文 | 定焦One，作者 | 王汉星，编辑 | 魏佳
在过去的12天里，宇树科技的股价连续反弹，累计涨幅超过70%，总市值回升至2800亿元以上，距离上市首日开盘还差大约1500亿元。
相比于市值的大举反攻，宇树在10月28日盘后发布的上市后首份财报却显得平淡很多，大多数财务数据都在预期之内，Q3宇树实现营收6.65亿元，同比增长69.93%，相较于之前增速有所回升；扣非后的净利润为1.05亿元，同比下降17.42%，今年以来，宇树明显提高了研发费用的支出，利润下滑的趋势也与之前几个季度一致。
一周前的2026世界物理AI大会暨宇树合作伙伴生态大会上，宇树创始人王兴兴给出了一个激进的判断，他认为GPT-6的一些能力已经可以解决机器人的部分规划问题，随着通用大模型能力的提升，快的话一到两年，慢的话两到三年，物理AI可能迎来重大突破。
很长一段时间里，宇树都被外界认为是一家偏硬件属性的公司，强于制造机器人本体和关节、运控系统，但弱点是机器人“大脑”。改变这一局面的不是宇树自己，而是OpenAI于9月4日发布的GPT-6 Astra，当大语言模型开始具备一定具身智能通用大脑的能力后，行业竞争的起跑线被重划，在出货量和成本控制能力上具备明显优势的宇树反而成为最大的受益者之一。
资本市场的态度也随之发生转变，摩根大通一个月前还在以300元/股的目标价看空宇树，一个月后将公司目标价直接上调至995元/股。
然而，这种态度转变和价值重估并没有发生在所有的具身智能公司身上，相反，整个Q3具身行业经历了一轮明显的降温，一级市场的融资明显收缩，通往二级市场的IPO门槛被抬高，甚至有创业公司倒闭。
从个股的市值狂欢，到公司经营业绩的按部就班，再到行业的遇冷，是什么原因造成了这种反差？宇树从上市即巅峰的每股1100元，一路跌至不足四成的404元附近，再大幅反弹，这家公司的估值危机已经过去了吗？
01.研发投入占比翻倍，毛利率下滑
单看财报，宇树依然保持着较快的收入增长和可观的盈利水平，只是增长的代价变得更清楚。
营收与之前略有差异的部分是增速，根据宇树此前公布的数据，今年Q1的营收增速为68.49%，1-6月营收11.52亿元，同比增长48.54%，呈现出下降趋势，到了Q3营收增速回升至69.93%，在经历了2025年的高速增长后依然保有接近70%的增速，说明宇树仍然有一定的韧性。
出货量提供了另一项参照。美国市场研究机构SAG发布的2026年上半年报告显示，宇树人形机器人出货量同比增长170%至5900台，半年的出货量已经超过公司此前披露的2025年全年5500台。销量扩张与收入增长相互印证，说明它在本体制造和交付上的优势还在延续。
但营收增长的同时，宇树上半年的毛利率出现了一定程度下滑。宇树上半年综合毛利率为56.01%，较上年同期下降4.15个百分点，前三季度进一步降至54.02%。按累计收入和成本减去上半年数据粗略测算，Q3单季毛利率约为50.6%，下探速度有所加快。
由于宇树没有披露分产品数据，无法直接确认原因。但宇树在此前的IPO问询回复中提供过一条信息，2024年至2025年前三季度，人形机器人毛利率由68.44%降至62.91%，主要由于中型产品G1销售占比提高，而G1的单价和毛利率均低于全尺寸H1、H1-2。公司当时对G1的定位，就是以性价比扩大销售规模。
这套机制可能延续到了2026年。宇树曾通过设计优化、供应链议价和规模效应，让单位成本降得比价格更快，如今毛利率持续回落，至少说明价格和产品组合的变化还未被成本优化完全对冲。
费用端的变化更加直观。Q3研发投入1.27亿元，同比增长249.58%，占营业收入的19.10%，同比提高9.82个百分点接近翻倍。前三季度研发投入2.63亿元，同比增长191.48%。公司将这些投入主要归因于具身大模型、算力、研发团队扩充和新产品开发。同期销售费用也增长202.45%至2.30亿元。研发与销售费用在毛利空间收窄后继续压缩净利润。
前三季度归母净利润达到3.87亿元，同比增长270.54%，这一高增速很大程度上源于上年同期一次性确认了较大金额股份支付费用。扣非后，前三季度净利润为3.49亿元，同比下降18.78%，更能反映当前收入扩张与利润承压并存的状态。
经营现金流也呈现出相似趋势。前三季度经营活动产生的现金流量净额为1.94亿元，同比下降54.69%。公司解释，购买商品、支付职工薪酬以及销售、研发等现金支出随业务扩张明显增加。
这份财报没有改变宇树仍在增长、持续盈利的基本事实，却也不足以单独解释两周内超过70%的股价反弹。
财报发布前，宇树已经因连续上涨两次披露异常波动公告。公司在10月26日晚间的公告中提示，当日滚动市盈率为465.22倍，显著高于通用设备制造业平均的38.54倍，物理AI从实验室演示走向规模化应用，仍面临感知精度、运动控制、数据积累和软硬件协同等挑战。
02.大模型当“大脑”，宇树先受益
当具身智能跨过量产的门槛后，行业内的所有人都知道下一个“Game Changer”是具身大脑，但很少有人会预见到这个“Game Changer”来自于大模型公司。
根据王兴兴透露，GPT-6 Astra的发布给行业带来了巨大的冲击，宇树在内部尝试用这一模型做了几项机器人实验，包括让机器人绘画、下围棋。宇树事先并没有针对这些任务编写复杂程序，仅仅是通过提示词，让机器人来完成，在“零训练”的条件下，Astra也能顺利完成其中的一部分任务。
在过去，机器人往往需要针对具体任务编写程序，或进行专门训练，而现在不同了，通用大语言模型已经具备了直接控制机器人完成部分任务的潜力。
这也改变了王兴兴对行业的预期，他认为具身智能行业临界点将很快到来，甚至可能是一两年之内。他还提到，“现在我个人对AI的信仰反而与日俱增，过去十几年我犯的最大错误就是对AI还不够相信。到现在可能很多人还没有充分使用或者拥抱AI。”
这也是宇树近期估值修复的原因之一，当具身智能的通用大脑能力掌握在一家大模型公司手里，而非其他任何一家具身智能公司，它们在短时间内不存在明显的竞争关系，这就使得拥有硬件优势的宇树率先受益。
但这只是一部分原因，Astra也并非无所不能，它的大脑能力主要集中在理解意图、空间理解和决策能力上，在末端的精细操作上依然存在短板，并且十分依赖云端运行，在具体应用过程中还要考虑延迟、网速、信号等一系列问题。
短时间内说大语言模型降维打击具身模型还为时尚早，但它给行业提供了一个解决当前瓶颈的可能性，王兴兴则认为，“明年GPT-7出来，可能直接把具身智能问题打穿。”
除了以上原因，近期宇树股价连续上涨还要考虑到流通性的问题，宇树上市后只有7.44%的股份可以交易，其余92.56%处于锁定状态，盘面价格实际上由极小一部分筹码决定，因此短期内出现市值的大幅波动是正常现象，是否真的存在价值重估还需要把观察的时间拉长。经历上市后的持续下跌，估值本身也存在修复空间。技术预期、机构上调评级与筹码结构相遇，反弹因此被进一步放大。
宇树上市之初备受市场关注还有一个重要的原因是它的稀缺性，它是A股具身智能第一股，此前整个赛道的火热大多停留在一级市场，二级市场始终缺少一个估值锚点。而近期具身智能行业整体的降温预期、IPO门槛提升反而增强了这种稀缺性。
首先是一级市场的融资减少，IT桔子数据显示，今年上半年国内具身智能赛道融资总金额达到935亿元，平均下来上半年每个季度融资大约为468亿元，而有媒体统计，到了今年第三季度具身智能领域发生大约185起融资事件，融资金额为333.35亿元，规模下降了100多亿元。
具身公司越来越难从投资人那里拿到钱。与此同时，在8月末具身智能PE/VC圈子里流传出一则消息，行业IPO将收紧，尤其是对于仍在亏损的企业，对细分赛道的待上市公司考核标准更加严格。
这些因素叠加在一起，一些创业公司开始扛不住了，其中不乏明星项目。今年2月成立的具身智能公司无境动力，6月还被当地媒体报道“拿下近10亿元订单”，近期被曝出8月起开始欠薪，三四十名员工被拖欠近800万元，据员工称公司已停工。曾吸引美团战投、字节跳动战投、金沙江创投和熊猫资本先后投资，估值最高冲到5亿美元的盈合机器人，于10月19日首次召开债权人会议，2个月前，法院发布破产受理及债权申报公告，它正式进入破产程序。
一边是行业遇冷，一边是近期宇树的市值狂欢，但还不能说宇树的估值危机已经过去，只是被这轮反弹暂时盖住了。推动这轮上涨的力量更多来自GPT-6带来的技术想象，而不是经营本身。
具身智能并未失去资本关注，但估值锚点已经转移到了客户、订单与现金流上。对宇树而言，技术预期推高了这轮反弹，经营质量将决定这轮重估能走多远。
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。
Wen, focus on One, by Wang Hanxing, editor Wei Jia
Over the past 12 days, the stock price of Waki technology has rebounded continuously, with a cumulative increase of more than 70 per cent. The total market value has risen to more than $280 billion, leaving about $150 billion short of the opening date.
In contrast to the massive countervailing of market value, Woo-ki ' s first post-marketed publish on 28 October was much more modest, with most of the financial data being expected, Q3-G received $665 million, up from 69.93 per cent, up from the previous rate of increase; net profit after deduction was $105 million, down from 17.42 per cent. Since this year, Woo-ki has significantly increased expenditure on R & D costs, and profits have declined in line with previous quarters.
A week ago, at the 2026 World Physics AI Congress and the Woo Tree Partnership Ecology Congress, the founder of Woo-hee, Wang Hing Xing, made a radical judgement that some of the capabilities of GPT-6 can already solve some of the robotic planning problems. As the capability of General LLM increases, one or two years soon, and two or three years later, physical AI could be a major breakthrough.
For a long time, Woo-chul was considered by the outside world to be a company with partial hardware properties, stronger than manufacturing robotics and joints, transport control systems, but the weakness was robotic “brain”. It was not Woo-ki himself, but Open AI’s GPT-6 Astra of publish on September 4th that changed the situation. When the large-language model began to have a certain intellectual general brain, the industry’s competitive start-up line was redrawn and the wild tree, which had a clear advantage in terms of volume and cost control, became one of the greatest beneficiaries.
The attitude of capital markets has also changed, with Chase Morgan looking at the space tree at a target price of $300 a month ago and moving the company ' s target price directly to $995 a month later.
However, this attitudinal shift and revaluation of value did not take place on all smart companies, but instead, the entire Q3 industry experienced a clear wave of cooling, financing for the first-tier market shrunk, and the IPO threshold to the second-tier market was raised and even start-ups collapsed.
What is the cause of this contrast, from the market value of a stock to the orderly performance of the company and the cold of the industry? Was the valuation crisis of the company over?
The ratio of R & D inputs doubled, under M & M Slide
In financial terms alone, the trees have continued to sustain faster income growth and considerable levels of profitability, although the costs of growth have become clearer.
This year, according to data previously published by Woo-ki, the Q1 battalion grew at a rate of 68.49 per cent, with a decrease of RMB 1,152 million in January-June, or 48.54 per cent over the same period, to a rate of 69.93 per cent for the Q3 battalion, which remained close to 70 per cent after a high rate of growth in 2025, indicating that the tree remained resilient.
Another reference is provided by the volume of deliveries. The first half of 2026 report of SAGpublish, the United States market research institute, indicates that Uyuki human robotics have grown by 170 to 5900 units over the same period, and that six months of deliveries have exceeded the 5,500 units reported by the company throughout 2025. Sales expansion and income growth are mutually evidence of its continuing advantages in home manufacture and delivery.
However, the increase was accompanied by a slight decline in the Maori rate in the first half of the year. The combined Maori ratio for the first half of the year was 56.01 per cent, a decrease of 4.15 percentage points compared to the same period of the previous year and a further decline to 54.02 per cent in the first three quarters. Based on a rough estimate of cumulative income and cost less data for the first half of the year, the Q3 single-season Māori rate is about 50.6 per cent, with an accelerated downturn.
Owing to the lack of disclosure of sub-product data, it is not possible to identify the reasons directly. However, in its previous IPO response, Woo-ki provided information that in the third quarter of 2024 to 2025, the Māori rate for human robots fell from 68.44 per cent to 62.91 per cent, mainly due to an increase in the share of G1 sales for medium-sized products, while the unit price and Māori rates for G1 were below the full size H1 and H1-2 rates. The company ' s positioning of G1 at the time was to expand sales in value terms.
This mechanism may last until 2026. With design optimization, supply chain bargaining and size effects, U-tree has reduced unit costs faster than prices, and the rate of Māori continues to fall, at least indicating that changes in prices and product portfolios have not been fully offset by cost optimization.
Cost-side changes are more visible. Q3 R & D investment was $127 million, representing an increase of 249.58 per cent over the same period and 19.10 per cent of operating income, almost double the increase of 9.82 percentage points over the same period. R & D investment in the first three quarters was $263 million, an increase of 19.48 per cent over the same period. The company attributed these inputs mainly to LLM, computing, expansion of the R & D team and new product development. Sales costs also increased by between $20.245 and $230 million over the same period. R & D and sales costs continue to compress net profits after the Maori space is narrowed.
The increase of 270.54 per cent over the same period in the first three quarters of net profits from the mother ' s return was due in large part to the one-time confirmation of the larger shares of paid during the same period of the previous year. The net profit for the first three quarters, after deductions, was $349 million, a decrease of 18.78 per cent over the same period, which is more reflective of the current state of income expansion in parallel with profit underlay.
A similar trend has been observed in operating cash flows. Net cash flows from operating activities in the first three quarters were $194 million, a decrease of 54.69 per cent over the same period. The company explained that cash expenditure on the purchase of goods, the payment of workers ' salaries and sales, research and development, among others, increased significantly with the expansion of business.
The financial statement does not alter the basic fact that the tree is still growing and making a sustained profit, but is not sufficient to explain alone the rebound of more than 70 per cent of the stock price within two weeks.
Before the financial paper publish, Woo-ki had already disclosed an announcement of abnormal fluctuations due to two consecutive increases. In its announcement of the evening of 26 October, the company advised that the roll-out rate on that day was 465.22 times higher than the average of 38.54 times in the generic equipment manufacturing industry, and that physical AI, moving from laboratory demonstrations to scale applications, still faced the challenges of sensory accuracy, motor control, data accumulation and hardware and software synergy.
LLM as "brain", the tree benefits first
When intelligence transcends the threshold of excess production, everyone in the industry knows that the next "Game Changer" is a brain, but very few will expect that "Game Changer" is from LLM.
According to Wang Xing Xing, GPT-6 Astra ' s publish has had a huge impact on the industry, and Woo-hee has tried to use this model internally to do several robotic experiments, including robot painting and chess. Woo-ki did not prepare complex procedures for these tasks in advance, simply through prompt, so that robots could do it, and Astra could do some of them successfully under “zero training”.
In the past, robots often needed mission-specific programming or specialized training, but now different, the universal large-language model already has the potential to directly control robots to perform some of their tasks.
This has also changed Wang Xing Xing's expectations for the industry, which, in his view, will soon have a critical point in the smart industry, possibly within a year or two. He also said, “Now that my personal faith in AI is growing, the biggest mistake I have made in the past decade or so is not enough to believe in AI. Many people may not yet have fully used or embraced AI.”
This is one of the reasons for the recent valuation restoration of Woo-tree, which has been the first beneficiary of its hardware advantage when its intellectual general brain capability is in the hands of a LLM company, rather than any other furniture-body-smart company.
However, this is partly due to the fact that Astra is not incompetent, its brain capacity is concentrated on understanding intent, spatial understanding and decision-making capabilities, with short panels still present at the end of the line, with heavy reliance on cloud operations, and a range of problems such as delays, web speed, signals, etc. being considered in specific applications.
It is still too early to say that a large-language model has a reduced impact on the body, but it offers the industry the possibility of addressing the current bottlenecks, while Wang Xing Xing believes that “next year the GPT-7 will come out, it will be possible to blow the body-smart problem directly through”.
In addition to the above, the recent consecutive increase in the value of the Woo-tree stock has to take into account the issue of negotiability, with only 7.44 per cent of the shares traded since Woo-tree's listing, the remaining 92.56 per cent being locked up and the price of the booking actually being determined by a very small portion of the chips, it is normal that there will be large fluctuations in market value in the short term, and whether there is a real revaluation of value will require longer observation. The valuation itself had room for repair as a result of the continuous decline in the market. The rebound was further amplified by technical expectations, institutional upgrades and leverage structures.
A further important reason for the market ' s early marketing attention is its scarcity, which is the first of A ' s smartness, after most of the heat of the entire track stayed in the first market, where a valuation anchor was always lacking. This scarcity has been exacerbated by recent cooling expectations for the entire intelligent industry and the rise of the IPO threshold.
First, there has been a decline in financing for the first level of the market, with data from the IT oranges showing that domestic smart track finance totalled 93.5 billion yuan in the first half of the year, with an average of 46.8 billion yuan per quarter in the first half of the year, while media statistics show that, by the third quarter of the year, about 185 financing events had occurred in the area of smarty, amounting to 33,335 million yuan, a decline of more than 10 billion yuan.
It is becoming increasingly difficult for companies to get money from investors. At the same time, in late August, there was a message in the smart PE/VC circle that the industry IPO would tighten, particularly for businesses still in deficit, and that the criteria for the examination of publicly listed companies for the break-down of the tracks would be stricter.
Taken together, these factors are beginning to overwhelm some start-up companies, including star projects. In February this year, smart companies were set up, and in June, local media reported that they had “take down nearly $1 billion of orders” and that they had recently been exposed to salary arrears since August. Thirty-four employees were owed nearly $8 million, and according to employees, the company had been shut down. Having attracted the United States Corps vote, the byte jump-off, the Golden Sands River venture and the panda capital investment, valued at up to $500 million, the first meeting of creditors was held on 19 October, two months before the court, publish, which officially entered the insolvency proceedings.
On the one hand, the industry was cold, and on the other hand, the market value of the recent Woo-tree party could not be said to have passed the valuation crisis, but was temporarily covered by this round of rebounds. The driving force of this round is more from the technological imagination of GPT-6 than from the operation itself.
Smarty did not lose capital attention, but valuation anchors had been shifted to customers, orders and cash flows. For Woo-tree, technology is expected to push the wheel back, and the quality of operations will determine how far this round can go.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.