刚刚，知名经济学家任泽平在其微信公众号发文谈及，人形机器人行业持续扩张，第一批独角兽却纷纷倒下。2026年上半年，全球人形机器人卖近2.5万台，同比增长432.1%；1至8月国内具身智能融资466起、1245.1亿元，达2025全年2.7倍，百亿独角兽从3家增至22家。另一边，公开可查的机器人公司破产、解散、清算超20家，人形整机、纯概念大脑、数采是重灾区。估值曾摸到223亿元的达闼机器人被法院裁定受理破产清算，硅谷K-Scale Labs账上只剩40万美元，解散退款，吉利系一星机器人存活不到半年。
Just now, well-known economist Ren Zeping wrote on his WeChat official account that the humanoid robot industry continues expanding while the first batch of unicorns falls. In the first half of 2026, global humanoid robot sales approached 25,000 units, up 432.1% year-over-year; January to August domestic embodied intelligence financing totaled 466 deals and 124.51 billion yuan, 2.7 times full-year 2025, with hundred-billion unicorns growing from 3 to 22. On the other side, publicly verifiable robot companies bankrupt, dissolved, or liquidated exceed 20, with humanoid complete machines, pure concept brains, and data collection as hardest hit areas. CloudMinds, whose valuation once touched 22.3 billion yuan, was accepted for bankruptcy liquidation by court ruling; Silicon Valley's K-Scale Labs had only $400,000 left, dissolved and refunded; Geely-affiliated Yixing Robot survived less than half a year.
任泽平把最大病根归结为短视，机器人忙于走秀，虚假繁荣。中国信通院口径下，2026年上半年全球出货里，文娱商演加科教数采合计超过六成，导览导购19%，智能制造13%，仓储物流只占5%。沙利文口径下，2026上半年全球卖出3.6万台，真正用于生产作业的只有0.12万台，只占3.3%。走秀来钱快，攻技术来钱慢，于是企业都去走秀，而没有按照第一性原理打造真正的具身智能。这个判断和前面讨论的万元级人形机器人买家画像高度吻合。前面讨论过，当前C端买家的画像高度集中，机器人爱好者、AI技术发烧友、高校科研团队、短视频创作者、科创培训机构。普通工薪家庭极少愿意花数万元，购买一台无法承担家务劳动的人形设备。简单来说，他们在玩，不在用。
Ren Zeping attributes the biggest root cause to short-sightedness—robots busy with shows, false prosperity. Under CAICT's scope, in 2026 first-half global shipments, entertainment performances plus science education data collection combined exceed 60%, guided shopping 19%, intelligent manufacturing 13%, warehousing logistics only 5%. Under Frost & Sullivan's scope, 2026 first-half global sales were 36,000 units, with only 1,200 units truly used for production operations—just 3.3%. Shows bring money fast; attacking technology brings money slowly—so enterprises all go to shows without building embodied intelligence according to first principles. This judgment aligns highly with the previously discussed ten-thousand-yuan humanoid robot buyer profiles. Earlier we discussed current C-end buyer profiles are highly concentrated—robot enthusiasts, AI tech fans, university research teams, short video creators, STEM training institutions. Ordinary working families rarely spend tens of thousands on a humanoid device that cannot handle housework. Simply put, they are playing, not using.
技术这道坎，少有人肯啃。正如马斯克所批评的，许多复杂演示是事先编排的，那不叫真正的AI，能应对没见过的情况才算。宇树王兴兴8月在世界机器人大会上披露，固定场景标定后成功率近100%，环境稍变就断崖下滑，卡在最后几毫米的触觉误差。Figure9月披露，30个陌生家庭实测叠毛巾、铺床，完整任务成功率56%，对照组9%。台上丝滑不等于实际能干。收入撑不起估值，只能靠PPT续命。达闼累计融资超50亿元，软银愿景、阿里都是股东，可法院口径下的被执行总额只有3530万元。深兰102名职工债权超2300万元，盈合2025年全年营收仅433万元，智橙动力死在回款链上。四家错的是同一件事，PPT融资大法，先按融到的钱定组织，再去找能付钱的客户。
The technology hurdle—few are willing to gnaw at it. As Musk criticized, many complex demonstrations are pre-choreographed; that is not真正的 AI; only handling unseen situations counts. Unitree's Wang Xingxing disclosed at the August World Robot Conference that after fixed-scenario calibration success rates approach 100%, but slight environmental changes cause cliff-like declines, stuck at the last few millimeters of tactile error. Figure disclosed in September that in 30 unfamiliar homes testing towel folding and bed making, complete task success rate was 56%, versus 9% control. Smooth on stage does not equal capable in practice. Revenue cannot support valuation, surviving only on PPT. CloudMinds cumulatively raised over 5 billion yuan, with SoftBank Vision Fund and Alibaba as shareholders, yet court-scope enforced total is only 35.3 million yuan. DeepBlue's 102 employee claims exceed 23 million yuan; Yinghe's full-year 2025 revenue was only 4.33 million yuan; Zhicheng Power died on the payment collection chain. All four got the same thing wrong—the PPT financing method: first set organization by money raised, then find customers who can pay.
贪大求全，战线太长。达闼同时做云端、配送、安防、人形，深兰铺向清洁、消毒、配送、园艺，盈合横跨应急、城市治理、仓储、医疗、工业，没一条跑通。灵巧手占整机成本14%到18%，塞满电机和触觉传感器，装配和标定复杂度压不下去，先亏钱抢量再靠量产拉毛利的老套路不成立，规模越大越亏。钱不给整机，给卖铲人。IT桔子口径下，2026年前8个月机器人赛道融资1245.1亿元，超过2025年全年两倍，但前十家拿走约59%。融资总额涨、整机吸收融资的占比在降，关节模组、力传感器、触觉感知升温，头部灵巧手企业估值两轮从30亿跳到80亿。资本不再赌谁整合得出一台好机器，转而下注绕不过去的零件。
Greedy for scope, front lines too long. CloudMinds simultaneously did cloud, delivery, security, and humanoids; DeepBlue spread into cleaning, disinfection, delivery, and gardening; Yinghe spanned emergency, urban governance, warehousing, medical, and industrial—none ran through. Dexterous hands account for 14% to 18% of complete machine cost, packed with motors and tactile sensors, with assembly and calibration complexity unable to be pushed down; the old playbook of losing money first to grab volume then relying on mass production to lift gross margin does not hold—the larger the scale, the more the loss. Money does not go to complete machines but to "shovel sellers." Under IT Juzi's scope, 2026 first-eight-month robot track financing was 124.51 billion yuan, over twice full-year 2025, but the top ten took about 59%. Total financing rose while complete machines' share of financing fell; joint modules, force sensors, and tactile perception heated up, with leading dexterous hand enterprises' valuation jumping from 3 billion to 8 billion in two rounds. Capital no longer bets on who can integrate a good machine but bets on parts that cannot be bypassed.
任泽平总结称，AI远超蒸汽机、电力、互联网，第四次科技革命，康波周期级别，人形机器人爆发在第二阶段，任重道远。这个判断和前面讨论的硅谷与五道口的两种造神运动、微亿智造的微利考验、智元估值500亿但IPO暂缓是同一个脉络。前面讨论过，当年轻一代最聪明的大脑，不再将精力用于突破物理世界的真实瓶颈，而是熟练地掌握了迎合资本FOMO情绪、制造估值泡沫的话术时，这个行业就会陷入危险的停滞。任泽平这篇文章，从经济学家的视角给出了同样的判断，第一批独角兽倒下不是行业衰退的信号，而是行业从讲故事、拼估值的阶段，转向拼技术、拼交付、拼收入的阶段的必经过程。走秀来钱快，攻技术来钱慢，但只有攻技术的公司才能活到第二阶段。
Ren Zeping concluded that AI far exceeds steam engines, electricity, and the internet—a fourth technological revolution at Kondratiev cycle level—with humanoid robot explosion in the second stage, a long and arduous task. This judgment follows the same thread as the previously discussed Silicon Valley and Wudaokou's two "god-making" movements, Micro Intelligence's thin-profit test, and AgiBot's 50 billion valuation with IPO postponed. Earlier we discussed that when the younger generation's brightest minds no longer devote energy to breaking through the physical world's real bottlenecks but skillfully master the rhetoric of catering to capital FOMO and manufacturing valuation bubbles, the industry will fall into dangerous stagnation. Ren Zeping's article gives the same judgment from an economist's perspective—the first batch of unicorns falling is not a signal of industry decline but an inevitable process as the industry shifts from storytelling and valuation competition to technology, delivery, and revenue competition. Shows bring money fast; attacking technology brings money slow—but only companies that attack technology can survive to the second stage.
