大财富，或许并非早慧的光环，而是一种“和最聪明的人一起坐冷板凳”的平常心[s47]。在那里，聪明不稀奇，耐得住寂寞才是真功夫；也正是这种平常心，支撑着他日后在无人问津的AI芯片方向上熬过近十年的冷门岁月[s1][s47]。2010年博士毕业，陈天石没有像许多同学那样出国或进入企业，而是选择进入中科院计算技术研究所，与哥哥成为同事[s5]。此时的他还没有想到，几年后兄弟俩会在一间30平方米的“小黑屋”里，做出一件让全球体系结构学术界侧目的大事[s1]。南昌城里那个沉迷游戏、靠抓阄定前程的少年，就此站上了一段长达十余年的冷门赛道的起点[s52]。

回头看，少年班经历给陈氏兄弟的馈赠，远不止一张文凭。那所学院以“强化数理、自由探索”著称，学生前两年不分专业，在数学、物理上打下几乎苛刻的底子，再凭兴趣选择方向——陈天石由数学转计算机、陈云霁由计算数学转芯片体系结构，都得益于此[s4][s5]。更重要的是一种氛围：身边全是各省最聪明的头脑，任何一点小聪明都无处炫耀，唯有长期、专注、坐得住冷板凳的人才能冒头。陈天石后来总结自己的性格底色——“脑子一根筋，耐得住寂寞”，种子正是在少年班的五年里埋下的[s47][s1]。

兄弟俩的分工也在求学期渐次成形。陈云霁2002年进入计算所后，一头扎进龙芯团队，从核级验证做到八核主架构师，把国产CPU从单核推向多核的全过程走了一遍，深知一颗芯片从定义到流片的每一道坎[s2][s4]。陈天石则在算法侧一路深耕，博士期间做神经网络与演化计算，对“智能需要什么样的计算”有第一手的体感[s5][s4]。一个懂芯片怎么造，一个懂智能怎么算——多年以后，正是这两套知识在一间小屋里的碰撞，催生了全球第一个深度学习处理器体系结构[s4][s46]。2010年陈天石走进计算所报到时，中国AI芯片的故事，其实已经悄悄翻开了扉页[s5][s6]。

**English:** In June 1985, Chen Tianshi was born in Nanchang, Jiangxi Province. Two years before him, his elder brother Chen Yunji had already established himself as the family's renowned bookish child. The two brothers grew up in an ordinary Nanchang household in which their parents set great store by education, and both sons were sent, one after the other, down the same road that led to the Special Class for the Gifted Young at the University of Science and Technology of China.[s52][s53] As the Jiangnan Metropolis Daily later recalled in its reporting, the two 1980s-born brothers who walked out of Nanchang were, from childhood, the neighbors' stock example of "other people's children"—yet their temperaments and paths to success were strikingly different: the elder brother Chen Yunji was single-minded from the start and excelled at every turn, while the younger Chen Tianshi was far more relaxed and easygoing.[s53]

In 1997, at fourteen, Chen Yunji entered USTC's Special Class for the Gifted Young from Nanchang No. 10 Middle School, becoming one of the youngest university students in Jiangxi Province that year; four years later, in 2001, sixteen-year-old Chen Tianshi followed him into the same class from Nanchang No. 2 Middle School, assigned to the "'01 Gifted Young" cohort.[s2][s52] Founded in 1978, USTC's Special Class for the Gifted Young is China's most famous experimental ground for the education of the exceptionally gifted; there, precocity is merely the admission ticket, since everyone around you was the number one of their home region. In a long feature, China Daily later paired the two brothers as the "twin stars" who fitted a Chinese chip into artificial intelligence.[s4]

Yet Chen Tianshi's years in the gifted class were far less glittering than the word "prodigy" suggests. He later mocked himself more than once as undistinguished in a class packed with geniuses, even describing himself as an "academic underdog": for a time at university he was addicted to video games, his grades were mediocre, and he had no clear plan for the future.[s52][s53] The Henan Business Daily recorded the detail, rich in contrast, that the greatest worry of a chip billionaire worth tens of billions in later life was, as a teenager, simply that he could not get through a video-game level.[s52] Around 2005, facing the choice of a graduate direction, he made the decision in a manner close to a joke—by drawing lots: he wrote several majors on scraps of paper, grabbed one at random, and unfolded it to find the words "computer science." So he stayed on at USTC for further study and took his PhD in computer science in 2010.[s1][s52] Years later the media returned again and again to that episode: the man who chose computer science by drawing lots ended up building China's best-known artificial-intelligence chip.

Behind the joke lay genuine accumulated substance. The gifted class has long followed a tradition of "science first, engineering later, with broad-based training": for the first two years students declare no major and drill hard into mathematics and physics. That allowed Chen Tianshi first to take a bachelor's degree in mathematics in 2005 before switching to computer science for his doctorate; the mathematical training later became the bedrock of his understanding of neural networks and computer architecture.[s4][s5] In his doctoral years he studied under Chen Guoliang, a doyen of parallel computing in China and an academician of the Chinese Academy of Sciences, and Professor Yao Xin, focusing on artificial intelligence and neural-network algorithms.[s4][s5] As CNR's "Talents of the World" column recounted, the young doctor had an almost instinctive interest in "making machines smarter"; his eyes lit up when he talked about neural-network algorithms, and he had concluded early on that intelligent computing was the direction most worth a decade of commitment.[s5]

The elder brother's road was the harder one. While still in the gifted class, Chen Yunji had begun research at the Institute of Computing Technology under Hu Weiwu, the researcher known as the "father of Loongson"; he formally joined the ICT in 2002 and, at twenty-five, became chief architect of the eight-core Loongson 3, one of the youngest technical leaders in China's general-purpose processor program.[s2][s4] One brother leaned toward algorithms and intelligence, the other toward computer architecture and chips; their professional divergence was foreshadowed already in their student years. Chen Tianshi is reserved, speaks little, and is accustomed to speaking through papers and products; Chen Yunji is more outgoing and decisive, a natural at marshaling forces in the world of chip architecture.[s4][s53]

An elder brother can be a teacher. During the years Chen Yunji worked on Loongson at the ICT, Chen Tianshi often visited him in Beijing, and the brothers' conversations kept circling back to the "Chinese chip." The elder brother's experience building general-purpose CPUs taught Chen Tianshi an early lesson: chips are the foundation of the information industry, and that foundation had long been held in other people's hands.[s2][s4] China Economic Net wrote in a profile that a stubborn determination ran through that generation of Chinese computer students—if Loongson had proved that Chinese people could build a CPU, then in the coming age of intelligence, could they not define a new chip of their own, a chip belonging to AI?[s6]

Looking back years later, the greatest gift the gifted class gave Chen Tianshi may not have been the halo of precocity but an equanimity about "sitting on the cold bench together with the smartest people."[s47] There, cleverness was nothing remarkable; what counted was the capacity to endure loneliness—and it was precisely that equanimity that sustained him through nearly a decade in the cold, neglected corner of AI chips.[s1][s47] When he graduated with his doctorate in 2010, Chen Tianshi did not, like many classmates, go abroad or join a corporation; instead he chose the Institute of Computing Technology of the Chinese Academy of Sciences, becoming his brother's colleague.[s5] He did not yet imagine that, a few years later, the two brothers would, in a thirty-square-meter "little black room," accomplish something that would make the world's computer-architecture community sit up and take notice.[s1] The Nanchang boy who had been addicted to games and settled his future by drawing lots now stood at the starting line of a cold track that would stretch on for more than a decade.[s52]

In retrospect, the gifted-class experience gave the Chen brothers far more than a diploma. The college is known for "intensive mathematical and scientific training and free exploration": students spend their first two years without a declared major, laying an almost merciless foundation in mathematics and physics before choosing a direction by interest—Chen Tianshi's move from mathematics into computer science and Chen Yunji's from computational mathematics into chip architecture both drew on this.[s4][s5] More important still was the atmosphere: surrounded by the sharpest minds of every province, there was no room to show off petty cleverness, and only those who could work long, focus deeply and endure the cold bench ever emerged. Chen Tianshi later summed up the ground color of his character—"single-minded to the point of obstinacy, able to endure loneliness"—and the seed was planted in those five years in the gifted class.[s47][s1]

The brothers' division of labor also gradually took shape during their student years. After entering the ICT in 2002, Chen Yunji plunged straight into the Loongson team, working his way from core-level verification to chief architect of the eight-core design, living through the entire process of pushing China's domestic CPU from single core to many cores and learning every hurdle a chip must cross from definition to tape-out.[s2][s4] Chen Tianshi, meanwhile, dug deep on the algorithm side; during his doctorate he worked on neural networks and evolutionary computation, gaining first-hand feel for "what kind of computing intelligence requires."[s5][s4] One understood how chips are built, the other how intelligence computes—and years later it was precisely the collision of those two bodies of knowledge in a single small room that gave birth to the world's first deep-learning-processor architecture.[s4][s46] When Chen Tianshi walked through the ICT's gates to report for duty in 2010, the story of China's AI chip had, in fact, already quietly turned to its title page.[s5][s6]

### 2. 小黑屋岁月：DianNao与深度学习处理器的开山时刻 / The Years in the Little Black Room: DianNao and the Founding Moment of the Deep-Learning Processor

2008年前后，在中科院计算所，陈云霁和陈天石兄弟凑到一起，聊出了一个大胆的念头：人工智能算法正变得越来越重，而无论CPU还是GPU，跑起神经网络来能效都低得惊人——通用芯片把绝大部分晶体管浪费在了与智能无关的功能上，能不能反其道而行之，为人工智能专门设计一颗处理器？[s1][s6] 这个念头在当时近乎异想天开。深度学习尚未爆发，人工智能本身是冷门，为AI做专用芯片更是冷门中的冷门。陈天石后来回忆：“我们刚开始做AI芯片这个方向的时候，AI芯片其实一点都不热——首先，AI不热；给AI做一个专门的芯片，就更不热。”[s1]

计算所给了这支十来人的团队一间约30平方米的办公室，设备简陋、空间逼仄，大家戏称其为“小黑屋”[s1][s2]。兄弟俩带着团队在这里开始了长达数年的探索：白天各自完成龙芯、算法等本职工作，晚上和周末扑在AI芯片上。投向顶级学术会议的论文接连被拒，评审意见大同小异——没人相信为神经网络专门造一颗芯片在学术和工程上成立[s1][s47]。陈天石把这段日子比作在“AI的石器时代保留火种”：没有人看得见前景，只能自己相信自己[s6]。

转机出现在2014年。这年3月，国际体系结构顶级会议ASPLOS宣布，陈天石作为第一作者、与法国INRIA研究员Olivier Temam及陈云霁等人合作的论文《DianNao：一种小尺寸、高吞吐的深度学习加速器》获得最佳论文奖[s44][s46]。这是ASPLOS历史上中国大陆、也是亚洲机构首次摘得最佳论文。论文提出的DianNao架构，在仅3.02平方毫米、485毫瓦的规模下实现了每秒4520亿次运算的神经网络处理能力，能效远超同期通用处理器[s44][s46]。一个用中文拼音“电脑”命名的芯片架构，就这样登上了世界计算机体系结构研究的最高讲坛。

此后两年，兄弟团队一发不可收。2014年12月，面向大规模神经网络、以多芯片互联扩展算力的DaDianNao论文在MICRO会议上再获最佳论文；2015年，可支持七类主流机器学习算法的PuDianNao登陆ASPLOS，贴近传感器端、主打超低功耗的ShiDianNao发表于ISCA[s44][s46]。2016年，团队在ISCA发表《Cambricon：一种面向神经网络的指令集架构》，第一作者为刘少礼；这篇提出深度学习专用指令集DianNaoYu（寒武纪指令集）的论文，创下了ISCA审稿评分的历史纪录，在十篇高分论文中排名第一[s45][s46]。DianNao、DaDianNao、PuDianNao、ShiDianNao加上Cambricon指令集，构成了全球首个成体系的深度学习处理器学术路线，被业界称为“寒武纪谱系”[s44][s46]。

论文之外，团队更早迈出了工程化的一步。2015年，全球首款深度学习专用处理器原型芯片在他们手中完成流片，实测能效达到传统芯片的近百倍，第一次在硅片上验证了学术架构的工程可行性[s2][s6]。2018年2月，《Science》杂志在相关报道中评价，中国科学家在人工智能芯片这一新兴领域作出了“开创性贡献”[s6]。从一间小黑屋里的被拒论文，到被全世界同行反复引用的指令集，陈氏兄弟用七八年时间证明：在一个全新的计算范式里，中国人可以从最底层开始做原创，而不只是在别人定义的赛道上追赶[s2][s46]。

学术巅峰之时，一个现实的分岔也摆在面前：论文里的芯片再好，不变成货架上的产品，就永远是纸面文章。最终兄弟俩做了分工——陈云霁留在计算所，继续从事基础研究，后来担任计算所智能处理器研究方向的首席科学家；陈天石则脱下“科研编制”，带着技术成果离岗创业[s2][s4]。多年后陈天石说，哥哥守在学术源头、自己冲向产业战场，这种安排让寒武纪既能扎根前沿、保持技术敏感，又能放手商业化、不必在体制内束手束脚[s4]。

小黑屋的岁月也沉淀了寒武纪最早的人才班底：刘少礼、郭崎等一批论文合作者后来陆续加入公司，构成了研发体系的核心[s46][s44]。这间30平方米的小屋，后来被媒体反复书写为中国AI芯片产业的“摇篮”——五亿年前寒武纪生命大爆发的隐喻，正是从这里开始孕育[s6][s3]。火种既已点燃，陈天石要做的，就是把它带出计算所的院墙，等待一场真正的产业“大爆发”[s6]。

学术圈给了这支团队最高规格的承认。DianNao论文获ASPLOS最佳论文，是中国大陆机构三十余年来首次在该会议摘得这一荣誉；DaDianNao再夺MICRO最佳论文，使兄弟团队成为史上少有的在同一年内拿下体系结构两大顶级会议最佳论文的研究组[s44][s46]。腾讯云开发者社区在技术解读中指出，DianNaoYu/Cambricon指令集首次以load-store形式定义了矩阵、向量、池化等神经网络专用指令，软硬件接口的设计思想被此后几乎所有AI芯片所借鉴[s46]。ACM数字图书馆收录的ISCA 2016论文页面显示，Cambricon指令集论文由刘少礼、陈天石等人共同完成，至今仍是神经网络指令集方向引用量最高的文献之一[s45]。

陈云霁后来留在计算所，先后担任智能处理器研究中心主任、研究所副所长级别的学术领军者，并获全国五一劳动奖章等荣誉；他的身份始终是科学家，不直接在寒武纪持股，却在技术源头上与公司保持着千丝万缕的联系[s2][s4]。陈天石则带着论文、专利和一支十余人的核心队伍走向市场。2016年公司成立时，他在内部说过一句话：我们是在AI的石器时代出发，手里没有别的，只有一颗火种和一张别人看不懂的地图[s6]。日后所有的高光与低谷，都能在这间小黑屋里找到源头——包括那台被戏称为“小黑屋”的办公室里，贴在墙上的一张芯片架构草图[s1][s2]。

**English:** Around 2008, at the Institute of Computing Technology, the brothers Chen Yunji and Chen Tianshi found themselves in the same conversations and talked their way to a bold idea: artificial-intelligence algorithms were growing ever heavier, yet whether run on CPUs or GPUs, neural networks achieved astonishingly poor energy efficiency—general-purpose chips wasted the vast majority of their transistors on functions unrelated to intelligence. Could they do the opposite, and design a processor dedicated to artificial intelligence?[s1][s6] At the time the notion bordered on fantasy. Deep learning had not yet exploded; artificial intelligence itself was a backwater, and building a specialized chip for AI was the coldest corner of a cold field. Chen Tianshi later recalled: "When we first started working on AI chips, the field wasn't hot at all—first, AI itself wasn't hot; and making a dedicated chip for AI was even colder."[s1]

The ICT gave the team of a dozen or so people an office of about thirty square meters—cramped, with primitive equipment—which they jokingly called the "little black room."[s1][s2] There the brothers led the team through years of exploration: by day each completed their regular work on Loongson or algorithms; by night and on weekends they threw themselves into the AI chip. Papers submitted to the top academic conferences were rejected one after another, the reviews striking much the same note—no one believed that building a chip dedicated to neural networks was sound, academically or engineering-wise.[s1][s47] Chen Tianshi compared those days to "keeping the flame alive in the Stone Age of AI": no one could see the prospect ahead, so they could only believe in themselves.[s6]

The turning point came in 2014. In March that year, ASPLOS, a top international conference on computer architecture, announced that the paper "DianNao: A Small-Foot-print, High-Throughput Accelerator for Ubiquitous Machine-Learning and Deep Learning," with Chen Tianshi as first author in collaboration with Olivier Temam, a researcher at France's INRIA, and Chen Yunji among others, had won the best-paper award.[s44][s46] It was the first time in ASPLOS's history that an institution from mainland China—and indeed from Asia—had taken the best-paper honor. The DianNao architecture proposed in the paper delivered neural-network processing of 452 billion operations per second within a footprint of just 3.02 square millimeters and a power budget of 485 milliwatts, with energy efficiency far beyond that of contemporary general-purpose processors.[s44][s46] A chip architecture named in the Chinese pinyin "dian nao"—the words for "computer"—had thus climbed onto the highest podium of world computer-architecture research.

Over the next two years the brothers' team was unstoppable. In December 2014, the DaDianNao paper, aimed at large-scale neural networks and scaling compute through multi-chip interconnection, won a further best-paper award at MICRO; in 2015, PuDianNao, which supported seven mainstream categories of machine-learning algorithms, landed at ASPLOS, while ShiDianNao, designed to sit close to the sensor and pursue ultra-low power, was published at ISCA.[s44][s46] In 2016 the team published at ISCA the paper "Cambricon: An Instruction Set Architecture for Neural Networks," with Liu Shaoli as first author; that paper, which put forward DianNaoYu, the dedicated instruction set for deep learning (the Cambricon instruction set), set an all-time record for review scores at ISCA, ranking first among the ten highest-scored papers.[s45][s46] DianNao, DaDianNao, PuDianNao and ShiDianNao, together with the Cambricon instruction set, constituted the world's first systematic academic roadmap for deep-learning processors—the lineage the industry came to call the "Cambricon family."[s44][s46]

Beyond the papers, the team had taken an earlier step toward engineering. In 2015, the world's first prototype chip for a dedicated deep-learning processor completed tape-out in their hands; measured energy efficiency reached nearly a hundred times that of traditional chips, validating for the first time in silicon the engineering feasibility of the academic architecture.[s2][s6] In February 2018, the journal Science, in related coverage, judged that Chinese scientists had made a "pioneering contribution" in the emerging field of AI chips.[s6] From rejected papers in a little black room to an instruction set cited again and again by peers across the world, the Chen brothers spent seven or eight years proving that in a brand-new computing paradigm, Chinese researchers could do original work from the bottom layer up, rather than merely chasing along a track defined by others.[s2][s46]

At the summit of academic success, a practical fork in the road also presented itself: however good the chips in the papers were, unless they became products on the shelf, they would remain forever words on paper. In the end the brothers divided the labor—Chen Yunji stayed at the ICT to continue basic research, later becoming chief scientist for the institute's intelligent-processor direction; Chen Tianshi shed his research-establishment status and left his post to start a company, taking the technical achievements with him.[s2][s4] Years later Chen Tianshi said that the arrangement—his brother guarding the academic source while he charged into the industrial battlefield—let Cambricon stay rooted in the frontier and keep its technical sensitivity, while freeing it to commercialize without being trammeled inside the system.[s4]

The little-black-room years also settled Cambricon's earliest core of talent: a group of paper collaborators including Liu Shaoli and Guo Qi later joined the company one after another and formed the backbone of its R&D system.[s46][s44] That thirty-square-meter room was later written up again and again in the press as the "cradle" of China's AI-chip industry—the metaphor of the Cambrian life explosion 500 million years ago began to gestate precisely here.[s6][s3] With the flame lit, what Chen Tianshi had to do was carry it beyond the walls of the ICT and wait for a genuine industrial "explosion."[s6]

The academic world granted this team its highest-grade recognition. The DianNao paper's best-paper award at ASPLOS was the first time in more than thirty years that an institution from mainland China had taken the honor at the conference; DaDianNao's subsequent best-paper award at MICRO made the brothers' team one of the very few research groups in history to win best-paper honors at the two top architecture conferences within the same year.[s44][s46] As the Tencent Cloud Developer Community pointed out in its technical analysis, the DianNaoYu/Cambricon instruction set was the first to define neural-network-specific instructions for matrix, vector and pooling operations in load-store form, and the design thinking behind its hardware-software interface was borrowed by nearly every AI chip that followed.[s46] The ISCA 2016 paper page in the ACM Digital Library shows that the Cambricon instruction-set paper, completed jointly by Liu Shaoli, Chen Tianshi and others, remains to this day one of the most-cited works in the neural-network instruction-set direction.[s45]

Chen Yunji later stayed at the ICT, successively serving as director of its Intelligent Processor Research Center and as an academic leader at deputy-director-general level of the institute, and receiving honors including the National May First Labor Medal; his identity has always been that of a scientist—he holds no direct equity in Cambricon, yet maintains a thousand threads of technical connection with the company at the source.[s2][s4] Chen Tianshi, meanwhile, carried the papers, the patents and a core team of a dozen-plus people toward the market. When the company was founded in 2016, he said internally: we set out in the Stone Age of AI, with nothing in hand but a single flame and a map no one else could read.[s6] Every highlight and every trough that followed can be traced back to that little black room—right down to the sketched chip-architecture diagram pinned to the wall of the office nicknamed for its darkness.[s1][s2]

### 3. 寒武纪开眼：从麒麟970的高光到达芬奇分手 / Cambricon Opens Its Eyes: From the Kirin 970 Highlight to the Da Vinci Breakup

2016年3月，北京中科寒武纪科技有限公司注册成立，初始注册资本仅90万元，其中陈天石出资63万元，中科院计算所旗下中科算源出资27万元[s1][s43]。公司名取自五亿年前的寒武纪生命大爆发——陈氏兄弟相信，人工智能也将迎来一场物种式的爆发，而他们要做这场爆发最底层的物质载体[s3]。同年，寒武纪推出寒武纪1A智能处理器IP，这是全球首款商用的终端智能处理器IP，主打在手机等终端设备上高效运行深度学习任务，让AI计算从云端下沉到每个人的口袋[s5][s10]。

资本很快闻风而来。成立当年，寒武纪获得中科院旗下基金数千万元天使投资；2016年8月完成Pre-A轮融资，元禾原点、科大讯飞、涌铧投资等入局；2017年8月18日，公司宣布完成1亿美元A轮融资，由国投创业领投，阿里巴巴、联想创投、国科投资、中科图灵等跟投，估值一跃达到10亿美元，跻身全球AI芯片领域少有的“独角兽”[s12][s13]。2018年6月，寒武纪再获数亿美元B轮融资，估值升至25亿美元，投资方包括中国国有资本风险投资基金等国家级基金[s13]。一家成立两年的芯片公司，成了中国AI赛道最炙手可热的标的，格隆汇在上市前的盘点中称其为“科创板最受期待的AI选手”[s10]。

真正让寒武纪一夜成名的，是华为。2017年9月2日，德国柏林IFA消费电子展上，华为发布全球首款内置独立神经网络处理单元（NPU）的手机芯片麒麟970——这颗采用台积电10纳米工艺、集成55亿个晶体管的SoC，其AI算力核心正是寒武纪1A[s9][s10]。华为在发布会上宣称，集成NPU后，麒麟970的AI运算密度相较同期方案有数量级提升，图像识别、语音翻译等端侧智能体验显著跃迁[s9]。搭载麒麟970的Mate 10等十余款机型陆续上市，寒武纪作为幕后IP供应商声名鹊起；此后麒麟980采用双核的寒武纪1H，支撑Mate 20系列等机型[s9][s43]。据寒武纪招股书披露，2017年、2018年公司终端智能处理器IP授权业务收入占比分别高达98.33%和99.69%——几乎可以说，那两年的寒武纪是“靠华为一家养活的”[s38][s43]。

这桩合作的牵线人之一，是2017年加入寒武纪出任CTO的梁军。梁军2000年加入华为，曾任海思半导体麒麟SoC芯片总架构师，既懂华为的体系，又懂寒武纪的技术，是把寒武纪IP引入麒麟970的关键人物[s25][s28]。与全球顶级手机厂商的联合打磨，也让寒武纪第一次见识到消费电子级芯片的工程标准与交付节奏[s9]。

高光之下，裂痕早已埋下。对华为这样年出货数亿台设备的终端巨头而言，核心AI引擎长期握在一家外部初创公司手里，本就不是长久之计；而对寒武纪来说，单一大客户贡献近乎全部收入，也是一把悬顶之剑[s33][s37]。2018年10月，华为在全联接大会上公布自研的达芬奇AI芯片架构；2019年，搭载自研NPU的麒麟810问世，华为终端逐步切换到自研路线，寒武纪的IP授权订单随之锐减[s9][s33]。2019年，寒武纪终端IP业务收入占比从上一年的99.69%骤降至15.49%，公司在招股书中坦承了大客户流失的风险[s38]。

这场“分手”在当时被不少人视为寒武纪的生死劫，陈天石后来却看得很平淡：华为这样的公司走向自研是产业必然，与其把命运绑定在一个客户身上，不如及早开辟自己的主战场[s33][s7]。事实上，早在2018年，寒武纪就已启动向云端智能芯片的战略转型——终端IP做名气，云端芯片打基业[s10][s22]。回头看，与华为的短暂蜜月给了寒武纪三样东西：一笔可观的早期收入、一个被全球顶级手机厂商验证过的技术品牌，以及一记让全公司清醒过来的警钟[s43]。麒麟970里那颗指甲盖大小的NPU，是寒武纪的成人礼；而达芬奇架构的转身，则逼着这家年轻公司在三年内完成从“IP供应商”到“全栈芯片厂商”的惊险一跃[s9][s33]。

融资节奏几乎踩在了每一个市场热点上。智东西在B轮融资的报道中梳理，寒武纪成立两年间完成四轮融资，背后既有国投创业、国有资本风投基金这样的国家级长线资本，也有阿里巴巴、联想这样的产业方，还有科大讯飞这类AI应用公司——资本结构里已经写满了“国产算力自主可控”的时代预期[s13]。中国证券报在申购报道中计算，若按发行价中一签500股，上市首日浮盈即可超过6万元，市场情绪之热可见一斑[s12]。

但隐患也藏在这份热闹里。寒武纪招股书显示，2017年、2018年公司前五大客户集中度极高，其中华为海思贡献的IP授权收入占比分别达到98.33%、99.69%；这意味着公司本质上还是一家“单点支撑”的技术供应商，离一家独立成熟的芯片公司距离尚远[s38][s43]。格隆汇在上市前的分析文章中也提醒，市场给寒武纪的估值里，既有对“AI芯片第一股”稀缺性的溢价，也有对客户集中、持续亏损、商业化不确定的担忧[s10]。陈天石对此并非没有预判——华为自研的传闻在2018年下半年已在业内流传，公司一方面继续交付1H IP、站好最后一班岗，另一方面把募投资源大规模倾斜到云端芯片研发上[s33][s7]。事后看，这是寒武纪成立以来最重要的一次战略转向：没有被“华为供应商”的舒适区留住，才有了后来思元系列的故事[s22][s43]。

**English:** In March 2016, Beijing Zhongke Cambricon Technology Co., Ltd. was registered with an initial registered capital of just 900,000 yuan, of which Chen Tianshi contributed 630,000 yuan and Zhongke Suanyuan, an investment vehicle under the ICT, contributed 270,000 yuan.[s1][s43] The company's name was taken from the Cambrian life explosion 500 million years ago—the Chen brothers believed that artificial intelligence would likewise witness an explosion of species, and that they wanted to be the most fundamental layer of material substrate for that explosion.[s3] In the same year, Cambricon launched the Cambricon 1A intelligent-processor IP, the world's first commercial terminal intelligent-processor IP, designed to run deep-learning tasks efficiently on terminal devices such as smartphones and push AI compute down from the cloud into everyone's pocket.[s5][s10]

Capital soon caught the scent. In its founding year, Cambricon received tens of millions of yuan in angel investment from a fund under the Chinese Academy of Sciences; in August 2016 it completed a Pre-A round with investors including Yuanhe Origin, iFlytek and Yonghua Investment; and on August 18, 2017, the company announced a US$100 million Series A led by SDIC Venture Capital, with Alibaba, Legend Capital, CAS Investment and Zhongke Turing among the followers, lifting its valuation at a stroke to US$1 billion and placing it among the rare "unicorns" in the global AI-chip field.[s12][s13] In June 2018 Cambricon raised a further several-hundred-million-dollar Series B, at a valuation of US$2.5 billion, with investors including national-level funds such as the China State-Owned Capital Venture Investment Fund.[s13] A chip company just two years old had become the hottest target on China's AI track; in a pre-IPO round-up, Gelonghui called it "the most anticipated AI contender on the STAR Market."[s10]

What made Cambricon famous overnight was Huawei. On September 2, 2017, at the IFA consumer-electronics show in Berlin, Huawei unveiled the Kirin 970, the world's first smartphone chip with a built-in, standalone Neural Processing Unit (NPU)—and the AI-compute core of that SoC, built on TSMC's 10-nanometer process with 5.5 billion transistors, was precisely the Cambricon 1A.[s9][s10] Huawei declared at the launch that, with the NPU integrated, the Kirin 970's AI-compute density was an order of magnitude higher than contemporary alternatives, with a marked leap in on-device intelligent experiences such as image recognition and voice translation.[s9] More than ten handsets powered by the Kirin 970, including the Mate 10, came to market one after another, and Cambricon's reputation soared as the behind-the-scenes IP supplier; the later Kirin 980 used the dual-core Cambricon 1H, powering the Mate 20 series and other models.[s9][s43] As Cambricon's IPO prospectus disclosed, IP licensing for terminal intelligent processors accounted for as much as 98.33% and 99.69% of the company's revenue in 2017 and 2018 respectively—one could almost say that in those two years Cambricon was "fed by Huawei alone."[s38][s43]

One of the matchmakers for that partnership was Liang Jun, who joined Cambricon as CTO in 2017. Liang had joined Huawei in 2000 and had served as chief architect of HiSilicon's Kirin SoC chips; he understood both Huawei's system and Cambricon's technology, and was the key figure who brought Cambricon's IP into the Kirin 970.[s25][s28] The joint polishing with the world's top handset maker also gave Cambricon its first taste of the engineering standards and delivery cadence of consumer-electronics-grade chips.[s9]

Beneath the glare, the cracks had long been laid. For a terminal giant like Huawei, shipping hundreds of millions of devices a year, leaving its core AI engine permanently in the hands of an external startup was never a durable arrangement; and for Cambricon, a single large customer contributing nearly all its revenue was likewise a sword hanging overhead.[s33][s37] In October 2018, Huawei unveiled its in-house Da Vinci AI-chip architecture at its Full Connection conference; in 2019 the Kirin 810, powered by a self-developed NPU, appeared, Huawei's terminal business gradually shifted to its own route, and Cambricon's IP licensing orders shrivelled accordingly.[s9][s33] In 2019, terminal-IP revenue's share of Cambricon's business plunged from the previous year's 99.69% to 15.49%, and the company acknowledged the risk of major-customer loss in its prospectus.[s38]

Many at the time saw the "breakup" as a life-or-death crisis for Cambricon, but Chen Tianshi later took a calm view: it was an industrial inevitability that a company like Huawei would move to in-house development, and rather than tie its fate to a single customer, Cambricon was better off opening its own main battlefield early.[s33][s7] In fact, as early as 2018 Cambricon had already launched its strategic pivot to cloud intelligent chips—terminal IP to make the name, cloud chips to build the foundation.[s10][s22] In hindsight, the brief honeymoon with Huawei gave Cambricon three things: a substantial early income, a technology brand validated by the world's top handset maker, and a wake-up call that sobered the whole company.[s43] The fingernail-sized NPU inside the Kirin 970 was Cambricon's coming-of-age ceremony; the turn toward the Da Vinci architecture, in turn, forced the young company to complete, within three years, the perilous leap from "IP supplier" to "full-stack chip vendor."[s9][s33]

The financing cadence landed almost on every market hotspot. As Zhidx sorted out in its reporting on the Series B, Cambricon completed four rounds of financing within two years of founding; behind it stood national-level long-term capital such as SDIC Venture Capital and the state-owned-capital venture fund, industrial players such as Alibaba and Lenovo, and AI-application companies such as iFlytek—the cap table already spelled out the era's expectation of "autonomous and controllable domestic compute."[s13] The China Securities Journal calculated in its subscription coverage that at the IPO price, a single winning lot of 500 shares would carry a floating profit of more than 60,000 yuan on the first day of trading—some measure of how fevered market sentiment had become.[s12]

But the dangers were hidden in the excitement. Cambricon's prospectus showed that in 2017 and 2018 customer concentration among the top five was extremely high, with IP-licensing revenue from Huawei HiSilicon alone reaching 98.33% and 99.69% of the total; this meant the company was in essence still a technology supplier propped up by a single point, far from being an independent, mature chip company.[s38][s43] Gelonghui's pre-IPO analysis likewise warned that the valuation the market gave Cambricon contained both a premium for the scarcity of the "first AI-chip stock" and worry over customer concentration, persistent losses and commercial uncertainty.[s10] Chen Tianshi was hardly without foresight—rumors of Huawei's in-house plans had circulated in the industry in the second half of 2018, and the company on the one hand continued delivering the 1H IP, standing its last watch, and on the other tilted the resources raised from investors heavily toward cloud-chip R&D.[s33][s7] In hindsight this was the most important strategic turn since Cambricon's founding: by not lingering in the comfort zone of "Huawei supplier," the company made possible the later story of the MLU series.[s22][s43]

### 4. 云端转型：思元谱系与云边端车的全栈布局 / The Pivot to the Cloud: The MLU Lineup and the Full-Stack Cloud-Edge-Device-Car Layout

失去华为这个大客户之后，寒武纪把全部身家押向了云端AI芯片。这是一条更难、也更宽的路：云端训练与推理芯片是英伟达的腹地，技术门槛极高、生态壁垒极厚，但市场空间远非手机IP可比。陈天石的判断是，人工智能的算力消耗终将集中在数据中心，谁能在云端站住脚，谁才拥有真正的未来[s33][s7]。他在2020年的专访中坦言，转型云端意味着同时挑战硬件、软件和生态三座大山，但“赛道足够长，值得用十年去赌”[s7]。

2018年5月，寒武纪发布思元100（MLU100）云端智能芯片，这是中国第一款高峰值性能的云端AI芯片，可同时支持云端推理与训练，标志着公司正式从终端IP供应商转向云端芯片厂商[s22][s43]。2019年6月20日，第二代云端芯片思元270发布，采用台积电16纳米工艺，INT8峰值算力达128TOPS，搭载公司自主研发的MLUv02指令集，并支持INT4等低精度量化，在互联网与安防的云端推理场景中开始规模落地[s22]。同年，面向边缘计算场景的思元220面世，凭借高性价比在智能安防、互联网边缘节点等场景放量，此后累计销量突破百万片，成为寒武纪出货量最大的“走量”产品[s30][s42]。

进入2021年，寒武纪的芯片节奏明显加快。1月21日，公司在上交所发布公告，推出思元290智能芯片及玄思1000智能加速器：思元290采用台积电7纳米工艺，集成460亿个晶体管，INT4算力达1024TOPS、INT8算力512TOPS，是当时国产AI芯片的巅峰之作，直接对标语境中的英伟达V100、A100与华为昇腾910；玄思1000在2U空间内集成四颗思元290，整体算力达4.1 PetaOPS，瞄准大规模训练集群[s23][s30]。11月3日，第三代云端芯片思元370发布，这是国内首颗采用chiplet（芯粒）先进封装技术的AI芯片，7纳米工艺、集成390亿个晶体管，INT8算力256TOPS，并在国内首次支持LPDDR5高速内存，兼顾训练与推理负载[s24][s30]。

2022年世界人工智能大会（WAIC）上，寒武纪首次披露新一代思元590的架构信息（MLUarch05）；这款7纳米芯片于2023年底正式上市，综合性能约为英伟达A100的八成，更关键的是，它在设计之初就面向大模型训练与推理，随后完成了对DeepSeek、通义千问（Qwen）、智谱GLM、月之暗面Kimi等国内主流大模型的适配[s42][s51]。在英伟达高端GPU对华出口持续受限的背景下，思元590成为国产大模型算力清单中最重要的选项之一[s32][s51]。华鑫证券在研报中评价，从100到590的五代演进，让寒武纪在云端训练、云端推理、边缘计算三条线上均具备了国内领先的产品力[s30]。

云端之外，寒武纪还把战线拉到了车内。2021年1月，公司全资设立行歌科技，专攻自动驾驶芯片，股东名单里出现了上汽集团、蔚来、宁德时代、博世等产业巨头的名字[s41]。行歌先后推出面向L2+辅助驾驶的SD5223（16TOPS算力）和面向L4自动驾驶的SD5226（7纳米工艺、400TOPS以上算力），画出“云、边、端、车”一体的全栈版图[s30][s41]。软件层面，寒武纪自研NeuWare基础系统软件栈与BANG编程语言，试图对标英伟达CUDA，尽可能降低开发者的迁移与适配成本[s33][s49]。2022年WAIC上，公司集中展示了覆盖云端、边缘、终端、车载的全算力产品矩阵，这在国产芯片厂商中独此一家[s42]。

从思元100到思元590，六年间五代云端芯片，外加边缘的220和车载的行歌系列，寒武纪成为国内产品线最完整的本土AI芯片公司[s30][s42]。陈天石在业绩说明会上反复强调，公司的策略是“云边端车”协同、以云端训练芯片牵引全栈能力[s49][s50]。这条路线烧钱惊人、回报迟缓，每一代芯片都要经历数亿元级的流片投入，却在大模型时代到来时显出了远见——当整个国家都在寻找英伟达之外的算力，那条在无人问津时就铺好的产品线，成了寒武纪最硬的底气[s33][s42]。

云端转型的每一步都代价不菲。思元100上市时，国内云端训练市场几乎被英伟达CUDA生态垄断，国产芯片“流得出、卖不动、用不顺”是普遍困境；寒武纪选择把硬件与软件栈同步推进，每一代芯片都配套迭代NeuWare，兼容主流深度学习框架，用“好用”一点点撬动客户[s22][s33]。思元290发布时，公司在公告中罕见地披露了460亿晶体管、1024TOPS等硬核指标，意在向市场证明：在7纳米节点上，中国团队有能力做出与国际一线对标的训练芯片[s23]。思元370引入chiplet芯粒技术，则是在先进制程受限预期下提前布局——用成熟封装拼接多颗芯粒，在不依赖最尖端单die工艺的前提下继续提升算力密度[s24][s30]。

行歌的车载尝试同样体现了“全栈”野心。2021年成立时，行歌被定位为寒武纪智能化版图的“第四块拼图”，上汽、蔚来、宁德时代、博世等股东几乎覆盖了整车制造、动力电池和Tier1全链条[s41]。SD5223与SD5226两款产品分别瞄准辅助驾驶量产市场和高阶自动驾驶，路线图一度排到5纳米节点[s30]。尽管2023年行歌因集团收缩战略而裁员近半，但车载产品线的技术积累与客户关系并未清零，被视为寒武纪在云端主业造血之后仍可重新激活的期权[s41]。甲子光年评价，寒武纪的“云边端车”布局是国产AI芯片公司中少见的完整形态，成则全栈协同，败则处处掣肘——考验的是掌门人分配资源的定力[s41]。

**English:** After losing the big customer that was Huawei, Cambricon staked everything on cloud AI chips. It was a harder road, but also a wider one: cloud training and inference chips are Nvidia's home ground, with extraordinarily high technical barriers and formidably thick ecosystem moats, yet the market space is beyond comparison with smartphone IP. Chen Tianshi's judgment was that the compute consumption of artificial intelligence would ultimately concentrate in data centers, and that whoever held a foothold in the cloud would own the real future.[s33][s7] In a 2020 interview he admitted frankly that pivoting to the cloud meant taking on three mountains at once—hardware, software and ecosystem—but that "the track is long enough to be worth a ten-year bet."[s7]

In May 2018, Cambricon released the MLU100 (Siyuan 100) cloud intelligent chip, China's first cloud AI chip with high peak performance; it could support both cloud inference and training, and marked the company's formal shift from terminal IP supplier to cloud-chip vendor.[s22][s43] On June 20, 2019, the second-generation cloud chip, the MLU270, was released; built on TSMC's 16-nanometer process, it delivered peak INT8 compute of 128 TOPS, carried the company's self-developed MLUv02 instruction set, supported low-precision quantization such as INT4, and began to be deployed at scale in cloud-inference scenarios across internet services and security surveillance.[s22] The same year saw the MLU220 for edge-computing scenarios; with its strong price-performance ratio it shipped in volume across intelligent surveillance and internet edge nodes, later surpassing a million units in cumulative sales to become Cambricon's highest-volume, "bread-and-butter" product.[s30][s42]

Entering 2021, Cambricon's chip cadence quickened markedly. On January 21, the company filed an announcement with the Shanghai Stock Exchange launching the MLU290 intelligent chip and the XuanSi 1000 intelligent accelerator: the MLU290 was built on TSMC's 7-nanometer process, integrated 46 billion transistors, and delivered 1,024 TOPS of INT4 compute and 512 TOPS of INT8—at the time the pinnacle of domestic AI chips, benchmarked directly against the Nvidia V100 and A100 and Huawei's Ascend 910; the XuanSi 1000 packed four MLU290 chips into a 2U form factor for overall compute of 4.1 PetaOPS, aimed at large-scale training clusters.[s23][s30] On November 3, the third-generation cloud chip, the MLU370, was released: China's first AI chip to use advanced chiplet packaging, on a 7-nanometer process with 39 billion transistors, 256 TOPS of INT8 compute, and the country's first support for LPDDR5 high-speed memory, balancing training and inference workloads.[s24][s30]

At the 2022 World Artificial Intelligence Conference (WAIC), Cambricon disclosed for the first time architectural details of the new-generation MLU590 (MLUarch05); this 7-nanometer chip formally went on sale at the end of 2023, with overall performance of roughly 80% of the Nvidia A100. More crucially, it was designed from the outset for large-model training and inference, and was subsequently adapted for China's mainstream large models, including DeepSeek, Qwen (Tongyi Qianwen), Zhipu GLM and Moonshot Kimi.[s42][s51] Against the backdrop of continuing restrictions on high-end Nvidia GPUs for China, the MLU590 became one of the most important options on the domestic large-model compute list.[s32][s51] China Fortune Securities, in its research reports, judged that the five-generation evolution from the 100 to the 590 gave Cambricon domestically leading product strength on all three lines of cloud training, cloud inference and edge computing.[s30]

Beyond the cloud, Cambricon also extended the battle line into the car. In January 2021, the company set up the wholly owned subsidiary Xingge Technology, focused on autonomous-driving chips, with the names of industrial giants including SAIC Motor, NIO, CATL and Bosch appearing on its shareholder list.[s41] Xingge successively launched the SD5223 for L2+ assisted driving (16 TOPS) and the SD5226 for L4 autonomous driving (7-nanometer process, more than 400 TOPS), drawing out a full-stack map integrating "cloud, edge, device and car."[s30][s41] On the software side, Cambricon built the in-house NeuWare base-system software stack and the BANG programming language, seeking to benchmark Nvidia's CUDA and lower developers' migration and adaptation costs as far as possible.[s33][s49] At WAIC 2022 the company showcased in one place a full-compute product matrix covering cloud, edge, terminal and automotive—something no other domestic chip vendor could match.[s42]

From the MLU100 to the MLU590, five generations of cloud chips in six years, plus the edge 220 and the automotive Xingge series, made Cambricon the domestic AI-chip company with the most complete product line in China.[s30][s42] At earnings briefings Chen Tianshi repeatedly stressed that the company's strategy was coordinated development across "cloud, edge, device and car," with cloud training chips pulling the full-stack capability along behind them.[s49][s50] The route burned astonishing amounts of money and repaid slowly; every chip generation required tape-out investment in the hundreds of millions of yuan—yet when the large-model age arrived it proved foresighted: when an entire country went looking for compute beyond Nvidia, the product line that had been laid down while no one was paying attention turned out to be Cambricon's hardest source of confidence.[s33][s42]

Every step of the cloud pivot came at a high price. When the MLU100 launched, China's cloud-training market was almost entirely monopolized by the Nvidia CUDA ecosystem, and domestic chips faced the common predicament of "taping out fine, but not selling and not running smoothly"; Cambricon chose to push hardware and software stack in lockstep, iterating NeuWare alongside each chip generation, staying compatible with mainstream deep-learning frameworks, and prying customers open little by little with "ease of use."[s22][s33] When the MLU290 launched, the company did something rare in its announcements, disclosing hard-core metrics such as 46 billion transistors and 1,024 TOPS, to prove to the market that at the 7-nanometer node a Chinese team could build a training chip benchmarked against the global first tier.[s23] The MLU370's adoption of chiplet technology, in turn, was an early layout in anticipation of constraints on leading-edge processes—stitching multiple chiplets together with mature packaging to keep raising compute density without depending on the most cutting-edge single-die process.[s24][s30]

Xingge's automotive attempt likewise embodied the "full-stack" ambition. When founded in 2021, Xingge was positioned as the "fourth piece of the puzzle" in Cambricon's intelligent landscape, with shareholders such as SAIC, NIO, CATL and Bosch covering nearly the whole chain of vehicle manufacturing, power batteries and Tier-1 suppliers.[s41] The SD5223 and SD5226 targeted respectively the mass-production assisted-driving market and high-level autonomous driving, with the roadmap at one point extending to the 5-nanometer node.[s30] Although Xingge cut nearly half its workforce in 2023 as the group retrenched, the technical accumulation and customer relationships of the automotive line were not wiped out; they are seen as an option Cambricon can reactivate once its cloud main business generates cash.[s41] Jazzy Gravity's assessment was that Cambricon's cloud-edge-device-car layout is a complete form rarely seen among domestic AI-chip companies—success means full-stack synergy, failure means constraints everywhere—and the test it sets is the leader's steadiness in allocating resources.[s41]

### 5. 科创板第一股：千亿开场与46.59元的冰点 / The First Stock of the STAR Market: A 100-Billion Opening and the 46.59-Yuan Ice Point

2020年，科创板开市刚满一年，市场就在等待一个标志性的标的。寒武纪从受理到上会仅用了约117天，刷新了科创板的审核速度，被媒体称为“闪电过会”[s10][s43]。7月20日，寒武纪（688256）正式登陆科创板，发行价64.39元，发行新股募资25.82亿元；开盘即报250元，较发行价大涨288%，最终收于212.4元，涨幅229.86%，首日市值一举突破千亿元[s11][s12]。7月23日，股价盘中冲上297.77元的上市初期高位[s38]。申购阶段，市场就流传“中一签或赚10万”的说法，打新热度空前[s12]。“AI芯片第一股”的名号，从此与寒武纪牢牢绑定[s11]。

然而资本市场的高光，掩盖不了财报的尴尬。寒武纪是带着巨额亏损上市的：2017年至2019年，公司营业收入分别仅为784.3万元、1.17亿元和4.44亿元，归母净利润分别为-3.81亿元、-0.41亿元和-11.79亿元[s38][s43]。上市之后，亏损并未收窄反而扩大：2020年至2023年，营收分别为4.59亿、7.21亿、7.29亿、7.09亿元，归母净利润分别为-4.35亿、-8.25亿、-12.57亿、-8.48亿元，其中2022年12.57亿元的亏损额创下纪录[s37][s39]。亏损的主因清晰而残酷：芯片研发是吞金兽，公司每年研发投入占营收比例长期居高不下，部分年份甚至超过100%，而云端芯片从流片到规模商用需要数年爬坡，收入迟迟跟不上烧钱的速度[s39][s43]。界面新闻在报道中算了一笔账：高研发投入压顶、政府补助贡献相当部分利润、商业化进度低于预期，是那些年寒武纪财报的三大特征[s37]。

股价随基本面与情绪剧烈摆动。上市初期的千亿市值很快回落，2022年更是雪上加霜——CTO离职、实体清单冲击、业绩亏损三重压力叠加，4月27日股价盘中触及46.59元，较64.39元的发行价深度破发，较297.77元的高点跌去逾八成[s38][s37]。36氪在一篇报道中写道：上市五年，寒武纪累计亏损数十亿元，市值却长期在千亿上下浮动，多空分歧之大在A股罕见[s43]。创投股东也开始用脚投票，多家早期机构披露减持计划，第一财经报道了股东拟清仓式减持的动向，“亏损扩大、股价暴涨、股东离场”成为那段时间最常见的标题组合[s38]。

为支撑持续烧钱的研发，寒武纪两度向资本市场伸手。2022年度，公司向特定对象发行股票募资16.72亿元，发行价121.10元/股、发行1380.6万股，2023年4月完成登记；彼时市场低迷，定增机构一度浮亏，直到行情回暖后才转为浮盈近120%[s39][s40]。2025年8月15日，公司拟募资不超过49.8亿元的定增申请获科创板上市委会议通过，被称为科创板AI芯片领域最大单笔定增；注册稿中募投规模调整为不超过39.85亿元，分别投向先进芯片平台（20.54亿元）、软件平台（14.52亿元）及补充流动资金（4.79亿元）[s54]。

回头看，2020年到2023年是寒武纪的“荒野时期”：顶着第一股的光环，却年年交出亏损的财报；被国产替代的叙事托举，又被迟迟不来的规模出货反复证伪[s37][s43]。陈天石在这几年里极少公开辩解，只是把融来的钱一笔笔砸进思元290、370、590的流片和软件生态里[s7][s49]。46.59元的冰点时刻，公司账上靠着定增募资和理财维持运转，很少有人相信，不到三年后这只股票会以万亿市值重新成为全场焦点[s40][s43]。那段时间他常对团队说的一句话是：“战略就是老实干活、搬砖。”[s7]

上市本身的细节至今仍是科创板的经典案例。证券时报记录了首日交易：开盘集合竞价即报250元，较64.39元发行价上涨288%，全天成交逾百亿，收盘价212.4元对应总市值约860亿元，盘中一度突破千亿[s11]。中国证券报的申购报道显示，公司发行价对应市值已超过200亿元，而市场给予的首日定价接近发行价四倍，资金对“国产AI芯片少有的上市标的”的稀缺性给出了极致定价[s12]。但从第二个交易周开始，股价便随着每一份亏损财报逐级走低[s37]。

财报里的另一组数字更能说明那些年的窘迫：2022年公司归母净利润-12.57亿元，而当年营收仅7.29亿元，亏损额接近营收的两倍；中国经济网在年报报道中指出，上市三年公司两次直接融资合计约42.5亿元，几乎全部投入研发与补流，账面货币资金一度全靠募资支撑[s39]。界面新闻则注意到，政府补助和理财收益在那些年对利润表贡献显著，扣非后的亏损更为刺眼[s37]。创投股东的退出也从未间断——第一财经报道，股价每一轮反弹都伴随着早期机构减持公告，甚至出现“清仓式减持”计划[s38]。

2023年4月落地的定增，是那个低谷期里的关键输血：121.10元/股、1380.6万股、募资16.72亿元，认购机构包括多家国资背景基金[s39][s40]。金融界后来复盘，这批定增股份在发行后一度浮亏近三成，直到2023年底大模型行情启动才扭亏为盈，浮盈最高时接近120%——长线国资在最悲观时点的下注，与寒武纪后来的爆发形成了精准的呼应[s40]。而2025年8月过会的49.8亿元定增（注册稿调整为不超过39.85亿元），则是公司在景气高点为下一代芯片平台和软件生态储备的弹药[s54]。

**English:** In 2020, with the STAR Market barely a year old, the market was waiting for a landmark target. Cambricon took only about 117 days from acceptance of its filing to hearing, a record pace for STAR review that the media called a "lightning pass."[s10][s43] On July 20, Cambricon (688256) formally listed on the STAR Market at an issue price of 64.39 yuan, raising 2.582 billion yuan from newly issued shares; it opened at 250 yuan, a gain of 288% over the issue price, and closed at 212.4 yuan, up 229.86%, with first-day market value breaking through 100 billion yuan at a stroke.[s11][s12] On July 23, the share price reached an intraday early-listing high of 297.77 yuan.[s38] During subscription, the market had already circulated the claim that "a single winning lot could earn 100,000 yuan," and IPO-subscription fever was unprecedented.[s12] The title of "first AI-chip stock" was bound to Cambricon from then on.[s11]

Yet the capital-market glare could not hide the embarrassment of the financials. Cambricon listed carrying huge losses: from 2017 to 2019, revenue was a mere 7.843 million yuan, 117 million yuan and 444 million yuan respectively, while net profit attributable to shareholders was -381 million, -41 million and -1.179 billion yuan.[s38][s43] After listing, losses did not narrow but widened: from 2020 to 2023, revenue was 459 million, 721 million, 729 million and 709 million yuan, and net profit attributable to shareholders was -435 million, -825 million, -1.257 billion and -848 million yuan, with the 2022 loss of 1.257 billion yuan a record.[s37][s39] The main cause of the losses was clear and brutal: chip R&D is a money-devouring beast, and the company's R&D spending as a share of revenue stayed stubbornly high for years, in some years exceeding 100%, while cloud chips need years of climbing from tape-out to commercial deployment at scale, and revenue chronically failed to keep pace with the burn rate.[s39][s43] Jiemian News did the arithmetic in its reporting: crushing R&D spending, government subsidies contributing a substantial portion of reported profit, and commercialization progressing slower than expected were the three defining features of Cambricon's financials in those years.[s37]

The share price swung violently with fundamentals and sentiment. The early-listing market value of 100 billion-plus yuan soon fell back, and 2022 added insult to injury—three pressures stacked together: the CTO's departure, the shock of the Entity List and the loss-making results. On April 27 the share price touched 46.59 yuan intraday, deeply below the 64.39-yuan issue price and down more than 80% from the 297.77-yuan high.[s38][s37] 36Kr wrote in a report: five years after listing, Cambricon had lost several billion yuan cumulatively, yet its market value long hovered around the 100-billion mark, with a bull-bear divide rare in the A-share market.[s43] Venture-capital shareholders also began to vote with their feet; several early institutions disclosed reduction plans, and Yicai reported moves toward "liquidation-style" selling, with "losses widening, share price soaring, shareholders leaving" becoming the most common headline combination of the period.[s38]

To sustain the cash-burning R&D, Cambricon twice reached into the capital markets. For fiscal 2022, the company raised 1.672 billion yuan through a private placement to specific investors at 121.10 yuan per share for 13.806 million shares, with registration completed in April 2023; with the market then in the doldrums, the placement institutions at one point sat on floating losses, which turned into a floating gain of nearly 120% only after the market warmed.[s39][s40] On August 15, 2025, the STAR listing committee approved the company's application for a private placement to raise no more than 4.98 billion yuan, called the largest single placement in the STAR Market's AI-chip field; in the registered version the planned raise was adjusted to no more than 3.985 billion yuan, allocated between an advanced-chip platform (2.054 billion), a software platform (1.452 billion) and working capital (479 million).[s54]

In retrospect, 2020 to 2023 were Cambricon's "wilderness years": wearing the halo of the first stock yet filing loss-making results year after year; buoyed by the narrative of import substitution yet repeatedly disproven by scale shipments that never arrived.[s37][s43] In those years Chen Tianshi rarely offered public excuses; he simply poured the money raised, tranche by tranche, into the tape-outs of the MLU290, 370 and 590 and into the software ecosystem.[s7][s49] At the ice-point moment of 46.59 yuan, the company kept running on placement proceeds and wealth-management income, and few believed that less than three years later the stock would return to the center of attention with a trillion-yuan market value.[s40][s43] A line he often said to the team in that period was: "Strategy is simply working honestly, laying bricks."[s7]

The details of the listing itself remain a classic STAR Market case. Securities Times recorded the first day's trading: the opening call auction immediately produced a price of 250 yuan, up 288% from the 64.39-yuan issue price, with turnover exceeding 10 billion yuan through the day and the closing price of 212.4 yuan corresponding to a total market value of about 86 billion yuan, which briefly topped 100 billion intraday.[s11] The China Securities Journal's subscription coverage showed that the issue price already corresponded to a market value above 20 billion yuan, while the market's first-day pricing came close to four times the issue price—capital awarding an extreme premium to the scarcity of "the only pure-play domestic AI-chip target."[s12] But from the second trading week onward, the share price drifted lower, level by level, with every loss-making report.[s37]

Another set of figures in the financials speaks even more clearly to the straits of those years: in 2022 the company's net loss attributable to shareholders was 1.257 billion yuan against revenue of just 729 million, the loss nearly twice revenue; China Economic Net noted in its annual-report coverage that in the three years after listing the company's two direct financings totaled about 4.25 billion yuan, nearly all of it poured into R&D and working capital, with cash on the balance sheet at one point supported entirely by fundraising.[s39] Jiemian News, for its part, noted that government subsidies and wealth-management income contributed markedly to the income statement in those years, and that losses after deducting non-recurring items were even more glaring.[s37] The exit of venture shareholders never let up either—Yicai reported that every rebound came accompanied by reduction announcements from early institutions, and even "liquidation-style" reduction plans.[s38]

The placement that landed in April 2023 was the key transfusion of that trough: 121.10 yuan per share, 13.806 million shares, 1.672 billion yuan raised, with subscribers including several state-backed funds.[s39][s40] JRJ.com later reconstructed the story: the placement shares at one point carried a floating loss of nearly 30% after issuance, and only turned profitable when the large-model rally began at the end of 2023, with floating gains approaching 120% at their peak—the bets of long-term state capital at the most pessimistic point formed a precise counterpoint to Cambricon's later explosion.[s40] And the 4.98-billion-yuan placement approved in August 2025 (adjusted to no more than 3.985 billion in the registered version) was ammunition the company stockpiled at the top of the cycle for the next-generation chip platform and software ecosystem.[s54]

### 6. 至暗时刻：实体清单、梁军离职与42亿元股权诉讼 / The Darkest Hours: The Entity List, Liang Jun's Departure and the 4.287-Billion-Yuan Equity Lawsuit

2022年冬天，寒武纪接连遭遇两场重击。

第一场来自海外。美国东部时间2022年12月15日、北京时间12月16日，美国商务部工业与安全局（BIS）将寒武纪等36家中国实体列入“实体清单”，理由是所谓“支持中国军事现代化”；寒武纪及其部分子公司等21家实体还被同时纳入外国直接产品规则（FDPR）名单，意味着受美国设备和技术约束的境外代工厂，也可能被限制为其生产[s14][s37]。消息公布当日，寒武纪股价盘中跌超8%[s14]。对一家采用Fabless模式、高度依赖境外先进制程的芯片设计公司而言，实体清单的杀伤力直指供应链命门：先进工艺流片、EDA工具、第三方IP等环节都可能受到波及[s14][s33]。

第二场来自身边。2022年3月14日，寒武纪公告称，公司核心技术人员、CTO梁军因“与公司存在分歧”于2月10日离职[s29][s25]。梁军是中国芯片界的重量级人物：2000年加入华为，曾任海思半导体麒麟SoC芯片总架构师，2017年加入寒武纪出任CTO，正是他在华为与寒武纪之间牵线，推动寒武纪IP进入麒麟970；加盟后，他主导了思元290、370、590等几代云端芯片的架构设计[s25][s28]。公告次日，寒武纪股价大跌约18%，市值蒸发数百亿元，市场将其解读为“技术灵魂出走”[s29]。

关于分歧的实质，多家媒体的还原指向路线之争：陈天石与公司管理层希望在国产替代窗口期加快产品落地、抢抓市场机遇；梁军则更倾向于沉下心打磨底层技术、在微架构上做长期深耕[s25][s29]。钛媒体在报道中指出，科学家创业公司在商业化冲刺期，技术理想主义与市场机会主义的冲突几乎不可避免，梁军的离开是这种张力的集中爆发[s25]。量子位则用“高层震荡”形容此事对市场信心的打击[s29]。

余波远比想象中长。2024年10月，梁军提起劳动争议诉讼，向寒武纪索赔股权激励损失约42.87亿元——其依据是自己间接持有的1152.3万股限制性股票，按当时约372元/股的市价估算[s26][s27]。寒武纪方面则主张按双方协议约定，以约5.2万元的对价回购相关股权；在此之前的相关仲裁与诉讼程序中，梁军的诉求已两度败诉[s26][s28]。2025年1月，南方都市报披露了这场“5万元对42亿元”的离奇对峙；11月，央广网报道案件仍在审理中，42.87亿元的标的额使其成为国内芯片行业最大的劳动争议之一[s26][s27]。蓝鲨硬科技梳理了纠纷始末：核心争议在于离职后限制性股票的回购条件与定价——协议签署时股价低迷、按离职时净资产定价近乎象征性补偿，而此后股价上涨数十倍，让这笔股权成了双方都无法让步的巨额财富[s28]。

同一时期，车载业务也踩了刹车。2023年，行歌科技被曝裁员近半，自动驾驶芯片投入收缩；甲子光年在调查中指出，车载芯片研发周期长、车规认证门槛高、客户回款慢，在主业尚未造血的阶段，寒武纪不得不收缩战线、集中资源保云端基本盘[s41]。

实体清单、核心高管出走、子公司裁员，三件事在一年内接踵而至，把寒武纪推到了成立以来的最低谷[s14][s41]。但陈天石没有改变路线：他在内部讲话中要求团队“脑子一根筋，耐得住寂寞”，把全部注意力收回到芯片本身[s7]。面对供应链的不确定性，公司一边推进多渠道产能安排，一边加速与国产软件生态的适配[s42][s33]。事后看，正是这段至暗时刻里的不动摇，让思元590按计划在2023年底走向市场，接住了半年后大模型爆发带来的滔天需求[s42][s51]。

实体清单的冲击机制值得细说。对采用Fabless模式的寒武纪而言，自身不拥有晶圆厂，芯片设计完成后需委托代工厂流片；实体清单加FDPR规则的组合拳，理论上可以切断其获取先进工艺产能和部分美国技术组件的通道[s14][s37]。消息公布后，市场恐慌集中爆发，寒武纪当日盘中跌幅超过8%，随后数月股价一路探至46.59元的历史最低点[s14][s38]。但此后两年的事实表明，制裁在压缩供给弹性的同时，也把国内客户“逼”向了国产芯片——2023年底思元590如期上市并快速进入大模型客户的采购名单，公司通过供应链多源化和产能前置安排维持了出货节奏[s42][s33]。

梁军事件则给所有科学家创业公司上了一课。钛媒体复盘认为，梁军代表的是“海思式”工程体系文化——追求架构极致、长周期打磨；而2021年后寒武纪面对的是窗口期稍纵即逝的国产替代市场，管理层必须在产品完美与商业时机之间做取舍，两种理性的碰撞最终以核心高管出走收场[s25]。量子位在离职公告刊发当日的报道中记录，市场以18%的单日跌幅为这场分歧“投票”[s29]。而后续的股权诉讼，又把早期股权激励协议的模糊地带暴露在聚光灯下：离职时按约定近乎象征性回购的限制性股票，在股价上涨数十倍后变成价值数十亿的争议标的[s26][s28]。央广网2025年11月报道显示，42.87亿元索赔的劳动争议案仍在司法程序中，其判决结果将对科技行业股权激励的退出机制产生示范意义[s27]。

行歌收缩则是另一重理性的选择。甲子光年获悉，2023年行歌裁员比例接近一半，SD5226等高阶项目节奏放缓，资源向云端主业集中[s41]。在陈天石的算盘里，云端大模型芯片是必须赢下的主战场，车载是值得保留但不能拖垮现金流的侧翼——先活下来、先把思元590的仗打好，比四条战线同时开火更重要[s41][s7]。

**English:** In the winter of 2022, Cambricon took two heavy blows in succession.

The first came from overseas. On December 15, 2022, U.S. Eastern Time (December 16, Beijing time), the U.S. Commerce Department's Bureau of Industry and Security (BIS) added Cambricon and 35 other Chinese entities to the "Entity List," citing the claimed reason of "supporting China's military modernization"; Cambricon and some of its subsidiaries, together with 20 other entities, were simultaneously placed on the Foreign Direct Product Rule (FDPR) list, meaning that overseas foundries bound by U.S. equipment and technology could also be restricted from manufacturing for them.[s14][s37] On the day the news was published, Cambricon's share price fell more than 8% intraday.[s14] For a fabless chip-design company highly dependent on overseas leading-edge processes, the Entity List struck straight at the supply chain's lifeline: leading-process tape-outs, EDA tools and third-party IP could all be affected.[s14][s33]

The second blow came from close at hand. On March 14, 2022, Cambricon announced that Liang Jun, its core technical officer and CTO, had left the company on February 10 "because of disagreements with the company."[s29][s25] Liang is a heavyweight in China's chip world: he joined Huawei in 2000, served as chief architect of HiSilicon's Kirin SoC chips, joined Cambricon as CTO in 2017, and it was he who acted as the bridge between Huawei and Cambricon, driving Cambricon's IP into the Kirin 970; after coming aboard, he led the architecture design of several cloud-chip generations including the MLU290, 370 and 590.[s25][s28] On the trading day after the announcement, Cambricon's share price plunged about 18%, wiping tens of billions of yuan off its market value, in a market reading of the event as "the departure of the technical soul."[s29]

On the substance of the disagreement, multiple media reconstructions point to a contest over route: Chen Tianshi and the management wanted to speed products to market and seize market opportunities during the window of import substitution; Liang leaned toward settling down to polish the underlying technology and cultivating the microarchitecture for the long term.[s25][s29] TMTPost observed in its reporting that in a scientist-founded company during the commercialization sprint, conflict between technical idealism and market opportunism is almost unavoidable, and Liang's departure was the concentrated eruption of that tension.[s25] QbitAI described the blow the event dealt to market confidence as "high-level turmoil."[s29]

The aftermath stretched far longer than expected. In October 2024, Liang filed a labor-dispute lawsuit against Cambricon, claiming compensation of about 4.287 billion yuan for stock-incentive losses—based on the 11.523 million restricted shares he held indirectly through an equity platform, valued at the then market price of roughly 372 yuan per share.[s26][s27] Cambricon, for its part, maintained that under the agreement between the two sides, the relevant equity should be repurchased for consideration of about 52,000 yuan; in the prior related arbitration and litigation proceedings, Liang's claims had already lost twice.[s26][s28] In January 2025, the Southern Metropolis Daily's Wan Finance unit disclosed the bizarre confrontation of "50,000 yuan against 4.2 billion yuan"; in November, CNR reported that the case was still being heard, and that the 4.287-billion-yuan amount made it one of the largest labor disputes in China's chip industry.[s26][s27] BlueShark Hard Tech laid out the whole dispute: the core disagreement lies in the repurchase terms and pricing of restricted shares after departure—the agreement was signed when the share price was depressed, with pricing based on net assets at departure amounting to near-symbolic compensation, while the subsequent several-dozen-fold rise in the share price turned that equity into a vast fortune neither side could concede.[s28]

In the same period, the automotive business also hit the brakes. In 2023, Xingge Technology was reported to have cut nearly half its workforce and retrenched its autonomous-driving chip investment; Jazzy Gravity's investigation noted that automotive chips have long R&D cycles, high qualification barriers and slow customer payments, and that at a stage when the main business had not yet begun to generate cash, Cambricon had to contract its lines and concentrate resources on defending the cloud base.[s41]

Entity List, the departure of a core executive, and subsidiary layoffs—three events arriving on each other's heels within a single year pushed Cambricon to the lowest valley since its founding.[s14][s41] But Chen Tianshi did not change course: in internal talks he asked the team to be "single-minded to the point of obstinacy, able to endure loneliness," pulling all attention back to the chips themselves.[s7] Facing supply-chain uncertainty, the company on the one hand pushed multi-channel capacity arrangements and on the other accelerated adaptation with the domestic software ecosystem.[s42][s33] In hindsight, it was precisely that refusal to waver in the darkest hours that let the MLU590 go to market on schedule at the end of 2023, catching the towering demand brought by the large-model explosion half a year later.[s42][s51]

The mechanism of the Entity List's shock is worth spelling out. For a fabless Cambricon, which owns no fabs and must entrust foundries with tape-out after the design is complete, the combination of the Entity List and the FDPR could, in theory, cut off its access to leading-process capacity and some U.S.-technology components.[s14][s37] After the news broke, market panic burst out in concentration: Cambricon fell more than 8% intraday that day, and over the following months the share price explored all the way down to its all-time low of 46.59 yuan.[s14][s38] But the facts of the two years since show that, while compressing supply elasticity, the sanctions also " forced" domestic customers toward domestic chips—the MLU590 launched on schedule at the end of 2023 and quickly entered large-model customers' procurement lists, and the company maintained its shipment rhythm through supply-chain multi-sourcing and front-loaded capacity arrangements.[s42][s33]

The Liang Jun affair taught a lesson to every scientist-founded company. TMTPost's reconstruction holds that Liang represented the "HiSilicon-style" engineering-system culture—pursuit of architectural perfection, long-cycle polishing—whereas after 2021 Cambricon faced an import-substitution market whose window could vanish in an instant, and management had to trade off between product perfection and commercial timing; the collision of two kinds of rationality ended, in the event, with the departure of a core executive.[s25] QbitAI's report on the day the departure announcement was published recorded that the market "voted" on the disagreement with an 18% single-day fall.[s29] The subsequent equity lawsuit, in turn, exposed the gray zones of early stock-incentive agreements to the spotlight: restricted shares repurchased on departure for near-symbolic consideration became, after the share price rose several dozen times, a disputed asset worth billions.[s26][s28] CNR's November 2025 report shows the 4.287-billion-yuan labor-dispute case remains in the judicial process, and its verdict will set a precedent for exit mechanisms in tech-industry stock incentives.[s27]

The Xingge retrenchment was another rational choice. Jazzy Gravity learned that in 2023 Xingge cut close to half its staff, the cadence of high-end projects such as the SD5226 slowed, and resources were concentrated on the cloud main business.[s41] In Chen Tianshi's reckoning, cloud large-model chips were the main battlefield that must be won, while automotive was a flank worth preserving but one that must not be allowed to drag down cash flow—surviving first, and first fighting the MLU590 battle well, mattered more than opening fire on four fronts at once.[s41][s7]

### 7. 卖铲人：思元590、国产替代与中立第三方的生意 / The Shovel Seller: The MLU590, Import Substitution and the Business of the Neutral Third Party

2023年以来，大模型的爆发让算力成为这个时代最紧俏的战略物资，也彻底改写了寒武纪的命运曲线。当英伟达高端GPU因出口管制难以进入中国数据中心，当华为昇腾因华为云自身下场提供云服务而让部分互联网客户心存顾虑，一个独立、中立、只卖芯片不碰应用的本土供应商，突然成了市场最渴求的角色——这就是陈天石为寒武纪定下的“卖铲人”定位[s51][s33]。

“卖铲人”的逻辑并不复杂：淘金热里最稳赚的是卖铲子的人。大模型厂商千帆竞渡，无论谁最终胜出，都要买算力；寒武纪不做云服务、不做大模型、不与下游客户争利，只提供训练和推理所需的芯片与软件栈，因此可以成为所有淘金者共同的供应商[s51][s32]。21世纪经济报道在调研中发现，互联网大厂之所以愿意把寒武纪列为华为昇腾之外的第二选项，正是看中其中立性——采购一家纯Fabless芯片公司的产品，不必担心议价权和业务数据落入潜在竞争对手之手[s51][s33]。据多家媒体报道，字节跳动已成为寒武纪的第一大客户，公司云端产品线收入在2025年占总营收比重超过99%[s15][s16]。

承接这波需求的主力，是思元590。这款7纳米芯片综合性能约为英伟达A100的八成，而其真正的竞争力在于对国产大模型生态的全面适配：DeepSeek、通义千问、智谱GLM、月之暗面Kimi等主流模型均完成了在590平台上的部署与优化，客户“拿来即用”的成熟度在国产芯片中领先[s42][s51]。在CUDA生态难以撼动的现实下，寒武纪把NeuWare软件栈和BANG语言作为重中之重，尽可能降低开发者从英伟达平台迁移的成本[s33][s49]。陈天石在2024年度业绩说明会上表示，公司正“积极拓展市场份额，加速场景落地”，并把软件生态投入视为与硬件同等重要的长期工程[s50][s49]。

宏观环境站在了寒武纪一边。英伟达CEO黄仁勋公开表示，受出口管制影响，英伟达在中国AI芯片市场的份额已从约95%降至几乎为零[s33]；行业测算显示，2025年国产AI芯片在国内市场的占有率已提升至约35%至41%[s32][s51]。供给真空与国产替代政策共振，让寒武纪的订单在2024年下半年开始集中释放，2024年第四季度公司营收环比大幅跳升[s15][s36]。面对市场上关于公司产品路线图、客户名单和产能的种种传闻，寒武纪一度发布紧急公告，称网传产品、客户、供应及产能预测信息均为不实信息，提醒投资者以官方公告为准[s31]。

竞争同样激烈。华为昇腾依托自研生态和政企渠道强势扩张，海光、壁仞、摩尔线程等本土公司各有山头，阿里平头哥、百度昆仑芯等互联网大厂自研芯片也在分流需求——36氪把寒武纪“第一次赚钱”的同时被众多“中国版英伟达”包围的局面，形容为一场刚开场就已白热化的竞赛[s32][s33]。陈天石对此的态度是：赛道足够大，容得下多家供应商；寒武纪的护城河在于先发的指令集积累、完整的云边端车产品线和不碰下游的中立立场[s33][s51]。

2024年11月20日，陈天石在世界互联网大会乌镇峰会主论坛上发言：“我们已经身处有史以来最伟大的技术变革时刻。人工智能不仅是技术进步的工具，它更有潜力成为增进人类福祉的强大力量。”[s48] 这句宏大叙事的背后，是一个朴素的商业判断：智能时代的算力需求才刚起步，而寒武纪要做的，就是那个在所有淘金者身后安静递上铲子的人[s48][s51]。

“卖铲人”模式的威力，在2025年年报里体现为一组惊人的结构变化：云端产品线收入64.76亿元、占总营收99%以上，综合毛利率55.15%，而四年前公司营收还在7亿元量级徘徊、毛利率远低于此[s16][s15]。需求端，字节跳动等互联网大客户的集中采购，让公司从“到处找订单”变为“排产保交付”，年末49.44亿元存货对应的正是尚未确认收入的在手订单[s15][s16]。供给端，英伟达H100/H200级别的高端GPU无法正常进入中国，H20等降规产品也面临政策反复，国产客户被迫将“可用、好用”的国产芯片纳入核心算力规划[s33][s32]。

但“卖铲人”的位置并非高枕无忧。36氪的报道提醒，互联网大客户自研芯片的动机始终存在——阿里平头哥、百度昆仑芯都已迭代数代，一旦自研成熟，外部采购比例随时可能下调，这与当年华为自研NPU的逻辑如出一辙[s32]。此外，客户集中度过高也让单一客户的资本开支节奏直接决定寒武纪的季度业绩弹性[s15]。公司对此的应对，一是持续扩展客户名单、降低对单一客户的依赖，二是把训练芯片与软件生态做深做厚，提高切换成本[s50][s49]。2026年初，针对市场上关于客户、产能、产品路线的大量猜测，寒武纪发布澄清公告，指相关预测均为不实信息——这从侧面说明，“卖铲人”的生意好到连传闻都能搅动股价[s31]。

陈天石在乌镇峰会上把人工智能称为“有史以来最伟大的技术变革”，并强调技术应服务于人类福祉[s48]；在业绩说明会的问答里，他则反复回到最朴素的表达：把产品做好、把场景落地、把市场份额一步一步拿下来[s50]。宏大叙事与搬砖式务实之间的落差，正是这位“卖铲人”最真实的姿态[s48][s7]。

**English:** Since 2023, the explosion of large models has made compute the most sought-after strategic commodity of the age, and it has rewritten Cambricon's fortune curve entirely. When high-end Nvidia GPUs could barely enter Chinese data centers because of export controls, and when Huawei Ascend left some internet customers uneasy because Huawei Cloud itself competes in cloud services, an independent, neutral domestic supplier that sold only chips and touched no applications suddenly became the role the market craved most—this is the "shovel seller" position Chen Tianshi defined for Cambricon.[s51][s33]

The logic of the shovel seller is not complicated: in a gold rush, the surest money is made by the man who sells shovels. Large-model makers race by the thousand, and whoever ultimately wins must buy compute; Cambricon runs no cloud services, builds no large models, and does not compete with downstream customers for profit—it supplies only the chips and software stack needed for training and inference—and can therefore be the common supplier of every gold prospector.[s51][s32] In its field research, the 21st Century Business Herald found that the reason the big internet firms were willing to place Cambricon as their second option after Huawei Ascend was precisely its neutrality—buying products from a pure-play fabless chip company, they need not fear that bargaining power and business data will fall into the hands of a potential competitor.[s51][s33] According to multiple media reports, ByteDance has become Cambricon's largest customer, and in 2025 the company's cloud product-line revenue accounted for more than 99% of total revenue.[s15][s16]

The main force carrying this wave of demand is the MLU590. This 7-nanometer chip has overall performance of roughly 80% of the Nvidia A100, but its real competitiveness lies in comprehensive adaptation to the domestic large-model ecosystem: mainstream models including DeepSeek, Qwen, Zhipu GLM and Moonshot Kimi have all completed deployment and optimization on the 590 platform, and its "ready to use out of the box" maturity leads among domestic chips.[s42][s51] Facing the reality that the CUDA ecosystem is hard to shake, Cambricon treats the NeuWare software stack and the BANG language as top priorities, lowering as far as possible the cost of developers migrating from the Nvidia platform.[s33][s49] At the 2024 annual results briefing, Chen Tianshi said the company was "actively expanding market share and accelerating scenario deployment," and treated software-ecosystem investment as a long-term project equal in importance to hardware.[s50][s49]

The macro environment lined up on Cambricon's side. Nvidia's CEO Jensen Huang said publicly that, under export controls, Nvidia's share of China's AI-chip market had fallen from about 95% to virtually zero;[s33] industry estimates show that in 2025 domestic AI chips' share of the Chinese market had risen to roughly 35% to 41%.[s32][s51] The supply vacuum resonating with import-substitution policy let Cambricon's orders begin to release in concentration in the second half of 2024, and revenue jumped sharply quarter on quarter in the fourth quarter of 2024.[s15][s36] Faced with market rumors about the company's product roadmap, customer list and capacity, Cambricon at one point issued an urgent announcement saying that online information purporting to forecast products, customers, supply and capacity was all false, and reminding investors to rely on official announcements.[s31]

Competition is equally fierce. Huawei Ascend expands aggressively on the strength of its in-house ecosystem and government-enterprise channels; domestic firms such as Hygon, Biren and Moore Threads each hold their own ground; and the in-house chips of internet giants, including Alibaba's T-Head and Baidu's Kunlunxin, are also diverting demand—36Kr described Cambricon's situation of "making money for the first time" while surrounded by numerous "Chinas' Nvidias" as a race that had turned white-hot before it had even properly started.[s32][s33] Chen Tianshi's attitude is that the track is large enough to hold multiple suppliers; Cambricon's moat lies in its first-mover accumulation of instruction sets, its complete cloud-edge-device-car product line, and its neutral stance of not touching downstream business.[s33][s51]

On November 20, 2024, Chen Tianshi spoke at the main forum of the World Internet Conference Wuzhen Summit: "We are already living through the greatest moment of technological change in history. Artificial intelligence is not merely a tool for technological progress; it has the potential to become a powerful force for the improvement of human welfare."[s48] Behind that grand narrative lies a plain commercial judgment: the compute demand of the intelligent age has only just begun, and what Cambricon wants to be is the man who quietly hands out shovels behind all the gold prospectors.[s48][s51]

The power of the shovel-seller model showed up in the 2025 annual report as a set of startling structural changes: cloud product-line revenue of 6.476 billion yuan, more than 99% of total revenue, and an overall gross margin of 55.15%, whereas four years earlier the company's revenue had still been hovering in the 700-million-yuan range with a far lower gross margin.[s16][s15] On the demand side, concentrated procurement by big internet customers such as ByteDance turned the company from "hunting for orders everywhere" to "scheduling production to guarantee delivery," and the 4.944 billion yuan of inventory at year-end corresponded precisely to orders in hand not yet recognized as revenue.[s15][s16] On the supply side, high-end GPUs of the Nvidia H100/H200 class could not enter China normally, and downgraded products such as the H20 also faced policy reversals, forcing domestic customers to build "usable, easy-to-use" domestic chips into their core compute plans.[s33][s32]

Yet the shovel-seller position is not free from anxiety. 36Kr's reporting warned that the big internet customers' incentive to build their own chips never disappears—Alibaba's T-Head and Baidu's Kunlunxin have both iterated through several generations, and once in-house designs mature, external procurement can be dialed down at any moment, in a logic identical to Huawei's in-house NPU move in years past.[s32] Furthermore, excessive customer concentration lets a single customer's capital-expenditure cadence directly determine the elasticity of Cambricon's quarterly results.[s15] The company's response is, first, to keep broadening the customer list and reducing dependence on any single customer, and second, to deepen and thicken the training chips and software ecosystem so as to raise switching costs.[s50][s49] In early 2026, amid a flood of market speculation about customers, capacity and product roadmaps, Cambricon issued a clarification announcement stating the relevant forecasts to be false—which shows, from the side, that the shovel-seller's business had become so good that even rumors could move the share price.[s31]

At the Wuzhen Summit Chen Tianshi called artificial intelligence "the greatest technological change in history" and stressed that technology should serve human welfare;[s48] in the Q&A at earnings briefings, by contrast, he returned again and again to the plainest expressions: make the products well, land the scenarios, and take market share step by step.[s50] The gap between grand narrative and brick-laying pragmatism is precisely this shovel seller's truest posture.[s48][s7]

### 8. 万亿时刻：从首季盈利到科创板股王 / The Trillion-Yuan Moment: From First Quarterly Profit to King of the A-Share Market

转折发生在2024年第四季度。这一季度，寒武纪实现营业收入9.89亿元、归母净利润2.72亿元，上市以来首次单季度扭亏为盈，持续多年的亏损曲线就此拐头[s17][s32]。真正的爆发在2025年：根据公司在上交所披露的2025年年度报告，寒武纪全年实现营业收入64.97亿元（6,497,196,198.68元），同比增长453.21%；归母净利润20.59亿元，扣非净利润17.70亿元——这是公司2020年上市以来首个年度盈利[s16][s15]。分季度看，四个季度营收分别为11.11亿、17.69亿、17.27亿、18.90亿元，净利润分别为3.55亿、6.83亿、5.67亿、4.55亿元，盈利基础逐季夯实[s16]。

年报的细节勾勒出一家真正运转起来的芯片公司：云端产品线收入64.76亿元，占总营收99%以上，客户结构以大型互联网企业为主；全年综合毛利率55.15%；研发投入11.69亿元，占营收17.99%；研发人员887人，占员工总数80.13%；年末存货49.44亿元，对应着仍在放量的在手订单[s16][s15]。公司同时推出上市以来首次分红，合计派发约6.32亿元，并实施“10转4.9派15元”的权益分配[s17]。每日经济新闻在报道中也提示了隐忧：存货激增与客户高度集中，仍是悬在高增长之上的两大风险点[s15]。

增长势头在2026年延续。一季度，公司实现营收28.85亿元，同比增长159.56%；归母净利润10.13亿元，同比增长185.04%；扣非净利润9.34亿元[s18]。上半年，营收59.96亿元，同比增长108.13%；归母净利润23.11亿元，同比增长122.61%；扣非净利润21.66亿元，综合毛利率55.25%——半年净利润已超过2025年全年水平[s36]。华鑫证券在半年报点评中认为，业绩高增彰显盈利弹性，国产算力需求共振正在打开寒武纪的长期成长空间[s36]。

与业绩同频的是股价的狂飙。2025年8月27日，寒武纪股价盘中触及1464.98元，自上市以来首次超过贵州茅台，虽仅“半日游”便回落，却已点燃市场情绪[s19][s20]；8月28日，公司收报1587.91元，正式收盘登顶A股“股王”，次日市值达6643亿元[s19][s21]。21世纪经济报道算了一笔账：85后创始人陈天石的身家随之突破1700亿元[s21]。2026年5月6日，股价盘中最高冲至1966元、对应市值最高达8296亿元[s18]；6月30日，寒武纪市值盘中突破1万亿元，最高触及1607元、对应市值10040亿元，成为科创板历史上第一只万亿市值股票[s34][s35]。截至2026年9月6日，公司收报1072元，总市值约6741亿元，在大幅波动中仍是A股最受关注的科技标的之一[s18][s34]。

陈天石个人的财富曲线随之陡峭。上市时他持股约29.87%，2025年年中约为28.63%，2026年增持5857万股后持股约28.35%，始终是公司绝对控制人[s21][s17]。财富榜单上，他2020年以约202.8亿元身家登上福布斯中国富豪榜；2024年以320亿元成为胡润榜上的南昌首富；2025年以870亿元身家位列胡润全球富豪榜第195位；同年9月，福布斯全球富豪榜估算其财富约215亿美元，为江西首富[s3][s52][s21]。

对这一切，陈天石保持着一贯的冷静。在业绩说明会上，他谈得最多的仍是产品路线、生态建设和场景落地，而非市值[s50][s49]。从46.59元的冰点到万亿市值的王座，市场只用了不到三年；但造出这一切的人很清楚，真正支撑股价的，是数据中心里一颗颗正在运行的思元芯片，而不是K线图上的喧嚣[s16][s34]。

这场市值狂奔的每个节点都被媒体密集记录。2025年8月27日午后，寒武纪盘中股价达到1464.98元，首次超过同时刻贵州茅台的股价，现代快报称其“半日封王”，当天市值约5740亿元[s19][s20]；次日公司收报1587.91元，以收盘价正式加冕A股“股王”[s19]。2026年5月6日，股价盘中最高1966元、市值峰值8296亿元，长江商报用“冲击2000元未果仍是A股之王”描述市场对2000元整数关口的争夺[s18]。6月30日的万亿时刻则被齐鲁晚报、北京日报同步记录：科创板开市七年来，第一次有公司市值盘中突破一万亿元，而这家公司五年前还在46.59元的冰点上挣扎[s34][s35]。

财富榜上的跃迁同样惊人。河南商报援引胡润与福布斯榜单追踪：2020年公司上市当年，陈天石以约202.8亿元身家登上福布斯中国富豪榜；2024年，他以320亿元成为胡润榜上的南昌首富；2025年，870亿元身家让他位列胡润全球富豪榜第195位，同年9月福布斯全球富豪榜估算其财富约215亿美元、登顶江西首富[s52][s3]。21世纪经济报道则在股王诞生当日算出，陈天石的持股对应身家已突破1700亿元人民币[s21]。

伴随高估值的永远是多空之辩。看多方的逻辑是国产算力刚需+稀缺标的+业绩持续兑现：2025年营收增长4.5倍、2026年上半年再翻一倍，利润表从亏损五十余亿元到半年净赚23亿元的反转速度，超出了所有模型的预期[s16][s36]。看空方则提示，存货49.44亿元意味着重资产式的备货风险，第一大客户贡献过半收入意味着议价权脆弱，而万亿市值对应的估值水平已远超全球可比芯片公司[s15][s17]。长江商报还注意到一个细节：公司首次分红6.32亿元的同时，牛散章建平家族持股市值一度达到92亿元——财富效应的外溢，已成A股市场的一景[s17]。

**English:** The turn came in the fourth quarter of 2024. That quarter, Cambricon achieved revenue of 989 million yuan and net profit attributable to shareholders of 272 million yuan, its first single-quarter profit since listing, and the loss curve that had run for years bent at last.[s17][s32] The real explosion came in 2025: according to the company's 2025 annual report disclosed on the Shanghai Stock Exchange, Cambricon achieved full-year revenue of 6.497 billion yuan (6,497,196,198.68 yuan), up 453.21% year on year; net profit attributable to shareholders was 2.059 billion yuan, and net profit after deducting non-recurring items was 1.770 billion yuan—the company's first annual profit since its 2020 listing.[s16][s15] By quarter, revenue was 1.111 billion, 1.769 billion, 1.727 billion and 1.890 billion yuan, and net profit was 355 million, 683 million, 567 million and 455 million yuan respectively, with the profitability base consolidating quarter by quarter.[s16]

The details of the annual report sketch a chip company genuinely in motion: cloud product-line revenue of 6.476 billion yuan, more than 99% of total revenue, with a customer structure dominated by large internet enterprises; an overall gross margin of 55.15% for the year; R&D spending of 1.169 billion yuan, 17.99% of revenue; 887 R&D staff, 80.13% of the workforce; and inventory of 4.944 billion yuan at year-end, corresponding to orders in hand still ramping.[s16][s15] The company also launched its first dividend since listing, distributing about 632 million yuan in total, and implemented an equity distribution of "4.9 bonus shares for every 10 plus a 15-yuan cash dividend."[s17] The National Business Daily also flagged the worries in its reporting: surging inventory and heavy customer concentration remain the two risk points hanging over the high growth.[s15]

The momentum carried into 2026. In the first quarter, the company achieved revenue of 2.885 billion yuan, up 159.56% year on year; net profit attributable to shareholders of 1.013 billion yuan, up 185.04%; and net profit after deducting non-recurring items of 934 million yuan.[s18] For the first half, revenue was 5.996 billion yuan, up 108.13%; net profit attributable to shareholders 2.311 billion yuan, up 122.61%; net profit after deductions 2.166 billion yuan; and overall gross margin 55.25%—the half-year net profit already exceeded the full-year 2025 level.[s36] In its interim-report review, China Fortune Securities argued that the high growth demonstrates earnings elasticity and that the resonating demand for domestic compute is opening Cambricon's long-term growth space.[s36]

In step with the results came a runaway share price. On August 27, 2025, Cambricon's share price touched 1,464.98 yuan intraday, surpassing Kweichow Moutai for the first time since listing; though it fell back after a mere "half-day reign," the move ignited market sentiment;[s19][s20] on August 28 the company closed at 1,587.91 yuan, formally taking the closing-price crown as the A-share market's most expensive stock, and market value reached 664.3 billion yuan the next day.[s19][s21] The 21st Century Business Herald did the arithmetic: the fortune of the 1980s-born founder Chen Tianshi thereupon broke through 170 billion yuan.[s21] On May 6, 2026, the share price climbed as high as 1,966 yuan intraday, corresponding to a peak market value of 829.6 billion yuan;[s18] on June 30, Cambricon's market value broke through 1 trillion yuan intraday, touching as high as 1,607 yuan per share and 1.004 trillion yuan in market value, becoming the first 1-trillion-yuan stock in the STAR Market's history.[s34][s35] As of September 6, 2026, the company closed at 1,072 yuan with a total market value of about 674.1 billion yuan, remaining, amid heavy volatility, one of the most closely watched technology targets in the A-share market.[s18][s34]

Chen Tianshi's personal wealth curve steepened with it. He held about 29.87% of the company at listing, about 28.63% in mid-2025, and about 28.35% after increasing his holdings by 58.57 million shares in 2026, always remaining the company's absolute controller.[s21][s17] On the wealth lists, he appeared on the 2020 Forbes China Rich List with wealth of about 20.28 billion yuan; in 2024 he became Nanchang's richest man on the Hurun list with 32 billion yuan; in 2025 he ranked 195th on the Hurun Global Rich List with wealth of 87 billion yuan; and in September of the same year Forbes' World's Billionaires list estimated his fortune at about US$21.5 billion, making him Jiangxi's richest man.[s3][s52][s21]

Through it all, Chen Tianshi kept his customary calm. At earnings briefings, what he talked about most was still product roadmaps, ecosystem building and scenario deployment, rather than market value.[s50][s49] From the ice point of 46.59 yuan to the throne of a trillion-yuan market value took the market less than three years; but the man who built it all knows well that what truly undergirds the share price is the MLU chips actually running, one by one, in data centers—not the noise of the candlestick chart.[s16][s34]

Every node of this market-value stampede was densely recorded by the media. On the afternoon of August 27, 2025, Cambricon's intraday price reached 1,464.98 yuan, overtaking Kweichow Moutai's price at the same moment for the first time; the Modern Express called it a "half-day coronation," with market value that day around 574 billion yuan;[s19][s20] the next day the company closed at 1,587.91 yuan, formally crowned the A-share market's most expensive stock at the closing price.[s19] On May 6, 2026, the intraday high was 1,966 yuan with a peak market value of 829.6 billion yuan, and the Changjiang Times described the market's tussle over the 2,000-yuan round-number mark with the phrase "failing to break 2,000, yet still king of the A shares."[s18] The trillion-yuan moment of June 30 was recorded in parallel by the Qilu Evening News and the Beijing Daily: in the seven years since the STAR Market opened, a company's market value had broken 1 trillion yuan intraday for the first time—and five years earlier that same company had been struggling at the ice point of 46.59 yuan.[s34][s35]

The leap up the wealth lists was equally astonishing. As the Henan Business Daily tracked, citing Hurun and Forbes: in 2020, the year of listing, Chen Tianshi appeared on the Forbes China Rich List with wealth of about 20.28 billion yuan; in 2024 he became Nanchang's richest man on the Hurun list with 32 billion yuan; in 2025, wealth of 87 billion yuan placed him 195th on the Hurun Global Rich List, and in September Forbes' World's Billionaires list estimated his fortune at about US$21.5 billion, crowning him Jiangxi's richest man.[s52][s3] The 21st Century Business Herald, on the day the stock king was born, calculated that Chen Tianshi's holdings corresponded to a fortune already exceeding 170 billion yuan.[s21]

High valuations are forever accompanied by bull-bear debate. The bull case rests on rigid demand for domestic compute plus scarcity of targets plus sustained delivery of results: revenue grew 4.5-fold in 2025 and doubled again in the first half of 2026, and the speed of the income-statement reversal—from cumulative losses of more than 5 billion yuan to 2.3 billion yuan of net profit in half a year—exceeded every model's expectations.[s16][s36] The bears point out that inventory of 4.944 billion yuan implies heavy-asset stockpiling risk, that a single largest customer contributing more than half of revenue implies fragile bargaining power, and that the valuation level corresponding to a trillion-yuan market value already far exceeds that of comparable chip companies worldwide.[s15][s17] The Changjiang Times also noted one detail: as the company paid its first dividend of 632 million yuan, the holdings of the family of famed retail investor Zhang Jianping were at one point worth 9.2 billion yuan—the spillover of the wealth effect had become a spectacle of the A-share market.[s17]

### 9. 长跑哲学：科学家创业的中国样本 / The Philosophy of the Long Race: A Chinese Specimen of the Scientist-Entrepreneur

在所有关于陈天石的描述中，“低调”是出现频率最高的词。他常年背双肩包、穿格子衬衫，走在中科院计算所的园区里像个普通研究员；创办千亿市值公司后，仍不爱应酬、鲜少接受采访，被媒体追问身家与目标时，他的回答是：“我只是个普通的科研人员，各种机缘下创办了一家公司，远远没到大咖档位。我个人也没什么远大目标，没想过自己一定要成为一个多么杰出的人，只想做好本职工作。”[s8][s3] 每日经济新闻在一篇报道中把他称为“天才创始人”，却也注意到他面对镜头时那种与身份不符的局促与谦逊[s8]。

这种谦逊背后，是一套一以贯之的长跑哲学。2020年3月，公司成立刚四年、正值聚光灯下，陈天石在接受甲子光年专访时却说了一段后来被反复引用的话：“Intel今年52岁，AMD今年51岁，NVIDIA今年27岁。寒武纪只有4岁，和行业前辈比起来还只是个孩子。罗马并非一天建成，前辈标杆也都是筚路蓝缕走过来的，我们有远大的志向，但长跑才刚刚开始。”[s7][s47] 他把做芯片比作推土机作业：“沿大路而行，不抄小道，不搞奇袭”，战略上“老实干活、搬砖”，心态上“脑子一根筋，耐得住寂寞”[s7][s1]。

这正是科学家创业的典型范式：先在论文里定义问题，再在指令集里给出答案，然后用一代又一代流片把答案磨成产品，最后用软件生态把产品变成行业标准[s44][s45]。陈天石和陈云霁从DianNao到Cambricon的学术积累，是寒武纪区别于所有同行的底牌——全球能同时拿出ASPLOS、MICRO最佳论文和商用芯片产品线的团队，屈指可数[s2][s46]。但陈天石也清醒，学术血统不能自动兑换成商业胜利，他不讳言雄心：“智能时代将诞生硬件新巨头，我不讳言寒武纪想当这个‘将军’。”[s8] 只是这位将军选择的打法，是结硬寨、打呆仗[s7]。

他为寒武纪设定的产业角色同样透着克制——“希望大家有一天不需要知道我们，只要我们能够支持好下游的应用，这就是我们的贡献。”[s3] 不做云、不做模型、不与客户争利，甘当算力基础设施的底层供应商，这种“隐形”定位在热衷讲故事的AI行业并不讨巧，却在大模型时代转化为最宝贵的中立性，让字节跳动这样的超级客户敢于把核心算力压在一家“只卖铲子”的公司身上[s51][s33]。他甚至愿意把自己放在历史长河里衡量：英特尔52岁、AMD 51岁、英伟达27岁，芯片业的伟大公司都以十年为单位计分，寒武纪的赛程才刚开场[s7][s47]。

当然，前路并非坦途。英伟达近二十年筑起的CUDA生态，仍是寒武纪NeuWare需要长期追赶的高墙；第一大客户贡献过半收入的集中度、互联网大厂自研芯片的分流、先进制程的供给约束、万亿市值背后随时可能反转的情绪，每一项都是实打实的考题[s32][s33][s49]。2025年业绩说明会上，面对“能否实现百亿营收”的提问，管理层的回答谨慎而务实——积极拓展份额、加速场景落地[s49][s50]。陈天石深知，芯片行业没有一劳永逸的胜利：思元590之后还有下一代架构，大模型之后还有新的智能范式[s48]。

但无论市值如何起落，陈天石和寒武纪已经为中国科技产业留下了一个样本：两个从南昌走出的少年班兄弟，在一间30平方米的小黑屋里坐了七八年冷板凳，把一个用拼音命名的芯片架构写进了全球体系结构的教科书，又把它变成了数据中心里支撑中国大模型的国产算力[s2][s46][s55]。2024年乌镇，他说“我们已经身处有史以来最伟大的技术变革时刻”[s48]；而对这位自称“普通科研人员”的企业家来说，最幸运的事或许是——他不仅亲历了这场变革，还亲手造出了变革所需的那种最基础、也最坚硬的物质[s3][s8]。长跑没有终点，搬砖仍在继续[s7]。

理解陈天石的经营哲学，有三个细节常被身边人提起。其一，公司成立十年，他的办公室始终贴着一张纸，上书“沿大路而行，不抄小道，不搞奇袭”——在动辄追逐风口的AI行业，寒武纪从未跨界做过矿机、未炒过概念、未借壳套利，每一轮融资都投向芯片和软件[s1][s7]。其二，他把研发组织成“推土机”模式：不押注单点奇技，而是按指令集、微架构、工艺、软件栈几条线齐头并进，一代芯片流片时下一代已在定义之中，思元100到590的稳定迭代节奏就是证据[s23][s24]。其三，在华为分手、梁军离职、实体清单这些关口，他从不在公开场合抱怨或辩解，回应市场的方式始终是下一份财报、下一颗芯片[s14][s29]。

中国新闻网在“天才兄弟”的报道中写道，陈天石最打动人的品质是“分寸感”：他清楚寒武纪的边界——不做应用、不碰数据、不与客户竞争，把“隐形”当作商业模式的一部分[s3]；也清楚自己的边界——把兄长陈云霁留在学术源头，把梁军这样的工程大将放在架构一线，自己只做方向、资源和节奏的取舍[s4][s25]。每日经济新闻的“多事之秋”报道则呈现了另一面：这位温和的科学家在董事会上同样会为路线分歧拍板，梁军离职、行歌收缩、客户押注，每一个重大决定背后都有他的签字[s8][s41]。科学家的耐心与企业家的决断，在他身上以一种少见的比例共存[s7]。

从更长的时间尺度看，陈天石与寒武纪的样本意义在于回答了一个问题：中国的硬科技创业，能不能走“从论文到产品、从书架到货架”的原生路径？龙芯一代人用二十年证明中国人能造CPU，而陈氏兄弟证明了在一个全新的计算范式里，中国人可以参与定义游戏规则——DianNao架构被全球同行引用，Cambricon指令集的思想渗透进后续各类AI加速器，思元芯片则在数据中心里承接真实的大模型负载[s44][s45][s55]。这条路没有捷径，只有小黑屋、冷板凳和一代接一代的流片[s47]。当被问及希望外界如何记住寒武纪，陈天石的回答依旧朴素：“希望大家有一天不需要知道我们，只要我们能够支持好下游的应用，这就是我们的贡献。”[s3] 对一位万亿市值公司的创始人而言，把自己的终极目标定义为“不被知道”，这或许是整个故事里最有格局的一句话[s3][s51]。

**English:** Of all the descriptions applied to Chen Tianshi, "low profile" is the word that appears most often. He carries a backpack year round and wears checked shirts, looking like an ordinary researcher as he walks through the ICT campus; even after founding a company worth hundreds of billions, he still dislikes socializing and rarely gives interviews. When the press presses him about his fortune and his goals, his answer is: "I'm just an ordinary researcher who, through a series of coincidences, founded a company. I'm nowhere near the ranks of the big names. I don't have any grand personal goals either—I never set out to become some outstanding person. I only want to do my own job well."[s8][s3] The National Business Daily called him a "genius founder" in one report, yet also noticed the awkwardness and modesty, out of keeping with his status, that he shows in front of the camera.[s8]

Behind that modesty lies a consistent philosophy of the long race. In March 2020, with the company just four years old and squarely in the spotlight, Chen Tianshi, in an interview with Jazzy Gravity (published by Jiemian News), spoke a passage that has been quoted again and again since: "Intel is 52 years old this year, AMD is 51, and NVIDIA is 27. Cambricon is only four—compared with the industry's predecessors, we are still just a child. Rome was not built in a day, and every benchmark company made its way up from scratch through great hardship. We have lofty ambitions, but the long race has only just begun."[s7][s47] He compares making chips to operating a bulldozer: "Stay on the main road; take no shortcuts, spring no surprises"—in strategy, "work honestly, lay bricks"; in mindset, "single-minded to the point of obstinacy, able to endure loneliness."[s7][s1]

This is precisely the classic paradigm of the scientist-entrepreneur: first define the problem in a paper, then give the answer in an instruction set, then grind the answer into a product through generation after generation of tape-outs, and finally turn the product into an industry standard through a software ecosystem.[s44][s45] The academic accumulation of Chen Tianshi and Chen Yunji, from DianNao to Cambricon, is the hidden card that distinguishes Cambricon from every peer—teams anywhere in the world that can simultaneously produce ASPLOS and MICRO best papers and a commercial chip product line can be counted on one's fingers.[s2][s46] Yet Chen Tianshi is also clear-eyed that an academic bloodline does not automatically redeem into commercial victory, and he does not hide his ambition: "The intelligent age will give birth to new hardware giants, and I make no secret of the fact that Cambricon wants to be that 'general.'"[s8] It is just that this general has chosen the tactic of building fortified camps and fighting stolid battles.[s7]

The industrial role he set for Cambricon likewise breathes restraint: "I hope that one day people won't need to know who we are. As long as we can support the applications downstream of us well, that is our contribution."[s3] No cloud, no models, no competing with customers for profit; willingly serving as the bottom-layer supplier of compute infrastructure—this "invisible" positioning is not a crowd-pleaser in an AI industry fond of storytelling, but in the large-model age it has converted into the most precious neutrality, letting a super-customer such as ByteDance dare to stake its core compute on a company that "only sells shovels."[s51][s33] He is even willing to measure himself against the long river of history: Intel at 52, AMD at 51, Nvidia at 27—the chip industry's great companies are scored in decades, and Cambricon's race has only just started.[s7][s47]

The road ahead, of course, is not smooth. The CUDA ecosystem that Nvidia built over nearly twenty years remains a high wall that Cambricon's NeuWare must chase for a long time to come; the concentration of more than half of revenue in the largest customer, the diversion from the internet giants' in-house chips, the supply constraints on leading-edge processes, and the sentiment behind a trillion-yuan market value that could reverse at any moment—each is a genuine examination question.[s32][s33][s49] At the 2025 earnings briefing, faced with the question of whether the company could achieve 10 billion yuan in revenue, management's answer was cautious and pragmatic—actively expanding share and accelerating scenario deployment.[s49][s50] Chen Tianshi knows well that there is no once-and-for-all victory in the chip industry: after the MLU590 comes the next-generation architecture, and after large models comes a new paradigm of intelligence.[s48]

But however the market value rises and falls, Chen Tianshi and Cambricon have already left Chinese technology industry a specimen: two brothers of the gifted class who walked out of Nanchang, sat on the cold bench for seven or eight years in a thirty-square-meter little black room, wrote a chip architecture named in pinyin into the global textbooks of computer architecture, and then turned it into the domestic compute that underpins China's large models in the data centers.[s2][s46][s55] At Wuzhen in 2024, he said that "we are already living through the greatest moment of technological change in history";[s48] and for this self-described "ordinary researcher," perhaps the luckiest thing is that he has not only lived through the change but built with his own hands the most basic, and hardest, kind of matter the change requires.[s3][s8] The long race has no finish line; the brick-laying continues.[s7]

To understand Chen Tianshi's management philosophy, those close to him often raise three details. First, in the company's ten years of existence, his office has always borne a sheet of paper reading "Stay on the main road; take no shortcuts, spring no surprises"—in an AI industry that chases every wind, Cambricon has never strayed into crypto-mining rigs, never hyped concepts, never used backdoor listings for arbitrage, and every round of financing has gone into chips and software.[s1][s7] Second, he organizes R&D on a "bulldozer" model: rather than betting on a single-point wizardry, several lines—instruction set, microarchitecture, process, software stack—advance abreast, with the next generation already being defined while the current generation tapes out; the steady iteration cadence from MLU100 to MLU590 is the evidence.[s23][s24] Third, at the junctures of the Huawei breakup, Liang Jun's departure and the Entity List, he never complained or made excuses in public; his response to the market was always the next financial report and the next chip.[s14][s29]

China News Service wrote in its "genius brothers" report that Chen Tianshi's most moving quality is his sense of proportion: he is clear about Cambricon's boundaries—no applications, no data, no competition with customers, treating "invisibility" as part of the business model;[s3] and clear about his own boundaries—keeping his brother Chen Yunji at the academic source, placing engineering generals such as Liang Jun on the architecture front line, and reserving for himself only the trade-offs of direction, resources and cadence.[s4][s25] The National Business Daily's "troubled autumn" report presents the other side: this gentle scientist will just as firmly bang the table on route disagreements in the boardroom, and his signature lies behind every major decision—Liang Jun's departure, the Xingge retrenchment, the customer bets.[s8][s41] The scientist's patience and the entrepreneur's decisiveness coexist in him in a rare proportion.[s7]

On a longer timescale, the specimen significance of Chen Tianshi and Cambricon lies in answering a question: can China's hard-technology entrepreneurship follow an indigenous path "from paper to product, from bookshelf to shelf"? The Loongson generation spent twenty years proving that Chinese people can build CPUs, while the Chen brothers proved that in a brand-new computing paradigm, Chinese people can take part in defining the rules of the game—the DianNao architecture is cited by peers worldwide, the ideas of the Cambricon instruction set have seeped into every kind of AI accelerator that followed, and MLU chips take on real large-model workloads in the data centers.[s44][s45][s55] There are no shortcuts on this road, only the little black room, the cold bench, and tape-out after tape-out across the generations.[s47] When asked how he hopes the outside world will remember Cambricon, Chen Tianshi's answer remains plain: "I hope that one day people won't need to know who we are. As long as we can support the applications downstream of us well, that is our contribution."[s3] For the founder of a trillion-yuan company to define his ultimate goal as "not being known" may be the most magnanimous sentence in the whole story.[s3][s51]

## 卷尾 Editorial Conclusion

五亿年前，寒武纪生命大爆发，地球在极短的地质时间里涌现出绝大多数动物门类，奇虾睁开了那个时代最先进的复眼，三叶虫披上钙质的硬甲，海口鱼长出了最初的脊索[s3]。陈氏兄弟用这个地质年代为公司命名，是相信人工智能也会迎来这样一场物种大爆发——而他们要做的，是这场爆发里最底层的那层“物质载体”。2026年回头看，这个隐喻竟一语成谶：大模型以月为单位迭代，智能算力以年为单位翻倍，而提供算力的芯片，真的成了这个时代最稀缺、最昂贵、也最具战略意义的“器官”。

陈天石常说自己“只是个普通的科研人员”，说寒武纪和英特尔、AMD、英伟达这些前辈比起来“还只是个孩子”，“长跑才刚刚开始”[s7][s8]。这句在2020年公司四岁时讲的话，在市值万亿之后依然成立——甚至更加成立。因为芯片产业的历史反复证明：这是一场以十年为单位计分的马拉松，一时的市值喧嚣终会退潮，真正留下的是指令集、微架构、软件生态和一代代被客户用起来的芯片。英伟达用近二十年筑起CUDA生态，寒武纪的NeuWare软件栈才刚学会走路；互联网大客户的自研芯片、华为昇腾的强势扩张、先进制程的供给约束，每一项都是悬在前方的考题[s32][s33][s49]。万亿市值不是终章，而是一张更难的考卷的扉页。

但陈天石和寒武纪的意义，早已不能用一家公司的盈亏来衡量。他们证明了一件事：中国人可以在一个全新的计算范式里，从最底层的指令集开始做出原创性贡献，而不只是在别人定义的赛道上追赶。DianNao论文里那个以中文拼音命名的芯片架构，如今被全世界的体系结构研究者引用；从一间30平方米小黑屋里走出来的技术路线，支撑起了中国大模型产业相当一部分的国产算力[s2][s55]。陈天石本人的财富在数年间膨胀十余倍，2025年他以870亿元身家登上胡润全球富豪榜[s3][s52]，可他最在意的评价或许仍是当年那句朴素的话：“希望大家有一天不需要知道我们，只要我们能够支持好下游的应用，这就是我们的贡献。”[s3]

地质学上的寒武纪，结束于一次物种大灭绝之后的漫长演化；而硅基世界的“寒武纪”，才刚刚拉开帷幕。当奇虾的复眼在数据中心里闪烁、三叶虫的硬甲化作国产芯片的防线、海口鱼的脊索撑起智能时代的脊梁，那个在南昌城里沉迷游戏、靠抓阄选中计算机专业的少年，已经把自己的名字写进了中国科技史[s52]。长跑没有终点。对陈天石而言，万亿市值只是路边一块里程碑——前方的路，仍然是那条他走了十八年、并且打算一直走下去的冷板凳。

More than 500 million years ago, the Cambrian explosion unfolded: in a geological instant, the Earth produced the great majority of its animal phyla. Anomalocaris opened the most advanced compound eyes of its age, trilobites clothed themselves in calcareous armor, and Haikouichthys grew the earliest notochord.[s3] When the Chen brothers named their company after that geological period, they were expressing a belief that artificial intelligence would likewise witness an explosion of species—and that what they wanted to build was the lowest, most fundamental layer of "material substrate" for that explosion. Looking back from 2026, the metaphor has proved uncannily accurate: large models iterate by the month, intelligent compute doubles by the year, and the chips that supply that compute have indeed become the scarcest, most expensive and most strategically vital "organ" of the age.

Chen Tianshi often says he is "just an ordinary researcher," and that compared with predecessors such as Intel, AMD and Nvidia, Cambricon is "still just a child" and "the long race has only just begun."[s7][s8] Those words, first spoken in 2020 when the company was four years old, still hold after the trillion-yuan valuation—indeed, they hold all the more. For the history of the chip industry proves again and again that this is a marathon scored in decades: the noise of market value eventually recedes, and what remains is instruction sets, microarchitectures, software ecosystems and generation after generation of chips that customers actually use. Nvidia spent nearly twenty years building the CUDA ecosystem; Cambricon's NeuWare software stack is only just learning to walk. The in-house chips of the big internet customers, Huawei Ascend's aggressive expansion, and the constraints on access to leading-edge process capacity—each is an examination question hanging over the road ahead.[s32][s33][s49] A trillion-yuan market value is not a final chapter; it is the title page of a harder examination paper.

Yet the significance of Chen Tianshi and Cambricon can no longer be measured by the profits and losses of a single company. They proved something: that Chinese engineers can make original contributions in an entirely new computing paradigm, starting from the very bottom layer of the instruction set, rather than merely chasing along a track defined by others. The chip architecture in the DianNao paper, named in Chinese pinyin, is now cited by computer-architecture researchers all over the world; the technical route that walked out of a thirty-square-meter room now underpins a substantial share of the domestic compute powering China's large-model industry.[s2][s55] Chen's own fortune swelled more than tenfold in a few years, and in 2025 he appeared on the Hurun Global Rich List with wealth of 87 billion yuan;[s3][s52] yet the judgment he cares about most may still be that plain sentence from years ago: "I hope that one day people won't need to know who we are. As long as we can support the applications downstream of us well, that is our contribution."[s3]

In geology, the Cambrian period ended in a long evolution after a mass extinction; in the silicon world, the "Cambrian" is only just raising its curtain. As Anomalocaris's compound eyes flicker inside data centers, as the trilobite's armor hardens into the defensive line of domestic chips, and as Haikouichthys's notochord becomes the backbone of the age of intelligence, the boy who was obsessed with video games in Nanchang and chose his computer-science major by drawing lots has written his name into the history of Chinese science and technology.[s52] The long race has no finish line. For Chen Tianshi, a trillion-yuan market value is no more than a milestone at the roadside—and the road ahead remains the same cold bench he has sat on for eighteen years, and intends to keep sitting on.

## 金句 Pull Quote

> Intel今年52岁，AMD今年51岁，NVIDIA今年27岁。寒武纪只有4岁，和行业前辈比起来还只是个孩子。罗马并非一天建成，前辈标杆也都是筚路蓝缕走过来的，我们有远大的志向，但长跑才刚刚开始。

> Intel is 52 years old this year, AMD is 51, and NVIDIA is 27. Cambricon is only four—compared with the industry's predecessors, we are still just a child. Rome was not built in a day, and every benchmark company made its way up from scratch through great hardship. We have lofty ambitions, but the long race has only just begun.

## 履历时间线 Career Timeline

- **1985-06** 出生于江西南昌 / Born in Nanchang, Jiangxi
  - 陈天石出生于江西南昌一个普通家庭，兄长陈云霁长他两岁，后于1997年14岁时考入中国科学技术大学少年班。兄弟俩先后走上中科大少年班之路，日后被媒体并称为为AI装上中国芯的“双子星”[s1][s53][s4]。
  - EN: Chen Tianshi was born into an ordinary family in Nanchang, Jiangxi Province. His elder brother Chen Yunji, two years his senior, would be admitted at fourteen in 1997 to the Special Class for the Gifted Young at the University of Science and Technology of China. The two brothers followed one another down the road to USTC's gifted class, and were later paired by the media as the "twin stars" who fitted a Chinese chip into artificial intelligence.[s1][s53][s4]
- **2001** 16岁考入中科大少年班 / Enters USTC gifted class at 16
  - 陈天石从南昌二中考入中国科学技术大学少年班。大学期间他一度沉迷电子游戏、成绩平平，自嘲为“学渣”，读研时靠抓阄选中计算机方向，2005年获数学学士学位，2010年获中科大计算机博士学位，师从陈国良院士与姚新教授[s52][s53][s5]。
  - EN: He was admitted from Nanchang No. 2 Middle School to the Special Class for the Gifted Young at the University of Science and Technology of China. At university he was for a time addicted to video games with mediocre grades, mocking himself as an "academic underdog"; he chose the computer-science direction by drawing lots when applying for graduate study, took a bachelor's degree in mathematics in 2005, and earned his USTC doctorate in computer science in 2010, studying under Academician Chen Guoliang and Professor Yao Xin.[s52][s53][s5]
- **2008** 与陈云霁在计算所启动AI芯片研究 / Launches AI-chip research with brother
  - 在中科院计算所，陈天石与兄长陈云霁带领十余人团队，在一间约30平方米的“小黑屋”里启动人工智能与芯片交叉研究。彼时深度学习尚未爆发，为AI设计专用芯片属冷门中的冷门，论文接连被拒，团队只能在本职工作之余“坐冷板凳”[s1][s2][s47]。
  - EN: At the Institute of Computing Technology of the Chinese Academy of Sciences, Chen Tianshi and his elder brother Chen Yunji led a team of a dozen-odd people in launching cross-disciplinary research into artificial intelligence and chips in a roughly thirty-square-meter "little black room." Deep learning had not yet exploded, designing dedicated chips for AI was the coldest corner of a cold field, papers were rejected one after another, and the team could only "sit on the cold bench" in the margins of their day jobs.[s1][s2][s47]
- **2010** 博士毕业进入中科院计算所 / Joins ICT after PhD
  - 陈天石获中科大计算机博士学位后进入中科院计算技术研究所，与哥哥陈云霁成为同事，正式投入深度学习处理器方向研究。陈云霁此前已师从“龙芯之父”胡伟武，25岁出任八核龙芯3号主架构师，兄弟二人一个懂算法、一个懂芯片，开始深度协作[s2][s4][s5]。
  - EN: After taking his doctorate in computer science from USTC, Chen Tianshi joined the Institute of Computing Technology of the Chinese Academy of Sciences, becoming his brother Chen Yunji's colleague and formally throwing himself into deep-learning-processor research. Chen Yunji had earlier studied under Hu Weiwu, the "father of Loongson," and became chief architect of the eight-core Loongson 3 at twenty-five; one brother understood algorithms and the other chips, and they began to collaborate in depth.[s2][s4][s5]
- **2014-03** DianNao论文获ASPLOS最佳论文 / DianNao paper wins ASPLOS best paper
  - 陈天石作为第一作者、与Olivier Temam及陈云霁等合作的DianNao论文获国际体系结构顶级会议ASPLOS 2014最佳论文，为中国大陆及亚洲机构首次。DianNao在3.02平方毫米、485毫瓦下实现4520亿次运算每秒，能效远超同期通用芯片[s44][s46][s2]。
  - EN: The DianNao paper, with Chen Tianshi as first author in collaboration with Olivier Temam and Chen Yunji among others, won the best-paper award at ASPLOS 2014, a top international computer-architecture conference—the first time an institution from mainland China or Asia had received the honor. DianNao delivered 452 billion operations per second within 3.02 square millimeters and 485 milliwatts, with energy efficiency far beyond contemporary general-purpose chips.[s44][s46][s2]
- **2015** 全球首款深度学习处理器原型流片 / First deep-learning processor prototype taped out
  - 团队完成全球首款深度学习专用处理器原型芯片流片，实测能效达传统芯片近百倍，首次在硅片上验证了学术架构的工程可行性。同年PuDianNao、ShiDianNao论文分别发表于ASPLOS与ISCA，DaDianNao获MICRO 2014最佳论文[s2][s6][s44]。
  - EN: The team completed tape-out of the world's first prototype chip for a dedicated deep-learning processor; measured energy efficiency reached nearly a hundred times that of traditional chips, validating for the first time in silicon the engineering feasibility of the academic architecture. In the same year the PuDianNao and ShiDianNao papers were published at ASPLOS and ISCA respectively, while DaDianNao had won the MICRO 2014 best-paper award.[s2][s6][s44]
- **2016-03** 创办寒武纪，注册资本90万元 / Founds Cambricon with 900,000 yuan capital
  - 北京中科寒武纪科技有限公司注册成立，初始注册资本90万元，其中陈天石出资63万元、中科院计算所旗下中科算源出资27万元。公司名取自五亿年前的寒武纪生命大爆发，寓意人工智能的“物种大爆发”；陈天石离岗创业，陈云霁留在计算所[s1][s43][s3]。
  - EN: Beijing Zhongke Cambricon Technology Co., Ltd. was registered with an initial capital of 900,000 yuan, of which Chen Tianshi contributed 630,000 yuan and Zhongke Suanyuan, an investment vehicle under the ICT, contributed 270,000. The company's name was taken from the Cambrian life explosion 500 million years ago, symbolizing the expected "explosion of species" in artificial intelligence; Chen Tianshi left his post to start the company, while Chen Yunji stayed at the ICT.[s1][s43][s3]
- **2016** 发布全球首款商用终端智能处理器IP寒武纪1A / Launches Cambricon 1A terminal IP
  - 寒武纪推出寒武纪1A智能处理器IP，号称全球首款商用终端深度学习处理器IP，可在手机等终端设备高效运行AI任务。同年公司获中科院旗下基金数千万元天使投资，2016年8月完成Pre-A轮融资，元禾原点、科大讯飞、涌铧投资等入局[s5][s10][s13]。
  - EN: Cambricon launched the Cambricon 1A intelligent-processor IP, billed as the world's first commercial terminal deep-learning-processor IP, capable of running AI tasks efficiently on terminal devices such as smartphones. In the same year the company received tens of millions of yuan in angel investment from a fund under the Chinese Academy of Sciences, and completed a Pre-A round in August 2016 with investors including Yuanhe Origin, iFlytek and Yonghua Investment.[s5][s10][s13]
- **2017-08-18** 完成1亿美元A轮融资，估值10亿美元 / Raises US$100m Series A at US$1bn valuation
  - 寒武纪宣布完成1亿美元A轮融资，由国投创业领投，阿里巴巴、联想创投、国科投资、中科图灵等跟投，公司估值达到10亿美元，跻身AI芯片“独角兽”。2018年6月再完成数亿美元B轮融资，估值升至25亿美元[s12][s13]。
  - EN: Cambricon announced a US$100 million Series A led by SDIC Venture Capital, with Alibaba, Legend Capital, CAS Investment and Zhongke Turing among the followers, lifting its valuation to US$1 billion and placing it among the AI-chip "unicorns." In June 2018 it completed a further several-hundred-million-dollar Series B, at a valuation of US$2.5 billion.[s12][s13]
- **2017-09-02** 华为麒麟970集成寒武纪1A发布 / Huawei Kirin 970 integrates Cambricon 1A
  - 华为在德国柏林IFA展发布全球首款内置独立NPU的手机芯片麒麟970，其AI核心采用寒武纪1A处理器IP。该芯片采用台积电10纳米工艺、集成55亿个晶体管，搭载于Mate 10等十余款机型；此后麒麟980采用双核寒武纪1H。2017、2018年公司IP收入占比分别达98.33%、99.69%[s9][s38][s43]。
  - EN: At the IFA show in Berlin, Huawei launched the Kirin 970, the world's first smartphone chip with a built-in standalone NPU, its AI core using the Cambricon 1A processor IP. Built on TSMC's 10-nanometer process with 5.5 billion transistors, it was deployed in more than ten models including the Mate 10; the later Kirin 980 used the dual-core Cambricon 1H. IP revenue accounted for 98.33% and 99.69% of the company's revenue in 2017 and 2018 respectively.[s9][s38][s43]
- **2018-05** 发布中国首款高峰值云端AI芯片思元100 / Launches MLU100 cloud AI chip
  - 寒武纪发布思元100（MLU100）云端智能芯片，支持云端推理与训练，标志公司从终端IP供应商向云端芯片厂商转型。2018年2月《Science》杂志报道评价中国科学家在AI芯片领域作出“开创性贡献”[s22][s6][s43]。
  - EN: Cambricon released the MLU100 (Siyuan 100) cloud intelligent chip, supporting both cloud inference and training and marking the company's transformation from a terminal IP supplier into a cloud-chip vendor. In February 2018 the journal Science, in related coverage, appraised Chinese scientists' work in AI chips as a "pioneering contribution."[s22][s6][s43]
- **2019-06-20** 发布第二代云端芯片思元270 / Launches second-gen MLU270 cloud chip
  - 寒武纪发布思元270云端AI芯片，采用台积电16纳米工艺，INT8峰值算力128TOPS，搭载自研MLUv02指令集。同年面向边缘计算的思元220面世，此后累计销量突破百万片，成为公司出货量最大的产品[s22][s30][s42]。
  - EN: Cambricon released the MLU270 cloud AI chip, built on TSMC's 16-nanometer process with peak INT8 compute of 128 TOPS and carrying the self-developed MLUv02 instruction set. The same year saw the MLU220 for edge computing, whose cumulative sales later surpassed a million units to become the company's highest-volume product.[s22][s30][s42]
- **2019** 华为自研达芬奇架构，终端IP合作终止 / Huawei goes in-house with Da Vinci
  - 华为2018年10月公布自研达芬奇AI架构，2019年搭载自研NPU的麒麟810问世，逐步终止与寒武纪的IP授权合作。寒武纪终端IP收入占比从2018年的99.69%骤降至2019年的15.49%，公司坚定转向云端主战场[s33][s38][s9]。
  - EN: Huawei unveiled its self-developed Da Vinci AI architecture in October 2018, and in 2019 the Kirin 810 with its in-house NPU appeared, gradually ending the IP licensing partnership with Cambricon. Cambricon's terminal-IP revenue share plunged from 99.69% in 2018 to 15.49% in 2019, and the company committed firmly to the cloud as its main battlefield.[s33][s38][s9]
- **2020-07-20** 寒武纪登陆科创板，“AI芯片第一股” / Lists on STAR Market as first AI-chip stock
  - 寒武纪（688256）在上交所科创板上市，发行价64.39元、募资25.82亿元，开盘报250元涨288%，收报212.4元涨229.86%，首日市值突破千亿元；从受理到上市仅约117天。公司由此成为“AI芯片第一股”，但上市时仍连年亏损[s11][s12][s10]。
  - EN: Cambricon (688256) listed on the Shanghai Stock Exchange's STAR Market at an issue price of 64.39 yuan, raising 2.582 billion yuan; it opened at 250 yuan, up 288%, and closed at 212.4 yuan, up 229.86%, with first-day market value breaking 100 billion yuan. From filing acceptance to listing took only about 117 days. The company became known as the "first AI-chip stock," though it was still losing money year after year at listing.[s11][s12][s10]
- **2021-01** 设立行歌科技进军自动驾驶芯片 / Sets up Xingge for auto-driving chips
  - 寒武纪全资设立行歌科技，专攻自动驾驶芯片，上汽集团、蔚来、宁德时代、博世等产业巨头参与投资。行歌此后推出L2+级SD5223（16TOPS）与L4级SD5226（7纳米、400TOPS以上），构成“云边端车”全栈版图的第四块拼图[s41][s30]。
  - EN: Cambricon established the wholly owned subsidiary Xingge Technology, focused on autonomous-driving chips, with industrial giants including SAIC Motor, NIO, CATL and Bosch among its investors. Xingge later launched the L2+ SD5223 (16 TOPS) and the L4-class SD5226 (7-nanometer, over 400 TOPS), forming the fourth piece of the puzzle in its full-stack "cloud-edge-device-car" map.[s41][s30]
- **2021-01-21** 发布思元290芯片与玄思1000加速器 / Launches MLU290 chip and XuanSi 1000
  - 寒武纪发布思元290智能芯片及玄思1000智能加速器：思元290采用台积电7纳米工艺、460亿晶体管，INT4算力1024TOPS、INT8算力512TOPS，对标英伟达V100/A100；玄思1000在2U空间集成四颗290，算力4.1 PetaOPS[s23][s30]。
  - EN: Cambricon released the MLU290 intelligent chip and the XuanSi 1000 intelligent accelerator: the MLU290 used TSMC's 7-nanometer process with 46 billion transistors, delivering 1,024 TOPS of INT4 compute and 512 TOPS of INT8, benchmarked against Nvidia's V100 and A100; the XuanSi 1000 integrated four MLU290 chips in a 2U space for 4.1 PetaOPS of compute.[s23][s30]
- **2021-11-03** 发布第三代云端芯片思元370 / Launches third-gen MLU370 chiplet chip
  - 寒武纪发布思元370云端AI芯片，为国内首颗采用chiplet芯粒技术的AI芯片，7纳米工艺、集成390亿晶体管，INT8算力256TOPS，国内首次支持LPDDR5内存，兼顾训练与推理负载[s24][s30]。
  - EN: Cambricon released the MLU370 cloud AI chip, China's first AI chip to use chiplet technology: a 7-nanometer process with 39 billion transistors, 256 TOPS of INT8 compute, and the country's first support for LPDDR5 memory, balancing both training and inference workloads.[s24][s30]
- **2022-03-14** CTO梁军因“与公司存在分歧”离职 / CTO Liang Jun departs over disagreements
  - 寒武纪公告核心技术人员、CTO梁军于2月10日因与公司存在分歧离职。梁军曾任华为海思麒麟SoC总架构师，2017年加入寒武纪，是寒武纪IP进入麒麟970的关键牵线人，并主导思元290/370/590架构。公告次日股价大跌约18%[s29][s25][s28]。
  - EN: Cambricon announced that Liang Jun, its core technical officer and CTO, had left on February 10 because of disagreements with the company. A former chief architect of Huawei HiSilicon's Kirin SoCs, he had joined Cambricon in 2017 as the key matchmaker for bringing Cambricon's IP into the Kirin 970, and led the architecture of the MLU290/370/590. On the trading day after the announcement the share price fell about 18%.[s29][s25][s28]
- **2022-04-27** 股价触及46.59元历史冰点 / Share price hits all-time low of 46.59 yuan
  - 受连续亏损、核心高管离职及市场低迷情绪拖累，寒武纪股价盘中触及46.59元，较64.39元的发行价深度破发，较2020年盘中高点297.77元跌去逾八成，创投股东减持不断，为公司上市以来最黑暗时刻[s38][s37]。
  - EN: Dragged down by consecutive losses, the departure of a core executive and depressed market sentiment, Cambricon's share price touched 46.59 yuan intraday, deeply below the 64.39-yuan issue price and down more than 80% from its 2020 intraday high of 297.77 yuan; venture shareholders kept reducing holdings, in the darkest moment since the company's listing.[s38][s37]
- **2022-12-16** 被美国列入实体清单及FDPR名单 / Placed on U.S. Entity List and FDPR
  - 美国商务部BIS将寒武纪等36家中国实体列入实体清单，寒武纪及部分子公司等21家实体另被纳入外国直接产品规则（FDPR）名单，先进制程流片与供应链受限。消息当日公司股价盘中跌超8%[s14][s37]。
  - EN: The U.S. Commerce Department's BIS added Cambricon and 35 other Chinese entities to the Entity List, with Cambricon, some subsidiaries and 20 other entities also placed on the Foreign Direct Product Rule (FDPR) list, constraining leading-process tape-outs and the supply chain. On the day of the news the share price fell more than 8% intraday.[s14][s37]
- **2023-04** 完成16.72亿元定增输血 / Completes 1.672bn-yuan private placement
  - 寒武纪2022年度向特定对象发行股票完成登记，以121.10元/股发行1380.6万股、募资16.72亿元，认购方以国资背景机构为主。加上IPO募资，上市三年两次直接融资合计约42.5亿元，支撑思元590等研发投入[s39][s40]。
  - EN: Cambricon completed registration of its fiscal-2022 share issuance to specific investors: 13.806 million shares at 121.10 yuan each, raising 1.672 billion yuan, with subscribers dominated by state-backed institutions. Together with the IPO proceeds, the two direct financings in the three years after listing totaled about 4.25 billion yuan, supporting R&D spending on the MLU590 and other products.[s39][s40]
- **2023** 行歌科技裁员近半，车载业务收缩 / Xingge cuts nearly half its staff
  - 据甲子光年报道，寒武纪车载子公司行歌科技裁员近半，高阶自动驾驶芯片SD5226等项目节奏放缓，集团资源集中投向云端大模型芯片主业。同年公司营收7.09亿元、归母净利润亏损8.48亿元，为连续第七年亏损[s41][s43]。
  - EN: According to Jazzy Gravity, Cambricon's automotive subsidiary Xingge Technology cut nearly half its workforce and slowed the cadence of high-end projects such as the L4 SD5226, with group resources concentrated on the cloud large-model-chip main business. In the same year the company posted revenue of 709 million yuan and a net loss attributable to shareholders of 848 million yuan—its seventh consecutive year of losses.[s41][s43]
- **2023-12** 思元590云端芯片正式上市 / MLU590 cloud chip formally launched
  - 寒武纪新一代云端芯片思元590（2022年WAIC首披露MLUarch05架构）于2023年底上市，7纳米工艺，综合性能约为英伟达A100八成，面向大模型训练与推理，随后完成DeepSeek、通义千问、智谱GLM、月之暗面Kimi等主流大模型适配[s42][s51][s30]。
  - EN: Cambricon's new-generation cloud chip, the MLU590 (whose MLUarch05 architecture was first disclosed at WAIC 2022), went on sale at the end of 2023. On a 7-nanometer process, with overall performance of roughly 80% of the Nvidia A100 and designed for large-model training and inference, it subsequently completed adaptation to mainstream large models including DeepSeek, Qwen, Zhipu GLM and Moonshot Kimi.[s42][s51][s30]
- **2024-11-20** 乌镇世界互联网大会演讲：最伟大的技术变革时刻 / Wuzhen speech: greatest technological change
  - 陈天石受邀在2024年世界互联网大会乌镇峰会主论坛发言，他表示：“我们已经身处有史以来最伟大的技术变革时刻。人工智能不仅是技术进步的工具，它更有潜力成为增进人类福祉的强大力量。”这是他少有的高规格公开演讲之一[s48]。
  - EN: Chen Tianshi was invited to speak at the main forum of the 2024 World Internet Conference Wuzhen Summit, where he said: "We are already living through the greatest moment of technological change in history. Artificial intelligence is not merely a tool for technological progress; it has the potential to become a powerful force for the improvement of human welfare." It was one of his rare high-profile public speeches.[s48]
- **2024-12** 2024年第四季度首次单季盈利 / First quarterly profit in Q4 2024
  - 寒武纪2024年第四季度实现营业收入9.89亿元、归母净利润2.72亿元，为公司2020年上市以来首次实现单季度扭亏为盈，连续多年亏损的业绩周期迎来明确拐点；背后是思元590等云端AI芯片在大模型客户处开始规模放量[s17][s32]。
  - EN: In the fourth quarter of 2024, Cambricon achieved revenue of 989 million yuan and net profit attributable to shareholders of 272 million yuan—the company's first single-quarter profit since its 2020 listing, marking a clear inflection point after years of losses; behind it lay cloud AI chips such as the MLU590 beginning to ship in volume at large-model customers.[s17][s32]
- **2025-08-15** 49.8亿元定增过会，科创板AI芯片最大单笔 / 4.98bn-yuan placement approved by STAR
  - 寒武纪向特定对象发行股票申请获科创板上市委会议通过，拟募资不超过49.8亿元（注册稿调整为不超过39.85亿元），投向先进芯片平台20.54亿元、软件平台14.52亿元及补流4.79亿元，为重注大模型算力基建[s54]。
  - EN: Cambricon's application for a share issuance to specific investors was approved by the STAR listing committee: it planned to raise no more than 4.98 billion yuan (adjusted to no more than 3.985 billion yuan in the registered version), allocated between an advanced-chip platform (2.054 billion), a software platform (1.452 billion) and working capital (479 million)—a heavy bet on large-model compute infrastructure.[s54]
- **2025-08-28** 股价收盘登顶A股“股王”超越茅台 / Closes as A-share's most expensive stock
  - 8月27日寒武纪盘中触及1464.98元首次超过贵州茅台股价（“半日游”），8月28日收报1587.91元正式收盘登顶A股第一高价股，次日市值达6643亿元；陈天石身家突破1700亿元[s19][s20][s21]。
  - EN: On August 27 Cambricon touched 1,464.98 yuan intraday, surpassing Kweichow Moutai's share price for the first time (a "half-day reign"); on August 28 it closed at 1,587.91 yuan, formally taking the closing-price crown as the A-share market's most expensive stock, with market value reaching 664.3 billion yuan the next day. Chen Tianshi's fortune broke through 170 billion yuan.[s19][s20][s21]
- **2026-03** 2025年报：营收64.97亿元，上市首年盈利 / 2025 annual report: 6.497bn revenue, first annual profit
  - 寒武纪披露2025年年报：全年营收64.97亿元、同比增长453.21%，归母净利润20.59亿元、扣非17.70亿元，为2020年上市以来首个年度盈利；云端收入64.76亿元占比超99%，毛利率55.15%，研发投入11.69亿元，并首次分红6.32亿元[s16][s15][s17]。
  - EN: Cambricon disclosed its 2025 annual report: full-year revenue of 6.497 billion yuan, up 453.21% year on year; net profit attributable to shareholders of 2.059 billion yuan and net profit after deductions of 1.770 billion—the first annual profit since the 2020 listing. Cloud revenue was 6.476 billion yuan, over 99% of the total; gross margin was 55.15%, R&D spending 1.169 billion yuan, and the company paid its first dividend of 632 million yuan.[s16][s15][s17]
- **2026-06-30** 市值盘中突破1万亿元，科创板首只万亿股 / Market value tops 1 trillion yuan intraday
  - 寒武纪股价盘中最高触及1607元、总市值最高达10040亿元，成为科创板历史上首只市值突破一万亿元的股票。此前5月6日股价曾盘中最高1966元、市值峰值8296亿元；2026年一季度公司营收28.85亿元、净利10.13亿元[s34][s35][s18]。
  - EN: Cambricon's share price touched an intraday peak of 1,607 yuan and total market value a high of 1.004 trillion yuan, making it the first stock in STAR Market history to break 1 trillion yuan in market value. Earlier, on May 6, the share price had reached an intraday high of 1,966 yuan with a peak market value of 829.6 billion yuan; in the first quarter of 2026 the company posted revenue of 2.885 billion yuan and net profit of 1.013 billion yuan.[s34][s35][s18]
- **2026-08** 2026年半年报：营收净利双翻倍 / 2026 interim report: revenue and profit double
  - 寒武纪2026年上半年实现营业收入59.96亿元、同比增长108.13%，归母净利润23.11亿元、同比增长122.61%，扣非净利润21.66亿元，毛利率55.25%，半年净利润已超2025年全年。华鑫证券点评称国产算力需求共振打开成长空间[s36]。
  - EN: In the first half of 2026, Cambricon achieved revenue of 5.996 billion yuan, up 108.13% year on year; net profit attributable to shareholders of 2.311 billion yuan, up 122.61%; net profit after deducting non-recurring items of 2.166 billion yuan; and gross margin of 55.25%, with half-year net profit already exceeding the whole of 2025. China Fortune Securities commented that resonating demand for domestic compute is opening the company's growth space.[s36]

## 常问问答 FAQ

**Q1: 陈天石和哥哥陈云霁是什么关系？兄弟俩在寒武纪如何分工？**

A: 陈天石与陈云霁是亲兄弟，哥哥陈云霁1983年生，长陈天石两岁，两人先后从南昌考入中国科学技术大学少年班：陈云霁1997年14岁入学，陈天石2001年16岁入学[s2][s52]。陈云霁在中科院计算所师从“龙芯之父”胡伟武，25岁出任八核龙芯3号主架构师，是国产通用处理器的年轻领军者；陈天石2010年获中科大计算机博士学位后也进入计算所，师从陈国良院士，主攻神经网络算法[s2][s4][s5]。2008年前后，兄弟俩在计算所一间约30平方米的“小黑屋”里联手启动AI芯片研究，合作完成DianNao、DaDianNao、PuDianNao、ShiDianNao和Cambricon指令集等系列论文，两获体系结构顶级会议最佳论文[s1][s44][s46]。2016年创办寒武纪时，两人做了明确分工：陈天石离岗下海，担任公司董事长兼CEO，负责商业化与经营；陈云霁留在中科院计算所继续从事基础研究，担任智能处理器方向学术领军者，不直接持有寒武纪股份，在学术源头与公司保持技术联动[s2][s4]。媒体因此将两人称为为AI装上中国芯的“双子星”[s4]。

**Q1 (EN): What is the relationship between Chen Tianshi and his brother Chen Yunji, and how did the two divide their roles at Cambricon?**

A: Chen Tianshi and Chen Yunji are biological brothers. The elder brother Chen Yunji was born in 1983, two years before Chen Tianshi, and both entered the Special Class for the Gifted Young at the University of Science and Technology of China from Nanchang: Chen Yunji enrolled at fourteen in 1997, and Chen Tianshi at sixteen in 2001.[s2][s52] Chen Yunji studied at the Institute of Computing Technology under Hu Weiwu, the "father of Loongson," and became chief architect of the eight-core Loongson 3 at twenty-five, a young leader of China's general-purpose processor program; Chen Tianshi also joined the ICT after taking his USTC doctorate in computer science in 2010, studying under Academician Chen Guoliang and focusing on neural-network algorithms.[s2][s4][s5] Around 2008, in a roughly thirty-square-meter "little black room" at the ICT, the brothers joined forces to launch AI-chip research, co-authoring the DianNao, DaDianNao, PuDianNao and ShiDianNao papers and the Cambricon instruction-set series, and twice winning best-paper awards at the top computer-architecture conferences.[s1][s44][s46] When Cambricon was founded in 2016, the two made a clear division of labor: Chen Tianshi left his post to go into business as the company's chairman and CEO, responsible for commercialization and operations; Chen Yunji stayed at the ICT to continue basic research, serving as an academic leader in the intelligent-processor direction, holding no direct equity in Cambricon while maintaining technical linkage with the company at the academic source.[s2][s4] The media have accordingly called the pair the "twin stars" who fitted a Chinese chip into artificial intelligence.[s4]

**Q2: DianNao系列论文为什么被认为是深度学习处理器的开山之作？**

A: 在DianNao之前，深度学习主要运行在CPU和GPU上，通用芯片把绝大部分晶体管消耗在与智能计算无关的逻辑上，能效低下。2014年3月，陈天石作为第一作者、与法国学者Olivier Temam及陈云霁合作的DianNao论文获国际体系结构顶级会议ASPLOS最佳论文，这是中国大陆、也是亚洲机构首次获此荣誉；论文提出的DianNao架构在仅3.02平方毫米、485毫瓦下实现每秒4520亿次神经网络运算，能效远超同期通用处理器[s44][s46]。此后团队连续发表DaDianNao（获MICRO 2014最佳论文）、PuDianNao（ASPLOS 2015）、ShiDianNao（ISCA 2015），并于2016年在ISCA发表Cambricon（DianNaoYu）神经网络专用指令集论文，由刘少礼担任第一作者，创下ISCA审稿评分纪录[s44][s45][s46]。这套成果首次在体系结构层面系统回答了“神经网络该用什么样的硬件、什么样的指令集”问题，其专用指令集与数据复用的设计思想被此后几乎所有AI芯片借鉴[s46]。2015年团队完成全球首款深度学习处理器原型芯片流片，能效达传统芯片近百倍；2018年《Science》杂志评价其为“开创性贡献”[s2][s6]。

**Q2 (EN): Why are the DianNao-series papers regarded as the founding works of the deep-learning processor?**

A: Before DianNao, deep learning ran mainly on CPUs and GPUs, general-purpose chips that spent the vast majority of their transistors on logic unrelated to intelligent computing and were therefore extremely inefficient in energy terms. In March 2014, the DianNao paper, with Chen Tianshi as first author in collaboration with the French scholar Olivier Temam and Chen Yunji, won the best-paper award at ASPLOS, a top international computer-architecture conference—the first time an institution from mainland China, and indeed from Asia, had received the honor; the DianNao architecture proposed in the paper delivered 452 billion neural-network operations per second within just 3.02 square millimeters and 485 milliwatts, with energy efficiency far beyond contemporary general-purpose processors.[s44][s46] The team then published in succession DaDianNao (best paper at MICRO 2014), PuDianNao (ASPLOS 2015) and ShiDianNao (ISCA 2015), and in 2016 published at ISCA the paper on Cambricon (DianNaoYu), the dedicated neural-network instruction set, with Liu Shaoli as first author, setting an all-time record for review scores at ISCA.[s44][s45][s46] This body of work was the first to answer systematically, at the architecture level, the questions of what hardware and what instruction set neural networks should use, and its design ideas of dedicated instructions and data reuse were borrowed by nearly every AI chip that followed.[s46] In 2015 the team completed tape-out of the world's first prototype deep-learning-processor chip, with energy efficiency nearly a hundred times that of traditional chips; in 2018 the journal Science appraised the work as a "pioneering contribution."[s2][s6]

**Q3: 寒武纪与华为的合作为何开始，又为何破裂？对公司影响有多大？**

A: 合作始于2017年：华为麒麟970在德国IFA展发布，集成寒武纪1A作为独立神经网络处理单元（NPU），这是全球首款内置NPU的手机芯片，采用台积电10纳米工艺、集成55亿个晶体管，搭载于Mate 10等十余款机型；后续麒麟980又采用双核寒武纪1H[s9][s43]。牵线人之一是2017年加入寒武纪任CTO的梁军，他曾任华为海思麒麟SoC总架构师[s25][s28]。凭借华为订单，寒武纪2017年、2018年终端IP授权收入占比分别高达98.33%和99.69%，公司也在2017年完成1亿美元A轮融资、估值10亿美元[s38][s13]。破裂源于华为自研：2018年10月华为公布达芬奇AI架构，2019年搭载自研NPU的麒麟810问世，终端逐步切换自研路线，寒武纪2019年IP收入占比骤降至15.49%[s9][s33]。对一家终端巨头而言，核心AI引擎长期依赖外部初创公司本就难以为继；对寒武纪而言，这次“分手”反而倒逼其全力转向云端芯片，思元系列由此诞生，公司从IP供应商蜕变为全栈芯片厂商[s33][s43]。

**Q3 (EN): How did Cambricon's partnership with Huawei begin, and why did it break up? How large was the impact on the company?**

A: The partnership began in 2017: Huawei's Kirin 970 was launched at the IFA show in Germany, integrating the Cambricon 1A as a standalone Neural Processing Unit (NPU)—the world's first smartphone chip with a built-in NPU, built on TSMC's 10-nanometer process with 5.5 billion transistors and deployed in more than ten models including the Mate 10; the later Kirin 980 adopted the dual-core Cambricon 1H.[s9][s43] One of the matchmakers was Liang Jun, who joined Cambricon as CTO in 2017 and had previously been chief architect of Huawei HiSilicon's Kirin SoCs.[s25][s28] On the strength of Huawei's orders, Cambricon's terminal-IP licensing revenue accounted for as much as 98.33% and 99.69% of total revenue in 2017 and 2018 respectively, and the company also completed its US$100 million Series A in 2017 at a valuation of US$1 billion.[s38][s13] The breakup stemmed from Huawei's in-house development: in October 2018 Huawei unveiled its Da Vinci AI architecture, in 2019 the Kirin 810 with a self-developed NPU appeared, and terminal products gradually switched to the in-house route, with Cambricon's 2019 IP revenue share plunging to 15.49%.[s9][s33] For a terminal giant, long-term reliance on an external startup for the core AI engine was never sustainable; for Cambricon, the "breakup" instead forced a full-scale pivot to cloud chips, from which the MLU series was born, transforming the company from an IP supplier into a full-stack chip vendor.[s33][s43]

**Q4: 寒武纪为什么被称为“AI芯片第一股”？上市后为什么长期亏损？**

A: 2020年7月20日，寒武纪（688256）登陆上交所科创板，发行价64.39元、募资25.82亿元，首日开盘涨288%、收涨229.86%，市值一度突破千亿元，是科创板首家以人工智能芯片为主业的上市公司，因此被称为“AI芯片第一股”[s11][s12]。长期亏损源于芯片行业的固有规律：高端AI芯片设计、流片与软件生态投入极大，一代7纳米芯片流片成本数以亿计，而收入要等产品被规模采用后才能兑现。2017至2023年，公司营收从784.3万元增长到7.09亿元，但归母净利润连续为负，2022年亏损一度扩大到12.57亿元，研发投入占营收比例长期居高不下[s38][s39][s37]。期间股价从2020年盘中高点297.77元跌至2022年4月46.59元的冰点，深度破发，创投股东多次减持[s38][s40]。公司靠IPO募资25.82亿元、2023年完成的16.72亿元定增等持续输血，把资金全部投入思元290、370、590的研发[s39][s40]。2024年第四季度公司首次单季盈利，2025年全年营收64.97亿元、归母净利润20.59亿元，上市五年后首次年度盈利[s16][s17]。

**Q4 (EN): Why is Cambricon called the "first AI-chip stock"? And why did it lose money for so long after listing?**

A: On July 20, 2020, Cambricon (688256) listed on the Shanghai Stock Exchange's STAR Market at an issue price of 64.39 yuan, raising 2.582 billion yuan; it opened up 288% and closed up 229.86% on its debut, with market value briefly breaking 100 billion yuan. It was the first STAR-listed company with artificial-intelligence chips as its main business, and is therefore known as the "first AI-chip stock."[s11][s12] The long losses arose from the inherent economics of the chip industry: the design, tape-out and software-ecosystem investment required for high-end AI chips is enormous—a single tape-out of a 7-nanometer chip costs hundreds of millions of yuan—while revenue arrives only after products are adopted at scale. From 2017 to 2023, the company's revenue grew from 7.843 million yuan to 709 million, but net profit attributable to shareholders was negative throughout, with the loss widening to as much as 1.257 billion yuan in 2022, and R&D spending as a share of revenue staying persistently high.[s38][s39][s37] In the meantime the share price fell from its 2020 intraday high of 297.77 yuan to the ice point of 46.59 yuan in April 2022, deeply below the issue price, and venture shareholders reduced holdings repeatedly.[s38][s40] The company kept itself funded by the IPO's 2.582 billion yuan and the 1.672-billion-yuan private placement completed in 2023, pouring all of it into the R&D of the MLU290, 370 and 590.[s39][s40] In the fourth quarter of 2024 it posted its first quarterly profit, and for full-year 2025 it reported revenue of 6.497 billion yuan and net profit attributable to shareholders of 2.059 billion yuan—its first annual profit five years after listing.[s16][s17]

**Q5: 美国实体清单对寒武纪有什么实际影响？**

A: 美国东部时间2022年12月15日（北京时间12月16日），美国商务部工业与安全局（BIS）将寒武纪等36家中国实体列入“实体清单”，寒武纪及其部分子公司等21家实体还被同时纳入外国直接产品规则（FDPR）名单，消息当日公司股价盘中跌超8%[s14][s37]。寒武纪采用Fabless模式，自身不设晶圆厂，芯片设计完成后委托代工厂流片，因此实体清单加FDPR的组合拳理论上会影响其获取先进制程产能、EDA工具及部分技术组件，直击供应链命门[s14][s33]。但此后两年的发展显示，制裁在压缩供给弹性的同时，也强化了国内客户采用国产芯片的紧迫性：公司通过供应链多源化与产能前置安排维持了出货节奏，2023年底思元590如期上市，并在2024年大模型算力紧缺中快速放量，2024年第四季度首次单季盈利[s33][s42][s17]。与此同时，英伟达高端GPU对华出口受限，黄仁勋称英伟达中国市场份额已从约95%降至近乎为零，国产芯片2025年国内市占率升至约35%至41%，制裁客观上把市场空间让渡给了寒武纪等本土厂商[s33][s32][s51]。

**Q5 (EN): What practical impact did the U.S. Entity List have on Cambricon?**

A: On December 15, 2022, U.S. Eastern Time (December 16 Beijing time), the U.S. Commerce Department's Bureau of Industry and Security (BIS) added Cambricon and 35 other Chinese entities to the Entity List, and Cambricon, some of its subsidiaries and 20 other entities were simultaneously placed on the Foreign Direct Product Rule (FDPR) list; on the day of the news, the company's share price fell more than 8% intraday.[s14][s37] Cambricon operates a fabless model, owning no fabs of its own and entrusting foundries with tape-out once designs are complete, so the combination of Entity List and FDPR could in theory affect its access to leading-process capacity, EDA tools and certain technology components, striking straight at the supply chain's lifeline.[s14][s33] Yet developments in the two years since show that, while compressing supply elasticity, the sanctions also sharpened domestic customers' urgency to adopt domestic chips: the company maintained its shipment rhythm through supply-chain multi-sourcing and front-loaded capacity arrangements, the MLU590 launched on schedule at the end of 2023 and shipped in volume during the large-model compute crunch of 2024, and the fourth quarter of 2024 brought its first quarterly profit.[s33][s42][s17] At the same time, exports of high-end Nvidia GPUs to China were restricted; Jensen Huang said Nvidia's China market share had fallen from about 95% to virtually zero, and domestic chips' share of the Chinese market rose to roughly 35% to 41% in 2025—the sanctions objectively ceded market space to domestic vendors such as Cambricon.[s33][s32][s51]

**Q6: 思元590和英伟达A100差距有多大？能替代英伟达吗？**

A: 思元590是寒武纪2023年底正式上市的新一代云端芯片，采用7纳米工艺，基于2022年世界人工智能大会上披露的MLUarch05架构，官方与多家券商研报口径显示其综合性能约为英伟达A100的八成，INT8等主流推理精度下能效表现突出[s42][s30][s51]。它的关键意义不在单项跑分，而在生态成熟度：公司已完成与DeepSeek、通义千问（Qwen）、智谱GLM、月之暗面Kimi等国内主流大模型的适配，客户可较平滑地迁移部署，配合自研NeuWare软件栈与BANG编程语言降低开发门槛[s51][s49][s33]。客观看，寒武纪与英伟达仍有代际差距：英伟达拥有最先进制程上的H100/H200/B系列产品，以及近二十年构筑的CUDA生态，这是寒武纪NeuWare需要长期追赶的高墙[s33][s32]。但在中国市场，高端英伟达GPU因出口管制无法正常供货，国产客户的选择标准从“最强”变为“可用、好用、可持续供应”，思元590由此成为大模型训练与推理国产替代清单中的主力选项之一，2025年公司云端产品线收入64.76亿元、占总营收99%以上[s16][s32]。陈天石在业绩说明会上的表态务实：持续拓展份额、加速场景落地，生态建设与硬件投入同等重要[s49][s50]。

**Q6 (EN): How big is the gap between the MLU590 and the Nvidia A100? Can it replace Nvidia?**

A: The MLU590 is Cambricon's new-generation cloud chip, formally launched at the end of 2023. Built on a 7-nanometer process and based on the MLUarch05 architecture disclosed at the 2022 World Artificial Intelligence Conference, it delivers overall performance of roughly 80% of the Nvidia A100 according to both the company and several brokerage research reports, with outstanding energy efficiency at mainstream inference precisions such as INT8.[s42][s30][s51] Its key significance lies not in any single benchmark score but in ecosystem maturity: the company has completed adaptation to China's mainstream large models, including DeepSeek, Qwen (Tongyi Qianwen), Zhipu GLM and Moonshot Kimi, customers can migrate and deploy relatively smoothly, and the in-house NeuWare software stack and BANG programming language lower the development threshold.[s51][s49][s33] Objectively, a generational gap with Nvidia remains: Nvidia has the H100/H200 and its Blackwell-generation products on the most advanced processes, together with the CUDA ecosystem built over nearly twenty years—a high wall that Cambricon's NeuWare must chase for a long time.[s33][s32] In the Chinese market, however, high-end Nvidia GPUs cannot be supplied normally because of export controls, and domestic customers' selection criterion has shifted from "the strongest" to "usable, easy to use and sustainably supplied"; the MLU590 has consequently become one of the mainstay options on the domestic-substitution list for large-model training and inference, and in 2025 the company's cloud product-line revenue was 6.476 billion yuan, more than 99% of total revenue.[s16][s32] Chen Tianshi's remarks at earnings briefings are pragmatic: keep expanding share and accelerate scenario deployment, with ecosystem building every bit as important as hardware investment.[s49][s50]

**Q7: 寒武纪为什么能在连续亏损多年后突然盈利？“卖铲人”模式指什么？**

A: 盈利拐点出现在2024年第四季度，当季营收9.89亿元、净利润2.72亿元，上市以来首次单季扭亏；2025年全年营收64.97亿元、同比增长453.21%，归母净利润20.59亿元、扣非17.70亿元，为上市首个年度盈利；2026年上半年营收59.96亿元、净利润23.11亿元，半年利润超2025年全年[s17][s16][s36]。爆发的直接原因是大模型带来的算力采购潮与英伟达供货受限形成共振，互联网大厂集中采购国产芯片，公司云端收入占比超过99%，毛利率升至55%左右[s15][s16]。“卖铲人”是陈天石为寒武纪定下的商业模式：公司只做芯片与软件栈，不做云服务、不做大模型、不与下游客户争利，像淘金热中向所有淘金者卖铲子的人，因此可以中立地服务包括字节跳动在内的全部互联网客户——字节跳动已成为公司第一大客户[s51][s33]。这种中立性正是华为昇腾之外的大客户选择寒武纪的核心原因：采购纯Fabless芯片公司的产品，不必担心业务数据与议价权落入竞争对手之手[s51][s33]。多年亏损则是这一模式的前置成本——思元100到590五代芯片的流片与NeuWare生态投入，全部在收入兑现前发生[s43][s7]。

**Q7 (EN): How did Cambricon suddenly turn profitable after years of losses? What is the "shovel seller" model?**

A: The profit inflection point came in the fourth quarter of 2024, when revenue was 989 million yuan and net profit 272 million yuan—the company's first single-quarter profit since listing; for full-year 2025, revenue was 6.497 billion yuan, up 453.21% year on year, with net profit attributable to shareholders of 2.059 billion yuan and net profit after deducting non-recurring items of 1.770 billion yuan, the first annual profit since listing; and in the first half of 2026, revenue was 5.996 billion yuan with net profit of 2.311 billion yuan, half a year's profit exceeding the whole of 2025.[s17][s16][s36] The direct cause of the explosion was the resonance between the compute-procurement boom brought by large models and restricted Nvidia supply: the big internet firms concentrated their procurement on domestic chips, the company's cloud revenue share rose above 99%, and gross margin climbed to around 55%.[s15][s16] The "shovel seller" is the business model Chen Tianshi set for Cambricon: the company makes only chips and a software stack—it runs no cloud services, builds no large models and does not compete with downstream customers for profit—like the man who sells shovels to every prospector in a gold rush, and can therefore serve all internet clients, including ByteDance, with neutrality; ByteDance has become the company's largest customer.[s51][s33] That neutrality is precisely the core reason the big customers beyond Huawei Ascend choose Cambricon: buying from a pure-play fabless chip company, they need not fear that business data and bargaining power will fall into a competitor's hands.[s51][s33] The years of losses were the up-front cost of this model—the tape-outs of the five chip generations from MLU100 to MLU590 and the NeuWare ecosystem investment all happened before revenue was realized.[s43][s7]

**Q8: 寒武纪市值一度突破一万亿元，这个估值是泡沫吗？**

A: 市值轨迹确实极端：2020年上市首日市值破千亿，2022年4月股价跌至46.59元冰点，2025年8月28日收报1587.91元收盘登顶A股“股王”，2026年5月6日盘中最高1966元、市值峰值8296亿元，6月30日盘中突破1万亿元（最高10040亿元），成为科创板首只万亿市值股票[s19][s18][s34][s35]。支撑高估值的多头逻辑有三：一是业绩兑现速度超预期，2025年营收增长453%、2026年上半年净利23.11亿元，扭亏弹性极强[s16][s36]；二是国产算力刚需下标的稀缺，A股纯正AI芯片设计公司极少，英伟达受限后国产替代空间以万亿计[s33][s51]；三是陈天石持股约28.35%，控制权稳定，2025年研发人员占比80%、研发投入占营收18%[s16][s17]。空头与监管关注的风险同样明确：第一大客户贡献过半收入、年末存货高达49.44亿元，客户集中与备货减值风险并存[s15][s17]；万亿市值对应的静态估值远超全球可比芯片公司，且互联网大厂自研芯片可能分流订单[s32][s37]。值得注意的是，公司在股价高位仍发布公告澄清网传客户与产能信息不实，2026年9月初市值回落至约6741亿元，波动本身说明市场仍在为其长期定价反复博弈[s31][s34]。

**Q8 (EN): Cambricon's market value once broke 1 trillion yuan—was that valuation a bubble?**

A: The market-value trajectory is indeed extreme: on its 2020 debut the company's market value broke 100 billion yuan; in April 2022 the share price fell to the ice point of 46.59 yuan; on August 28, 2025 it closed at 1,587.91 yuan to take the closing-price crown as the A-share market's most expensive stock; on May 6, 2026 it reached an intraday high of 1,966 yuan with a peak market value of 829.6 billion yuan; and on June 30 it broke through 1 trillion yuan intraday (peak 1.004 trillion), becoming the STAR Market's first 1-trillion-yuan stock.[s19][s18][s34][s35] The bull case for the high valuation has three strands: first, results were delivered faster than expected—revenue grew 453% in 2025 and first-half-2026 net profit was 2.311 billion yuan, with exceptionally strong loss-reversal elasticity;[s16][s36] second, scarcity of targets amid rigid demand for domestic compute—pure-play A-share AI-chip designers are few, and with Nvidia constrained the import-substitution space is measured in the trillions; [s33][s51] third, Chen Tianshi holds about 28.35%, giving stable control, and in 2025 R&D staff accounted for 80% of the workforce with R&D spending at 18% of revenue.[s16][s17] The risks that bearish investors and regulators watch are equally clear: the largest customer contributes more than half of revenue, year-end inventory was as high as 4.944 billion yuan, and customer concentration coexists with inventory-impairment risk;[s15][s17] the static valuation implied by a trillion-yuan market value far exceeds that of comparable chip companies worldwide, and the internet giants' in-house chips could divert orders.[s32][s37] Notably, even with the share price high, the company issued an announcement clarifying that online claims about customers and capacity were false, and in early September 2026 market value fell back to around 674.1 billion yuan; the volatility itself shows the market is still repeatedly bargaining over its long-term pricing.[s31][s34]

**Q9: 前CTO梁军向寒武纪索赔42.87亿元的股权诉讼是怎么回事？**

A: 梁军是寒武纪创始团队之外最重要的技术高管：他2000年加入华为，曾任海思麒麟SoC芯片总架构师，2017年加入寒武纪出任CTO，是推动寒武纪IP进入华为麒麟970的关键人物，并主导了思元290、370、590等几代云端芯片架构[s25][s28]。2022年3月14日公司公告，梁军因“与公司存在分歧”于2月10日离职；媒体复盘分歧指向路线之争——管理层希望抢抓国产替代窗口加快产品落地，梁军更倾向长期深耕底层架构，公告次日公司股价大跌约18%[s29][s25]。纠纷的焦点是限制性股票：梁军通过持股平台间接持有1152.3万股限制性股票，按其离职后股价上涨后的市价（约372元/股）估算价值约42.87亿元；2024年10月，梁军提起劳动争议诉讼，索赔股权激励损失约42.87亿元[s26][s27]。寒武纪则主张按双方协议约定，以约5.2万元对价回购相关股权；此前相关仲裁与诉讼程序中梁军已两度败诉[s26][s28]。2025年11月央广网报道案件仍在审理中，因标的额巨大，被视为国内芯片行业最大劳动争议之一，其判决将对科技公司股权激励退出机制产生示范影响[s27][s28]。

**Q9 (EN): What is the equity lawsuit in which former CTO Liang Jun claims 4.287 billion yuan from Cambricon?**

A: Liang Jun was the most important technology executive outside Cambricon's founding team: he joined Huawei in 2000, served as chief architect of HiSilicon's Kirin SoC chips, and joined Cambricon as CTO in 2017, acting as the key figure who drove Cambricon's IP into Huawei's Kirin 970 and leading the architecture of several cloud-chip generations including the MLU290, 370 and 590.[s25][s28] On March 14, 2022, the company announced that Liang had departed on February 10 "because of disagreements with the company"; media reconstructions pointed to a contest over route—management wanted to seize the import-substitution window and speed products to market, while Liang leaned toward long-term cultivation of the underlying architecture—and on the trading day after the announcement the share price fell about 18%.[s29][s25] The focus of the dispute is restricted shares: Liang held 11.523 million restricted shares indirectly through an equity platform, worth an estimated 4.287 billion yuan at the market price after the post-departure rise (roughly 372 yuan per share); in October 2024 Liang filed a labor-dispute lawsuit claiming about 4.287 billion yuan in stock-incentive losses.[s26][s27] Cambricon maintains that under the parties' agreement the relevant equity should be repurchased for consideration of about 52,000 yuan, and Liang had already lost twice in prior related arbitration and litigation proceedings.[s26][s28] In November 2025 CNR reported the case was still being heard; given its enormous amount in dispute, it is regarded as one of the largest labor disputes in China's chip industry, and its verdict will set a precedent for stock-incentive exit mechanisms at technology companies.[s27][s28]

**Q10: 客户高度集中、互联网大厂又在自研芯片，寒武纪未来最大的风险是什么？**

A: 第一类风险是客户集中。2025年年报显示，寒武纪云端产品线收入64.76亿元、占总营收99%以上，客户以大型互联网企业为主，字节跳动为第一大客户；每日经济新闻等报道明确提示客户高度集中是公司主要隐忧之一，单一客户资本开支节奏即可显著影响季度业绩[s16][s15]。第二类风险是大客户自研分流——这正是当年华为模式的重演：华为在采购寒武纪IP两年后推出自研达芬奇架构，导致寒武纪2019年IP收入占比从99.69%骤降至15.49%；如今阿里平头哥、百度昆仑芯等大厂自研芯片持续迭代，一旦成熟便可能减少外采[s38][s32]。第三类风险是供应链与制程约束：2022年底公司被列入美国实体清单及FDPR名单，先进制程产能存在不确定性[s14][s37]。第四类是估值与预期风险：万亿市值已计入大量乐观预期，而年末49.44亿元存货意味着备货减值压力，早期创投股东在股价反弹中持续减持[s15][s38][s17]。陈天石的应对思路是保持“卖铲人”的中立定位、扩展客户名单，并持续做厚软件生态提高切换成本，同时把39.85亿元定增募投资金投向新一代芯片与软件平台[s51][s50][s54]。他在业绩说明会上称，行业竞争是长期马拉松，“长跑才刚刚开始”[s49][s7]。

**Q10 (EN): With customers heavily concentrated and the internet giants building their own chips, what are Cambricon's biggest future risks?**

A: The first category of risk is customer concentration. The 2025 annual report shows Cambricon's cloud product-line revenue at 6.476 billion yuan, more than 99% of total revenue, with customers dominated by large internet enterprises and ByteDance as the largest; outlets such as the National Business Daily explicitly flag heavy customer concentration as one of the company's main worries, since a single customer's capital-expenditure cadence can materially swing quarterly results.[s16][s15] The second category is diversion by big customers' in-house chips—a reprise of the Huawei pattern: two years after buying Cambricon's IP, Huawei launched its in-house Da Vinci architecture, causing Cambricon's IP revenue share to plunge from 99.69% in 2018 to 15.49% in 2019; today the in-house chips of giants such as Alibaba's T-Head and Baidu's Kunlunxin keep iterating, and could reduce external purchasing once mature.[s38][s32] The third category is supply-chain and process constraints: at the end of 2022 the company was placed on the U.S. Entity List and FDPR list, leaving leading-process capacity subject to uncertainty.[s14][s37] The fourth is valuation and expectation risk: the trillion-yuan market value already prices in a great deal of optimism, while year-end inventory of 4.944 billion yuan means stockpiling-impairment pressure, and early venture shareholders have kept reducing holdings on rebounds.[s15][s38][s17] Chen Tianshi's response is to hold the neutral "shovel seller" position, broaden the customer list, keep thickening the software ecosystem to raise switching costs, and deploy the 3.985 billion yuan raised in the private placement toward the new-generation chip and software platforms.[s51][s50][s54] At earnings briefings he says industry competition is a long marathon: "the long race has only just begun."[s49][s7]

## 来源清单 Sources

1. [s1] 环球人物网 · 太燃了！从被美制裁到超越茅台，这对江西兄弟逆袭登顶 — https://www.globalpeople.com.cn/n4/2025/0829/c305917-21633281.html
2. [s2] 澎湃新闻 · 国产芯片寒武纪研发者陈云霁获全国五一劳动奖章，毕业于中科大少年班 — https://m.thepaper.cn/wifiKey_detail.jsp?contid=17912515
3. [s3] 中国新闻网 · 6000亿寒武纪背后的天才兄弟 — http://www.chinanews.com/m/cj/2025/09-01/10474869.shtml
4. [s4] 中国日报网 · 为AI装上中国芯的“双子星”——记寒武纪科技创始人陈云霁、陈天石 — https://chuangxin.chinadaily.com.cn/2018-03/20/content_35882398.htm
5. [s5] 央广网 · [天下英才]陈天石：为人工智能装上“中国芯” — http://m.cnr.cn/news/20170424/t20170424_523721978.html
6. [s6] 中国经济网 · 寒武纪科技陈天石：在AI“石器”时代保留火种 — http://www.ce.cn/cysc/newmain/pplm/czrw/xw/201802/05/t20180205_28049485.shtml
7. [s7] 界面新闻（甲子光年） · 对话寒武纪陈天石：通往伟大芯片公司的赛程很长 — https://m.jiemian.com/article/4121014.html
8. [s8] 每日经济新闻 · 业绩持续亏损、创投股东离场……寒武纪与天才创始人陈天石的“多事之秋” — http://m.toutiao.com/group/7284074989379240489/
9. [s9] 电子工程世界（全天候科技） · 华为联手寒武纪发布全球首款手机AI芯片，是颠覆还是噱头？ — https://news.eeworld.com.cn/qrs/article_2017090738121.html
10. [s10] 格隆汇 · 寒武纪来了！AI芯片第一股登“科”倒计时！ — https://m.gelonghui.com/p/352395
11. [s11] 证券时报网 · 寒武纪今日登陆科创板 首日开盘大涨288% — https://www.stcn.com/stock/djjd/202007/t20200720_2146989.html
12. [s12] 中国证券报 · 中一签或赚10万！寒武纪明日申购，概念股大涨 — http://m.toutiao.com/group/6846718705800315405/
13. [s13] 智东西 · 寒武纪获数亿美元B轮融资 估值达25亿美元 — https://m.zhidx.com/news/4430.html
14. [s14] 腾讯新闻（澎湃新闻） · 被美列入“实体清单”，寒武纪盘中跌超8% — https://new.qq.com/rain/a/20221216A01QM700
15. [s15] 中国经济网（每日经济新闻） · 寒武纪2025年营收大增453% 存货激增与客户集中存隐忧 — http://www.ce.cn/xwzx/gnsz/gdxw/202603/t20260313_2825102.shtml
16. [s16] 上海证券交易所 · 中科寒武纪科技股份有限公司2025年年度报告 — http://www.sse.com.cn/disclosure/listedinfo/announcement/c/new/2026-03-13/688256_20260313_GWH3.pdf
17. [s17] 长江商报（新浪财经） · 寒武纪年营收增4倍 拟分红6.3亿 牛散章建平持股市值92亿 — https://finance.sina.com.cn/wm/2026-03-16/doc-inhrcxwx8951371.shtml
18. [s18] 长江商报（新浪财经） · 寒武纪股价冲击2000元未果仍是A股之王 首季大赚逾10亿 — https://finance.sina.com.cn/roll/2026-05-07/doc-inhwzmmr8553953.shtml
19. [s19] 现代快报（腾讯新闻） · 寒武纪午后登顶A股“股王”超茅台，市值达5740亿元 — http://news.qq.com/rain/a/20250827A07P4200
20. [s20] 深圳商报（证券时报网） · 寒武纪盘中股价一度超越贵州茅台 — https://stcn.com/article/detail/3290699.html
21. [s21] 21世纪经济报道 · 1461元！寒武纪股价超越贵州茅台，85后创始人身家突破1700亿元 — http://m.toutiao.com/group/7543138540465701395/
22. [s22] 中国日报网财经 · 寒武纪推出新一代云端AI芯片思元270 — https://caijing.chinadaily.com.cn/chanye/2019-06/25/content_37484499.htm
23. [s23] 上海证券交易所 · 中科寒武纪科技股份有限公司关于发布思元290智能芯片及玄思1000智能加速器的公告 — http://www.sse.com.cn/disclosure/listedinfo/bulletin/star/c/688256_20210122_1.pdf
24. [s24] 寒武纪（今日头条） · 寒武纪发布第三代云端AI芯片思元370 — http://m.toutiao.com/group/7026919402029187621/
25. [s25] 钛媒体 · 寒武纪CTO梁军离职背后：技术与商业博弈间，大牛何去何从？ — http://m.toutiao.com/group/7076642883121447437/
26. [s26] 南方都市报·湾财社 · 5万元转让1150万股股票？寒武纪与前CTO梁军对簿公堂 — https://static.nfnews.com/content/202501/21/c10457078.html
27. [s27] 央广网 · 被前CTO起诉 寒武纪陷超40亿元劳动争议诉讼 — http://finance.cnr.cn/ycbd/20251101/t20251101_527416139.shtml
28. [s28] 蓝鲨硬科技（腾讯新闻） · 寒武纪陈天石身价超1100亿元，前CTO42亿股权纠纷始末 — http://news.qq.com/rain/a/20250814A05RUL00
29. [s29] 量子位（新浪财经） · 寒武纪高层震荡！CTO梁军因“与公司存在分歧”离职，市值蒸发 — https://finance.sina.cn/stock/ssgs/2022-03-15/detail-imcwiwss6130892.d.html
30. [s30] 证券之星 · 华鑫证券：寒武纪产业线多维发展，芯片技术创新突破居领先地位 — http://m.toutiao.com/group/7399269355042652709/
31. [s31] 证券时报网 · 寒武纪紧急公告：网传产品、客户、供应及产能预测信息均为不实信息 — https://stcn.com/article/detail/3525947.html
32. [s32] 36氪（青澄财经） · 寒武纪第一次赚钱，但已被“中国版英伟达”包围 — https://36kr.com/p/3729096279375234
33. [s33] 36氪 · 英伟达向左，寒武纪向右 — https://m.36kr.com/p/3807749556215812
34. [s34] 齐鲁晚报（新浪财经） · 科创板首只万亿市值股票来了 — https://finance.sina.com.cn/jjxw/2026-07-01/doc-inifihnu2995009.shtml
35. [s35] 北京日报（新浪财经） · 科创板首家！寒武纪市值突破万亿元 — https://finance.sina.com.cn/jjxw/2026-06-30/doc-inifczzw4943367.shtml
36. [s36] 华鑫证券 · 业绩高增彰显盈利弹性，国产算力需求共振打开成长空间——寒武纪2026年半年报点评 — https://pdf.dfcfw.com/pdf/H3_AP202609021828901452_1.pdf
37. [s37] 界面新闻 · “AI芯片第一股”寒武纪又亏了，高研发投入压顶，原始股东减持不停 — https://m.jiemian.com/article/8278493.html
38. [s38] 第一财经 · 亏损扩大、股价暴涨近200%，创投股东拟清仓减持寒武纪 — https://m.yicai.com/news/101709064.html
39. [s39] 中国经济网 · 寒武纪2022年亏损增至12.6亿 上市3年两募资共42.5亿 — http://finance.ce.cn/stock/gsgdbd/202305/05/t20230505_38532170.shtml
40. [s40] 金融界（证券之星） · 寒武纪重回千亿市值！定增机构浮盈近120% — https://stock.stockstar.com/IG2023042000025712.shtml
41. [s41] 甲子光年 · 寒武纪子公司行歌裁员，自动驾驶芯片还做吗？ — https://www.jazzyear.com/article_info.html?id=1065
42. [s42] 寒武纪官网 · 2022 WAIC：寒武纪展现全算力产品实力，用“芯”助力行业快速升级 — https://www.cambricon.com/index.php?m=content&c=index&a=show&catid=127&id=51
43. [s43] 36氪 · 连亏7年，寒武纪市值凭啥破千亿？ — https://m.36kr.com/p/2985250291359749
44. [s44] 中国科学院计算技术研究所 · Tianshi Chen 个人主页（DianNao系列论文目录） — http://novel.ict.ac.cn/tchen/
45. [s45] ACM Digital Library · Cambricon: An Instruction Set Architecture for Neural Networks (ISCA 2016) — https://dl.acm.org/doi/10.1109/ISCA.2016.42
46. [s46] 腾讯云开发者社区 · 中科院深度学习指令集DianNaoYu到底是什么？寒武纪4篇论文解读 — https://cloud.tencent.com/developer/article/1164457
47. [s47] 光明网文摘报 · “长跑才刚刚开始” — https://epaper.gmw.cn/wzb/html/2025-09/06/nw.D110000wzb_20250906_2-05.htm
48. [s48] 新京报 · 世界互联网大会乌镇峰会开幕 陈天石：我们身处有史以来最伟大的技术变革关键时期 — https://www.bjnews.com.cn/detail/1732104577129999.html
49. [s49] 上海证券报·中国证券网 · 今年能否实现百亿营收？寒武纪2025年度业绩说明会回应 — http://www.cnstock.com/commonDetail/680292
50. [s50] 每日经济新闻 · 寒武纪董事长陈天石：积极拓展市场份额，加速场景落地 — http://m.toutiao.com/group/7551412235173462569/
51. [s51] 21世纪经济报道 · 最强“卖铲人”！寒武纪靠AI芯片疯狂掘金 — http://m.toutiao.com/group/7551395324875538980/
52. [s52] 河南商报 · 中科大少年班诞生一个超级富豪，身家250亿，一度沉迷游戏 — http://m.toutiao.com/group/6851756115667976716/
53. [s53] 江南都市报（澎湃新闻） · 南昌这对80后兄弟太了不起！背后故事值得好好看看 — https://m.thepaper.cn/newsDetail_forward_8525564
54. [s54] 电子元件技术网 · 寒武纪获科创板AI芯片最大单笔定增通行证！50亿重注大模型算力基建 — http://ep.cntronics.com/market/14872
55. [s55] 与非网 · 从中科院到科创板：“中国版英伟达”市值超5200亿 — https://m.eefocus.com/article/1881207.html

---
本档案机器版由亚洲封面人物官网自动生成：https://coverfigure.com ｜ 永久档案：https://coverfigure.com/acf/ACF-00-00105/geju ｜ 生成时间 Generated：2026-09-11 00:42