文 | 华商韬略
10月21日，特斯拉在三季度财报里宣布：Optimus人形机器人已在加州弗里蒙特工厂的量产线上开工。这条产线原本属于Model S和Model X，今年7月，特斯拉只用46天就把旧线拆了个干净，然后换上了Optimus的生产线。
几乎同期，英伟达创始人黄仁勋在柏林GTC大会上高喊：物理AI，人形机器人，是下一个伟大的产业；王兴兴说自己犯的最大错误就是不够相信AI，并且提前了人形机器人的规模化应用时间表，市值被腰斩的宇树科技则开始了股价反攻。
如果把过去三年的大模型比作一个人，它更像一位被困在电脑里的“无所不知”：能创作、能编程、能回答你各种问题，陪你聊人生，但却从没走出电脑，哪怕为你亲手拿起过一只杯子。物理AI要补的，就是这一课——
让“无所不知”变成“无所不能”。
黄仁勋今年1月在CES上的表述是：AI正在越过理解数字信息的阶段，开始去理解重力、摩擦、惯性和因果。翻译成大白话，物理AI，就是让AI从“会说”变成“会做”。它要从屏幕里走出来，走进工厂、仓库、街道和家庭。
物理Al的核心是机器人。
会自己开的车，是带轮子的机器人；会自己飞的无人机，是带翅膀的机器人；产线上会自己找零件、拧螺丝的机械臂，是长在工位上的机器人……
人形机器人只是其中最难、也最被寄予厚望的那一个。原因是：楼梯、门把手、工具、工位，人类的世界，全都是按人的身材设计的，谁能长成人的样子，谁就能更无缝、直接地接手人的一切工作。
物理AI瞄准的，是人类劳动力本身这个最大的市场。
这是一笔大账。
全球制造业招工越来越难，人口老龄化让养老、护理、物流的人手缺口一年比一年大。花旗曾估算，到2050年全球人形机器人数量可达6.48亿台，相当于今天全球汽车保有量的四成多。
让一台人形机器人工作，就像让一个人工作，看似简单，背后却是一连串复杂的拆解。
假设你对它说：帮我从冰箱里拿一瓶可乐。
它先要用“眼睛”——摄像头和3D视觉传感器，看清冰箱在哪，门把手在哪，可乐在第几层。
然后是“大脑”。多模态大模型把这句话拆成一串动作：走过去、拉开门、找到可乐、拿起来、关上门、递过来。行业里管这类模型叫VLA，也就是“视觉-语言-动作”模型。
接着是“小脑”，也就是运动控制系统。两条腿走路，本质上是每一步都在摔倒的边缘，它要在毫秒之间调整重心，保证身体不倒。
最后是“肌肉和筋骨”。一台人形机器人通常有四十个左右的关节，每个关节里塞着电机、减速器、丝杠和传感器。手上还得有触觉和力觉：抓可乐瓶可以稍微用力，抓钢球得特别用力，拿鸡蛋就得收着劲。
听起来，每一步都不复杂，但对机器人来说，每一步都是一道坎。
人工智能界有个著名的“莫拉维克悖论”：让电脑下赢围棋世界冠军相对容易，让机器人像三岁孩子一样叠好一件衣服，却难上加难。
难，就难在数据。
语言大模型之所以突飞猛进，是因为人类在互联网上攒下了几十年的文字。但物理世界，没有现成的“互联网”。机器人的每一次抓取、每一次摔倒，都得在真实世界里一遍遍试出来，又慢又贵。
破局的办法有两条。一条是让真机大批量下场干活，边干边采数据；另一条，是在虚拟世界里“开驾校”。
后者就像飞行员先进模拟舱：先在虚拟世界里摔一百万次，再去真实世界迈第一步。英伟达去年3月公布过一组数据：用它的工具，11个小时就能生成78万条机器人合成动作轨迹，相当于人类示范6500个小时的数据量。
即便如此，今天的物理AI离“通用”仍有距离。
宇树科技创始人王兴兴8月在世界机器人大会上就讲过：在固定场景里，经过充分的数据采集和专项训练，机器人完成任务的成功率几乎能到100%；可只要换一件物品，或者把环境稍微挪一挪，成功率就会大幅下滑。
王兴兴给具身智能的“ChatGPT时刻”下的标准是：把一台机器人带到任意一个陌生环境，比如你家里，它基本能完成80%左右的任务。到那一天，产业才算真正走到爆发的临界点。当时，他对这个时间的预期是：快的话两三年，慢的话五年甚至十年。
但10月12日，在第二十六届中国国际工业博览会开幕式的圆桌对话上，王兴兴把这个预判往前提了。他认为突飞猛进的模型能力很快就会将壁垒击穿，这个临界点快的话一两年内就可能到来。他甚至放话：明年GPT-7一出来，可能直接把具身智能的问题打穿。
换句话说，几个月前，物理AI还是比较遥远的未来，但现在它已跑进冲刺前的最后一段直道；而它要改造的制造、物流和服务业，恰恰是中国最熟悉的主场。
全球物理AI的牌桌上，玩家很多，打法截然不同。
马斯克是最高调的布道者，也是下注最狠的那一个。
早在2024年6月的特斯拉股东大会上，他就算过一笔大账：全世界未来每年可能要造10亿台人形机器人，特斯拉的份额可能远超10%；规模足够大时，每台成本可降到1万美元左右，卖2万美元。
他把Optimus称作特斯拉“史上最大的产品”，还畅想过：有了它，每个人都能拥有一位顶尖的外科医生。
但马斯克的时间表，向来跑得比工厂快。他曾预计2025年会有上千台Optimus进厂干活，到了今年1月，又改口：Optimus还没有以实质性的方式在工厂里投入使用。然而，马斯克却是真把家当押上去了。
他停掉了Model S和Model X，把弗里蒙特的产线改造成Optimus工厂。今年三季度，特斯拉研发费用同比跳涨69%至27.6亿美元，净利润同比下降43%，钱主要砸向了特斯拉所说的超级智能、机器人等新业务。
早在7月1日，马斯克晒出和Optimus生产团队的合影时就打了预防针：量产初期会极其缓慢，因为一切都是新的，“这不像造汽车”。
10月21日，Optimus终于在量产线上开工。马斯克说，这条设计年产能100万台的产线，大部分设备已经运抵；明年4月前后，Optimus还要换上自研的AI5芯片。他的判断依旧宏大：
Optimus将成为地球上最大的产品，未来机器人的数量会超过人类。
黄仁勋走的是另一条路：不造机器人，要让全世界的机器人都长在英伟达的平台上。
英伟达的打法，是一整套从训练到上身的“全家桶”：Cosmos世界模型负责理解和生成物理世界，Isaac仿真平台负责“开驾校”，GR00T是给人形机器人准备的通用大脑，Jetson Thor则是装在机器人身上的芯片。
今年3月的GTC大会上，ABB、发那科、库卡、安川这些工业机器人巨头，加上美国的Figure、中国的智元，都出现在英伟达的合作伙伴名单里。
5月底的台北GTC大会上，英伟达发布Cosmos 3，号称全球首个完全开源的全模态模型，能把物理AI的训练和评估周期从数月压缩到数天。
10月21日的柏林GTC大会上，陪黄仁勋站在台上的，是几只25厘米高、鸭子模样的小机器人。他让它们翻跟头，又让它们“睡觉”，偏偏有一只背过身去，黄仁勋只好叮嘱它：注意力集中。台下一阵大笑。
逗完鸭子，他端出的是正经家伙：一款由德国Agile Robots打造、基于Jetson Thor和GR00T的参考人形机器人，Agile Robots计划明年就造1000到2000台。黄仁勋当天的判断是：机器人的ChatGPT时刻，正在快速逼近。
黄仁勋要做的，是把在数据中心赚得盆满钵满的那套“芯片+软件+生态”，原样搬进物理世界：不管最后谁家的机器人胜出，它们的大脑都得在英伟达的平台上长大。
硅谷的其他玩家，则把最多的钱押在了“大脑”上。
今年1月的CES上，谷歌DeepMind宣布与波士顿动力合作，把Gemini Robotics模型装进新一代Atlas人形机器人，按现代汽车的计划，新Atlas今年就进入其美国佐治亚州的工厂。
颇有戏剧性的是，2017年谷歌嫌机器人投入大、回报远，把波士顿动力卖给了软银；兜兜转转九年，谷歌又以“大脑供应商”的身份回来了。
创业公司的估值更是一路狂飙。Figure AI去年9月估值已达390亿美元；由前谷歌DeepMind研究员与斯坦福、伯克利学者联手创办、只做机器人通用模型的Physical Intelligence，今年3月被曝正以约110亿美元的估值洽谈新一轮约10亿美元融资，四个月估值翻倍。联合创始人谢尔盖·莱文给公司的定位，就是做“机器人的ChatGPT”。
9月28日，AMD宣布以约82亿美元全股票收购李飞飞创办的World Labs，这是AMD史上仅次于赛灵思的第二大收购。李飞飞的World Labs成立于2024年，做的是“空间智能”：从文字、图片、视频生成和模拟可交互的3D世界，被视为物理AI的上游。苏姿丰的说法是：要做下一代AI的计算平台，必须深刻理解模型如何演进。李飞飞本人将出任AMD执行副总裁兼首席科学家。这笔交易，也把芯片巨头的竞争，从卖卡打到了抢“世界模型”。
英伟达把Cosmos开源出去，让整个行业在它的平台上做物理AI；AMD则干脆把一个顶级实验室买回家，用模型反过来定义芯片。
然而，硅谷拥有最聪明的脑，最贵的估值，但真正能大规模量产的身体，却不在这里，而是在中国。中国的机器人或许大脑不够聪明，资本也更有限，但在量产上却真正做到了遥遥领先。
中国以人形机器人为核心的物理AI，其基本格局是：本体在跑量，车企在转身，巨头在搭台。
先看本体新势力。
据市场研究机构Smart Analytics Global统计，今年上半年全球人形机器人出货约1.91万台，中国厂商拿下了97%以上。其中，智元出货约8400台，全球份额44%，排名第一；宇树约5900台，份额31%。两家合计，占了全球四分之三。
智元的创始组合颇有看头：董事长邓泰华是华为前副总裁；总裁彭志辉，就是B站上的“稚晖君”，当年以“天才少年”身份加入华为。两个华为人，三年时间，把公司从首年营收30万元做到了2025年的10亿元以上。
6月28日，智元第15000台G2机器人下线，距离第10000台不到三个月。这台机器人没进展厅，而是直接被送往龙旗科技南昌工厂的平板电脑产线，和另外8台G2一起干质检。此前连续6天的直播里，这8台机器人完成了1.75万多次质检，成功率99.99%。
宇树是另一种气质。王兴兴从四足机器人起家，2016年创业，一做就是十年。
今年5月底，《每日经济新闻》记者走进杭州滨江的宇树总部：门口，研发人员正在调试一台四足机器人；楼顶上，一台轮足机器狗在无人巡逻；一楼展厅已经搬空，销售团队正往一栋还没装修好的新楼里挪。公司长得比房子还快。
智能汽车和人形机器人，在传感器、芯片、电机、电池、算法等领域，底层高度同源，据盖世汽车研究院测算，两者供应链重合度达到六到七成。中国新能源汽车产业链稍一延伸，就是机器人产业链。而事实上，中国车企也同样在积极布局机器人。
其中，转身最积极彻底的是小鹏。何小鹏在6月10日的内部信里说，这是小鹏从“智能汽车公司”向“物理AI公司”转型的关键一步。8月，小鹏机器人业务完成超9亿美元首轮融资，投后估值超63亿美元，腾讯、阿里以战略投资者身份入局。小鹏之外，理想的李想，则在3月放话：2026年是理想进化为“具身智能”企业的关键一年。
已经在新能源汽车上超越特斯拉的比亚迪，对机器人的理解还是把重点落到制造上。其执行副总裁李柯说：机器人的竞争，拼的是制造能力、软件能力和硬件能力，而汽车上积累的AI能力可以移植到机器人上。
中国科技巨头也同样在加大对机器人的布局。
华为云的具身智能平台CloudRobo去年6月发布、今年6月升级并开启公测，从数据、模型、仿真到部署，提供一整条流水线，已与30多家行业伙伴、20多家本体厂商合作。小米的雷军，则把2026年定义为“具身智能大模型元年”。
但中国能在人形机器人跑得快，靠的不只是一两家明星公司，而是产业链和生态。据工信部7月披露，目前中国人形机器人整机型号已有400余款，占全球一半以上；四足机器人全球销量份额接近70%。电机、减速器、电池、传感器、结构件，几乎每一个环节都能在长三角、珠三角一两个小时车程内找到供应商。
在深圳南山的机器人谷，大约10公里的留仙大道，就是一条完整的机器人关键技术研发和零部件制造体系，从大小脑模型等算法软件，触觉传感器、激光雷达等感知系统，到伺服电机、动力模组等执行系统，再到机械臂、灵巧手等核心部件，都能在机器人谷找到优质供应商，实现快速交付。
在这里，“上午出图、下午打样、次日迭代”是常态。一款灵巧手用的触觉传感器，过去靠进口要10万元，如今国产的199元起。今年1到8月，南山区机器人货运订单同比增长94.4%。
美国在讲最宏大的故事，造最聪明的大脑；中国在造最多的身体，跑最快的量产，进最真实的场景。中美合起来，整个人形机器人就能更快速地跑起来，AI也就更快走出屏幕，走向各行各业，家家户户。
产业热了，钱就来了。
据行业机构统计，今年上半年，中国人形机器人整机产量突破4万台，二季度环比增长50%；赛道融资总额达到1041亿元，已经超过2025年全年。工信部预计，今年全年整机产量有望突破10万台。花旗更是预测，到2050年，全球人形机器人市场规模将达到7万亿美元。
需求也在从“表演”走向“上岗”。工信部、国务院国资委6月联合印发通知，开展2026年度人形机器人与具身智能实景实训专项行动，要求到2026年底，人形机器人等重点产品在一批代表性场景中率先完成应用验证和常态部署，开启“作业模式”。上海10月12日发布的《上海市加快发展新一代智能制造行动》更是明确，要推动8000台人形机器人进工厂。
IDC的数据显示，今年上半年全球人形机器人出货约2.5万台，同比增长432%；其中科研教育、展览展示这类“不干活”场景的占比，已从去年全年的83.8%降到69%。这意味着，机器人正一台台从展台走向工位。
但越是热闹，越要冷静地问一句：物理AI的钱，最后会落进谁的口袋？
回看新能源汽车的历史，有一条规律值得记住：整车杀得头破血流时，最先赚到钱的往往是核心部件。宁德时代就是最好的注脚。物理AI大概率会重演这一幕。
整机格局远未定型，今天的冠军明天未必还在牌桌上；但不管谁家的机器人胜出，都绕不开关节、电机、减速器、丝杠和传感器。
物理AI接下来将产生的产业与投资机会，重点是三条线：
第一条线：先看“身体”，再看“大脑”。
以A股公司为例，今天真正被订单验证过的物理AI公司，大多集中在执行器环节。原因很现实：大脑的价值，眼下主要被英伟达和尚未上市的模型公司拿走；中国握着全球最强的制造底座，所以中国的机会首先藏在物理AI的身体里。
绿的谐波就是这条线上的代表。谐波减速器曾长期被日本企业垄断，绿的谐波率先撕开了口子，如今又把行星滚柱丝杠做到了规模化量产，全面切入具身智能的核心供应链。
三花智控、拓普集团代表着另一类玩家：汽车零部件巨头带着规模制造和降本能力切入机器人执行器，被市场视为特斯拉供应链的核心选手。
此外，双环传动的精密齿轮与减速器、恒立液压的丝杠、鸣志电器的空心杯电机、奥比中光的3D视觉、柯力传感的力传感器，都在各自的细分环节握有硬功夫。
第二条线：看主业能不能托底。
机器人放量是一场马拉松，中途必有颠簸。能熬到终点的，往往是主业现金流足够厚的公司。
汇川技术是典型。今年上半年营收246.75亿元，通用伺服系统国内市占率升至35.2%，稳居第一。人形机器人对它而言不是转型，而是伺服、电机、控制器能力的自然延伸。这类公司弹性未必最大，却最不怕题材退潮。
第三条线：看订单，不看概念。
雷赛智能就是一个用财报说话的样本。它聚焦关节模组和灵巧手两大品类，上半年营收12.47亿元，同比增长39.92%；归母净利润1.94亿元，同比增长62.79%，产品已拿到国内主流机器人厂商的批量订单。
整机层面，宇树科技已是A股最纯正的人形机器人标的，也是整个板块的估值锚；智元则通过入主上纬新材，在A股有了自己的资本平台。
反过来，有三类公司需要格外小心。
一类是“合作”“参股”“战略协议”的蹭热度者。6月8日，家居卖场居然智家因与优必选的合作被炒成物理AI概念，股价涨停；当晚公司就发公告澄清：那只是一份框架协议，没有实质订单，也没给收入贡献过一分钱。
一类是估值远远跑在业绩前面的。截至7月初，绿的谐波今年以来股价累计涨超150%，而它一季度的营收只有1.40亿元，市场已经提前买下了好几年的增长。
一类是造蓝图提概念，没有金刚钻却要揽瓷器活儿的。比如，一个既没有强大资本，也没有顶尖技术人才的公司，声称自己要做机器的大脑，甚至与英伟达竞争。
更深一层的考验在价格。机器人一旦放量，零部件单价必然被持续压低，就像当年的光伏和动力电池。最后的赢家，要么是真有独一无二核心技术壁垒，要么是能把成本打到最低、把规模做到最大的那一个。
说到底，投资物理AI，不必去赌谁能造出最好的机器人，而要找谁都离不开的东西。
这场变革来得比多数人预想得快。王兴兴在今年工博会上的一句自我反思，值得每个盯着这条赛道的人记住：
“过去十几年我犯的最大的错误，就是对AI相信得还不太够。”
参考资料：
[1] 新华社：工信部在2026世界人工智能大会新闻发布会上表示，2026年我国人形机器人整机产量有望突破10万台
[2] 特斯拉2026年第三季度财报及业绩电话会（新浪科技、每日经济新闻相关报道）
[3] 英伟达GTC 2026圣何塞、台北、柏林大会官方发布及科创板日报相关报道
[4] Smart Analytics Global、IDC：2026年上半年全球人形机器人出货数据
[5]第一财经、证券时报：王兴兴在2026世界机器人大会、第二十六届中国国际工业博览会的演讲报道
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, China, China.
On 21 October, Tesla announced in a three-quarter financial paper that Optimus human robots had begun work on the production line at the Fremont factory in California. The line originally belonged to Model S and Model X. In July of this year, Tesla removed the old line for 46 days and replaced it with Optimus.
In almost the same period, the founder of Young Weidar, Huang In-hoon, at the GTC conference in Berlin, shouted that physics AI, the human robot, was the next great industry; Wang Hing Xing said that his biggest mistake was not to trust AI, and that he had advanced the scale application schedule for human robots, and that the market-valued Oki technology had begun a stock-price fight.
If you compare LLM over the past three years to a person, it's more like a man who's been trapped in a computer's "intelligence": able to create, program, able to answer your questions, able to talk to you about life, but never walked out of the computer, even with a glass for you. That's what physics AI. needs to do.
Let “everything knows” become “everything”.
Hwang In-hoon's statement on CES in January this year is that AI is moving beyond the stage of understanding digital information to begin to understand gravity, friction, inertia and causality. Translating into big white, physics AI is going to change it from "can say" to "can do." It comes out of the screen and goes into factories, warehouses, streets and homes.
The core of physics Al is robotics.
The car that drives itself is a robot with wheels; the drone that flies by itself is a robot with wings; and the machine arm on the line of production that looks for parts, screws, is a robot that grows on a working position...
Man-shaped robots are only the hardest of them, and the least of them. The reason for this is that the stairs, the doorknobs, the tools, the jobs, the world of human beings are all designed in human form, and who can look like an adult can take over all his work more seamlessly and directly.
Physical AI targets the largest market for human labour itself.
It's a big deal.
Global manufacturing is becoming more and more difficult, and population ageing leaves a gap of one year to one year for old age, care and logistics. Citicorp had estimated that the number of human robots could reach 648 million globally by 2050, equivalent to more than 40 per cent of today ' s global vehicle holdings.
Letting a human robot work is like making a person work, it looks simple, but behind it is a complex series of dismantlings.
Suppose you say to it, "Get me a Coke from the fridge."
It starts with "eye" -- camera and 3D visual sensors to see where the refrigerator is, where the door handle is, and how many floors the Coke is.
And then the brain. The multi-modular LLM splits this sentence into a series of actions: walking, pulling open, finding Coke, picking it up, closing the door and handing it over. These models are known in the industry as VLA, which is the Visual-Language-Action model.
And then there's Little Brain, which is the motion control system. Walking on both legs is essentially the edge of every step falling, and it adjusts its weight between milliseconds to ensure that it does not fall.
And finally, the muscles and the bones. A human robot usually has about 40 joints, each of which is filled with electrics, speed-retarders, silk bars and sensors. You have to have touch and strength on your hands: you can grab a Coke bottle a little hard, you can grab a steel ball a little hard, you have to hold on to the eggs.
Sounds like every step is not complicated, but for robots, it's every step.
There's a famous Moravik paradox in artificial intelligence: it's relatively easy to win a world championship on a computer, and it's hard to make a robot fold a dress like a three-year-old.
Hard, hard data.
The language of LLM has evolved dramatically because humans have saved words on the Internet for decades. But there is no “Internet” in the physical world. Every grab and every fall of a robot has to be tried over and over again in the real world, slow and expensive.
There are two ways to break the rules. One is to get real machines to work in large quantities and to collect data while dry; the other is to “open a school” in the virtual world.
The latter is like the pilot's advanced simulator: one million times in the virtual world, then the first step to the real world. In March of last year, Yvette released a set of data: with its tools, 780,000 robotic synthetic action tracks could be generated in 11 hours, equivalent to 6,500 hours of human demonstration data.
Even so, today's physical AI is still far from universal.
Wang Xing Xing Xing, founder of Wow tree technology, said at the World Robotics Congress in August: In fixed scenes, with adequate data collection and specialized training, robots can perform almost 100 per cent of success rate; success rate can decline significantly if one object is replaced or the environment is moved slightly.
Wang Xing Xing gave a smart "ChatGPT Time" standard: to take a robot to any strange environment, like your home, it could almost do 80 percent of the job. By that date, the industry will have reached the real threshold of the outbreak. At that time, his expectation for that time was two to three years, five to ten years.
On 12 October, however, Wang Hing Xing took this premise in a round-table dialogue at the opening of the twenty-sixth China International Industrial Fair. He believed that a rapid advance in modelling capacity would soon break through the barrier, and that the threshold could come within a year or two. He even talks: next year, when the GPT-7 comes out, it's possible that the smart question will go right through.
In other words, a few months ago, physics AI was still a distant future, but now it has run into the last straight track before the rush; and its transformation of manufacturing, logistics, and services is China’s most familiar home.
There's a lot of players on Global Physics AI's card table. It's very different.
Mask is the best preacher and the worst bett.
As early as June 2024, at the Tesla Shareholders' Conference, he made a big deal: One billion man-made robots are likely to be built in the world each year, with Tesla ' s share likely to be well over 10 per cent; when it is large enough, the cost can be reduced to around $10,000 per unit and $20,000 sold.
He called Optimus Tesla “the biggest product in history” and wondered: with it, everyone could have a top surgeon.
But Mask's schedule is always faster than the factory. He had predicted that in 2025, thousands of Optimus would be working at the factory, and by January of this year he had changed his mouth: Optimus had not been used in a substantial way. However, Mask did indeed take home.
He stopped Model S and Model X and converted the Fremont line to Optimus. In the third quarter of this year, the cost of research and development in Tesla jumped by between $69 and $2.76 billion over the same year, net profits decreased by 43 per cent over the same period, and money was mainly targeted at new operations such as super-smart, robotics, etc., as Tesla described.
As early as July 1st, Mask tanned a photo shot with the Optimus production team: The beginning of the production would be extremely slow, because everything was new, “It was not like car making”.
On 21 October, Optimus finally started work on the production line. According to Mask, most of the 1 million production lines that are designed to produce annually have arrived; Optimus will have to switch to self-study A5 chips around April next year. His judgment remains ambitious:
Optimus will be the largest product on Earth, with more robots than humans in the future.
Wong In-hoon is going another way: without building robots, the world's robots will grow up on a platform in In-Wida.
Young Weida's play is a set of “family buckets” from training to upper body: Cosmos World Model is responsible for understanding and generating the physical world, Isaac Simulation Platform is responsible for “opening school”, GR00T is the universal brain for human robots, and Jason Thor is a chip for robots.
At the March GTC conference, the industrial robot giants ABB, Hanako, Kuka, Andokawa, together with the United States Figure and China's Ji-won, appeared on the list of partners in Wei Da.
At the GTC conference in Taipei at the end of May, publishCosmos 3, called the world ' s first fully open-source full-model model, could compress the physical AI training and evaluation cycle from months to days.
At the GTC conference in Berlin on 21 October, Huang In-hoon stood on the stage with a couple of little robots 25 cm tall and ducky. He turned them around, and he made them “sleep”, and one of them turned his back, and Hwang In-hoon told him to focus. Next giggle.
After the duck, he brought out a serious guy: a human-shaped robot built by Agile Robots in Germany, based on Jason Thor and GR00T, which plans to build 1,000 to 2,000 machines next year. Wong In-hoon's judgment on that day was that the robot's ChatGPT moment was fast approaching.
What Wong In-hoon would do is to move into the physical world the “chip+soft+ecology” that he earns in the data centre: whose robots win in the end, their brains grow up on the platform of England.
The rest of Silicon Valley's players place the most money on the brain.
On CES in January this year, Google DeepMind announced that it would work with Boston Dynamics to install the Gemini Robotics model into a new generation of Atlas human robots, and that the new Atlas would enter its Georgia factory this year, as planned in a modern car.
It's quite dramatic that in 2017 Google's robots went into big, far from paying off, selling Boston power to soft silver; and for nine years, Google returned as a “brain supplier”.
The valuation of start-up companies is even more dramatic. Figure AI ' s valuation reached $39 billion in September of last year; the Physical Intelligence, created by former Google DeepMind researchers in collaboration with Stanford and Berkeley scholars, which only serves as a universal model of robotics, was exposed in March of this year to a new round of approximately $1 billion in valuation negotiations, which doubled over four months. The co-founder, Sergei Levin, positioned the company as the "ChatGPT of Robots".
On 28 September, AMD announced the purchase of the World Labs, founded by Lee Fei Fei with a full stock of about $8.2 billion, the second largest acquisition in AMD history after Céline. The World Labs of Lee Fei was founded in 2024 as “space intelligence”: from text, pictures, video generation and simulation of an interactive 3D world, considered upstream of the physical AI. Sozfun says that to be a platform for the next generation of AI, it is important to understand how models evolve. Li will be Executive Vice President and Chief Scientist of AMD. The deal, too, hit the chip giant's competition from a sales card to a world model.
Ying Weidar took Cosmos open source and lets the whole industry do physical AI on its platform; AMD simply buys home a top-level laboratory, which in turn defines the chip with a model.
However, Silicon Valley has the smartest brain and the most expensive valuation, but it is not here, but in China. China’s robots may not be smart enough and have more limited capital, but they actually lead in volume.
China ' s physical AI, with human-shaped robots at its core, is based on the following pattern: the body is running, the car is turning around and the giant is setting up a stage.
Let's look at the new forces.
According to the market research institute Smart Analytics Global, in the first half of this year around 191 million human robots were exported globally, with Chinese manufacturers taking over 97 per cent. Of these, approximately 8,400 were shipped by Ji-won, with a global share of 44 per cent, ranked first, and about 5,900 by Woo-ju, with a share of 31 per cent. Together, the two accounted for three quarters of the world.
The founding combination of Ji-won was quite impressive: Chairman Deng Taehua was the former Vice President of China; President Peng Zhihui, the “childish prince” on Bilibili, joined as a “genius teenager”. Two Chinese, three years later, paid the company $300,000 from its first year of operation to over $1 billion in 2025.
On 28 June, the 15,000 G2 robots in Jiwon got offline less than three months before the 10,000th. Instead of a progress hall, the robot was sent directly to the tablet computer line at the Dragon Flag Tech Nanchang factory, with eight other G2s for dry-quality screening. The eight robots had completed more than 175,000 mass tests and success rate99.99 per cent of the six days of live broadcasts.
A woo tree is another form of gas. Wang Xing Heung started with four-foot robots and started a business in 2016, and it was 10 years.
At the end of May this year, the Daily Economic News reporter walked into the woo tree headquarters of the Hangzhou Riang: at the door, a four-foot robot is being tested by researchers; on the top of the building, a wheeled machine dog is not being patrolled; the first-floor gallery has been vacated and the sales team is moving to a new building that has not yet been renovated. The company grew faster than the house.
Smart cars and human-shaped robots, in the fields of sensors, chips, electrics, batteries, algorithms and so forth, are at the same base height and, according to the Gestapo Institute of Automobiles, the supply chain of the two is measured to be 60 to 70 per cent. China ' s new energy and automobile industry chain is a small extension of the robotics industry chain. In fact, Chinese cars are also actively deploying robots.
Of these, the most active and thorough turn was the peng. In an internal letter dated 10 June, Peng Ho said that this was a critical step in the transition of Peng from a “smart car company” to a “physical AI company”. In August, Peng robotics completed a first round of financing of over $900 million, with a post-investment valuation of over $6.3 billion, with Tencent and Ali entering as strategic investors. Beyond Peng, the ideal Li wants to speak in March: 2026 is the key year for the ideal to evolve into a “smart” enterprise.
The understanding of robots has moved beyond the Biadi of Tesla on new energy vehicles, and the focus has been on manufacturing. Its Executive Vice President, Li Ko, said that robots competed for manufacturing, software and hardware capabilities, while the AI capacity accumulated on cars could be transplanted to robots.
China’s technology giants are also increasing their distribution of robots.
In June last year, publish, the Claud Robo, China’s cloud-smart platform, upgraded and opened public surveys in June of this year, providing a whole line of water from data, models, simulations to deployments, has worked with more than 30 industry partners and 20 home-grown manufacturers. Mi's Thunder Army defined 2026 as “year of smart LLM”.
But China can run fast with human robots, not just by one or two star companies, but by industrial chains and ecology. According to the Ministry of Industry and Communications in July, the Chinese human robotic integrator model now accounts for more than 400, or more than half of the world ' s total; the share of four-foot robots in global sales is close to 70 per cent. Electrical, speed-retarders, batteries, sensors, structural components, almost every link, can be found within one or two hours of the long triangle.
In Shenzhen South Mountain, about 10 kilometres of Rixen Boulevard is a complete system of key robotic technology development and component manufacturing, ranging from algorithmic software such as large brain models, sensory sensors, laser radar, etc., to implementation systems such as servulating electrics, power modelling units, and core components such as mechanical arms, smart hands, to rapid delivery in the Hole.
Here, the “mornings of the morning, the drawings of the afternoon, the next-day overlaps” are normal. A tactile sensor with a smart hand, which used to cost $100,000 on imports, now produces 199. From January to August this year, robotic cargo orders in the southern mountains increased by 94.4 per cent over the same period.
The United States is telling the grandest story, building the smartest brain; China is making the most body, running the fastest volume, entering the most real scene. When China and America combined, the whole human robot could run faster, and AI would move out of the screen faster, moving to all walks of life and home.
When the industry gets hot, the money comes.
According to industry statistics, in the first half of this year, Chinese man-made robotics produced more than 40,000 units, a 50 per cent increase in the ring in the second quarter; the total amount of the track financing reached 104.1 billion yuan, more than in the whole of 2025. The Ministry of Industry and Communications predicts that the total capacity of the aircraft will exceed 100,000 this year. Citicorp is more than predicting that the global human robotics market will reach $7 trillion by 2050.
Demand is also moving from performance to “induction”. In June, the Ministry of Industry and Communications, together with the State Commission for State Administration, issued a joint circular to launch a 2026 human-shaped robotics and smart-based reality training initiative, requiring that, by the end of 2026, priority products, such as human-shaped robots, take the lead in completing application validation and normal deployments in a representative set of scenes and initiate “operational models”. The Shanghai Action Plan for Accelerated Development of a New Generation of Smart Manufacturing in Shanghai, publish of 12 October, is even more explicit in promoting 8,000 man-made robots into factories.
According to IDC data, in the first half of the year around 25,000 human robots were shipped globally, an increase of 432 per cent over the same period; the share of scientific education, exhibitions showing such “no-work” scenarios fell from 83.8 per cent to 69 per cent throughout the year. This means that the robot is moving from a stand to a position.
But the more fun it is, the more calmly you ask: "Who's money will end up in the pocket?"
Looking back at the history of the new energy car, it is worth remembering a rule: when the car is bleeding, the first to make money is often the core component. The Ningde era is the best footnote. Physical AI probably will repeat this.
The whole machine is far from fixed, and today's champions may not still be on the table tomorrow; but whoever wins the robot won't be able to turn on the joints, the electrics, the brakes, the wire bars and the sensors.
The next industrial and investment opportunities to be generated by Physical AI focus on three lines:
Line 1: Look at the body, then the brain.
In the case of Share A, for example, Physical AI, which is actually certified by an order today, is mostly concentrated in the enforcement chain. The reason for this is real: the value of the brain is now largely taken away by British Weidar and model companies that are not yet on the market; China holds the strongest manufacturing base in the world, so China's opportunity is first hidden in the body of Physical AI.
Green tunes represent this line. Harmonic decelerators, which had long been monopolized by Japanese firms, and green waves, which had opened their mouths, now have the planetary roller poles on a massive scale and are fully integrated into the smart core supply chain.
The Three Flower Brains, Tops Group represents another type of player: the auto parts giant, with its capacity to make and reduce their size, cuts into a robotic implementer and is considered by the market to be the core player in the Tesla supply chain.
In addition, two-ringed fine gears have a hard grip on their own subdivisions, as well as a constant hydraulic cord, an empty glass machine for the beeper, a 3D vision for the light in Obi, and a power sensor for the cleavage.
Line 2: See if the industry can get to the bottom.
It's a marathon, and there's a bump in the middle. It is often the main industry that is able to reach the end of the line with a sufficiently strong cash flow.
Sichuan technology is typical. In the first half of the year, 24,675 million yuan was spent, and the rate of the national market for the universal service system rose to 35.2 per cent, the highest. Human robots are not transformations for them, but natural extensions of the capacity to serve, power and control. Such companies are not necessarily the most resilient, but they are the least afraid to retreat.
Line 3: Look at the order, not the concept.
Reyser's smart is a sample that talks in financials. It focuses on the two main groups of joints and smart hands, with a profit of $1.247 billion in the first half of the year, an increase of 39.92 per cent over the same period; and a net profit of $194 million for the mother, an increase of 62.79 per cent over the same period, with products having been placed on bulk orders by the country ' s mainstream robot manufacturer.
On the whole machine level, WikiLeatech is already the most authentic human robot in unit A and an anchor in the valuation of the entire plate; the simulator has its own capital platform in unit A by taking up the upper latitude.
In turn, there are three types of companies that require extra care.
One category is “cooperative” “participating” “strategic agreement”. On 8 June, the stock price rose and the stock price of the shopkeepers was cut off from the physical AI concept for their collaboration with Zuo Cai; the same night, the company issued a communiqué clarifying that it was a framework agreement, had no substantive orders and had not contributed a penny to the income.
One is that valuations are well ahead of performance. By the beginning of July, the Green Harmonics had accumulated an increase of more than 150 per cent in stock prices since the beginning of the year, while it had earned only $140 million a quarter, and the market had bought several years of growth ahead of schedule.
One is the concept of a blueprint, and without a diamond, you're going to be working on the china. For example, a company that neither has strong capital nor the best technical talent to claim to be the brain of a machine or even to compete with Britain.
The deeper test is price. Once the robot is released, the unit price of the parts must be kept under constant pressure, like the photovoltaic and power cells of the year. The final winner is either the one with a unique core of technical barriers or the one who can minimize the costs and scale.
After all, investing in physics AI doesn't have to bet on who can make the best robot, but on what nobody can.
The change came sooner than most people expected. A self-reflection of Wang Xing Xing's speech at this year's Labour Fair deserves to be remembered by everyone who looks at this track:
"The biggest mistake I've made in the last decade or so is not enough to believe in AI."
References:
[1] Xinhua Society: In a publish conference on World Congress of Artificial Intelligence, 2026, the Ministry of Industry and Communications stated that our human-shaped robots were expected to produce more than 100,000 machines by 2026. Station
[2] Tesla Press and Performance Telephone Conference, third quarter of 2026 (New Wave, daily economic news related coverage)
[3] Reports on GTC 2026 San José, Taipei, publish, official Berlin Congress, and the Science and Technology Journal
[4] Smart Analytics Global, IDC: Global human robot departure data for the first half of 2026
[5] First Financial, Securities Times: coverage of Wang Heung's speech at the 2026 World Robotic Congress and the 26th China International Industrial Fair
Auto-fetched and translated by AI Atlas. All rights belong to the original author.