文 | 白马商评
9月底，市场研究机构沙利文和IDC先后发布人形机器人报告，两家都判断行业在放量，但口径差了四成：IDC统计上半年全球出货接近2.5万台、同比涨432.1%；沙利文统计同期3.6万台，收入5.4亿美元，半年已超2025全年。
机器人公司账本上是另一番景象，收入份额第一的优必选，上半年亏损3.39亿元；另一家头部公司宇树科技，2025年出货超5500台、扣非净利率约34.8%，包含四足机器人业务收入利润，人形机器人的主力是G1这类几万元级的半尺寸机型。
贵机器撑起了行业八成的收入，却还没为谁撑起利润。
上半年2.5万台的出货里，大多数是单价三万美元以下的“便宜货”，真正能干活、卖得贵的全尺寸机器，数量则少得多。这也解释了为什么行业规模持续暴增，企业却赚不到钱。
份额第一，也在亏钱
今年上半年，沙利文在报告里按收入给全球人形机器人企业排座次，前五名全是中国公司：优必选、智元、宇树、银河通用、乐聚。优必选拿了两个第一，整体人形机器人收入市占率16.5%，全尺寸具身智能赛道收入市占率26%。
全尺寸具身智能是沙利文定义里门槛最高的一类：非遥控、非预编程，身高160厘米以上，本体算力通常200TOPS以上。据沙利文统计，2025年优必选人形机器人总销量13838台，其中全尺寸具身智能只有1079台，不到一成。2026年上半年，全尺寸具身智能卖出921台，同期公司人形机器人总销量是16123台，全尺寸具身智能产品占比进一步下降至5.71%。从财务数据看，过去几年，优必选持续亏损，今年上半年仍未盈利。
宇树的情况稍好一些，2025年出货超5500台、营收约17亿元、扣非净利率约34.8%，主力是G1这类半尺寸机型，客户以科研教育为主。
其实，两家走的是同一条路，都是靠非全尺寸机型把量做起来。区别在于，宇树已经赚到了钱，优必选还没有。差别不在产品尺寸，而在模式——宇树卖的是下单发货的标准品，回款快；优必选的全尺寸机型要进工厂、进场景，交付周期长，服务成本高。全尺寸这条行业里最“值钱”的赛道，还没有跑出盈利的样本。
冲量的是便宜货，收钱的是贵机器
按IDC的数据折算，单台均价从去年同期的3.7万美元降到今年上半年的2.96万美元，但这个变化主要来自结构，便宜机型放量，把加权均价拉了下来，并不是每台机器都便宜了两成。销量增速432.1%远高于收入增速322.7%，因此这轮放量的增量集中在低价机型。
根据沙利文的数据，具身智能机型（能自己感知、决策、干活的机器）2025年卖了0.88万台、收入3.75亿美元，占行业收入的75%；2026年上半年卖了1.05万台、收入4.48亿美元，占比升到83%。同期，预编程和遥控机型销量平稳，收入还在往下走。
量价倒挂在两把尺子下都成立。按尺寸看，2026年上半年全尺寸机型卖了0.84万台——只占销量约两成，却拿走了4.08亿美元收入，占行业收入的75%以上；小尺寸机型卖了1.97万台、占出货一半以上，收入只有0.19亿美元。
按智能程度看，门槛最高的全尺寸具身智能（非遥控、非预编程）半年只卖出0.58万台、收入3.2亿美元；拿今年上半年的数据和去年全年（0.53万台、2.6亿美元）相比，销量多了不到一成，收入却多出了近四分之一，单价在往上走。
数量上的反差更直接。摩根士丹利近期的调研显示，2026年的出货里，半尺寸机型（比如宇树G1）约占70%，全尺寸操作型只占30%。出货量靠便宜机器撑着，收入靠贵机器撑着，不同单价的产品被算进了同一个“出货量”。
“会干活的”，收入不到一成五
行业走到哪一步了，看场景比看数量更清楚。
沙利文把应用场景分成四类：娱乐表演、科研教育、数据采集、生产作业。2026年上半年，娱乐表演仍是体量最大的一块，科研教育小幅回落，数据采集大体持平，只有生产作业在收入上明显上量，半年收入0.78亿美元，比2025年全年的0.47亿美元还多65.9%。但0.78亿美元放进行业5.4亿美元的半年收入里，占比不到15%。
按照沙利文商业化成熟度的五个阶段，生产作业要经过技术展示、客户试点、付费部署、常态化运营、规模化复制这五步。大多数订单还停在客户试点和付费部署之间，走到常态化运营的很少。
9月的几条朋友圈把这个话题掀到了台面上。9月1日上市的梅卡曼德，其创始人邵天兰连发多条朋友圈，点名批评行业里的“攒局型”企业：把机器人卖给地方政府建的数采中心，再把数据买回来，“钱在‘自己人’手里转一圈，账面收入就‘做’出来了”。他在评论区直接点了银河通用的名，限定词是“北京、上海、名气很大、估值很高、上过春晚”。银河通用回了句“不参与无谓的口水战”，两天后发布声明称网络流传的相关内容均属不实信息，并已向公安机关报案。
谁对谁错，眼下没有审计结论。但这场争论点出的问题和IDC、沙利文的报告咬得很紧：一边统计出货，六成九来自科研、表演和数采；一边统计收入，真正干活的作业场景不到一成五。
数采中心本身不是坏事。具身智能训练需要千万小时级的真机数据，而全球高质量数据只有数十万小时级，缺口约95%，训练场填的正是这个坑。据信通院统计，截至2026年年中，国内建成投用的具身智能训练场已超70家，在建或规划的还有46家。
问题出在交易结构上。数采中心采购机器人，但采购方可能与厂商存在股权关联，交易在关联方之间完成，而不是来自真实终端需求。灵御智能CEO金戈9月接受媒体采访时称，灵御智能的TA系列机器人售价11.99万至19.99万元，在跨区域远程操作场景下，单台每月可产生约4000元人力套利空间，按双班倒运营则翻倍，据此测算，客户可单年回本；单年数据采集综合成本可控制在30万元以内，折合每小时约100至150元。
金戈算的是客户侧的回本账，这个账成立的前提是终端客户有真实、持续的使用需求。如果买方本来就是股东的关联方，产品是否真正进入场景发挥作用恐怕就得打个问号了。
根据公开信息，2026年上半年披露的人形与具身机器人中标项目共218个、总金额17.23亿元，其中工业及技术型企业相关项目仅占21.1%（教育科研占55.5%，政府及公共平台占20.6%，其他占2.8%）。该统计不含K12小型桌面设备、工程建设类项目及10万元以下小额租赁项目，且“中标”不等于交付验收或确认收入。
需要说明的是，沙利文这份报告在统计收入时，剔除了厂商内部交易和不代表第三方真实需求的关联交易，口径对标IPO审计。也正因为如此，沙利文统计的生产作业收入只有0.78亿美元，占比不到15%——挤掉水分之后，真实第三方需求的天花板就摆在这里。
不同口径，撑起不同的故事
即便不去看这些台面上的问题，人形机器人的统计数据就大有“文章”可做。
据IDC的数据，2026年上半年全球出货接近2.5万台，同比增长432.1%；市场规模7.4亿美元，同比增长322.7%，单台均价约2.96万美元。而沙利文的数据显示，上半年全球卖了3.6万台，收入5.4亿美元，单台约1.5万美元。同一个“全球人形机器人市场”，出货量差了四成，单价差了近一倍。主要差异是口径，IDC算的是“人形机器人”整机，沙利文把中尺寸、小尺寸、遥控机型也算了进去，同时剔除了关联交易。低价机型一进来，均价立刻被拉低。
全年预测的口径更乱。工信部说的是整机产量，“有望突破10万台”；摩根士丹利说的是出货，把国内预测从2.8万台上调到5万台。一个说产量、一个说出货，本来就不该直接比较，但这些数字经常被并排引用，去论证同一个结论。
同一个半年、同一个市场，两份报告能算出一倍的均价差；到了全年预测，不同机构之间差距巨大。这不是统计技术问题，而是行业还没有一套被共同接受的标准，数据由各家自己定义，口径之差，就是讲故事的空间。
谁能率先盈利？
白马认为，现阶段看人形机器人行业，“出货量”是最容易骗人的指标。这个数字把入门产品和全尺寸高端机型加在一起，算出来432.1%的增速很“热闹”，却看不出谁在赚钱或谁最有希望赚钱。
因此，要看清人形机器人行业，有三个更实在的口径。
第一，全尺寸占多少？摩根士丹利判断，全尺寸操作型的出货占比会从2026年的30%升到2027年的50%、2028年的70%。这个比例真走上去，才能说明机器人真的“能干活”了。沙利文的数字已经在往这个方向指，具身智能机型只占出货的三成左右，却拿走了83%的收入。
第二，场景应用走到哪一步？按照沙利文商业化成熟度的五个阶段，现在大部分订单还卡在客户试点和付费部署；什么时候能看到一批企业稳定跑到常态化运营，商业化才算真正落地。
第三，单台价格往哪走？当前，均价变化主要来自结构，不是机器本身便宜了两成，未来真正值得看的是同类机型的价格有没有往下走。如果连全尺寸这种高价品类也开始降，说明竞争的焦点已经从“能不能做出来”转到“多少钱一台”。不过，在一个还没跑出盈利样本的行业里打价格战，才是要警惕的信号。
IDC在报告里把2030年的预测上调到“全球出货超过75万台”，比上次高了约50%。这个数字能否兑现，取决于上面三个指标未来怎么走。回到当下，出货量涨四倍，盈利还是很难，而这恰恰是决定行业走向的关键问题。
文章内容和观点仅供参考，不构成投资建议。投资有风险，决策需谨慎。
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 Zheng, White Horse Business Review
At the end of September, the market research institutions Sullivan and the IDC reported, after publish, that both judged that the industry was quenching, but that the calibre was 40 per cent different: IDC counted nearly 25,000 global deliveries in the first half of the year, an increase of 432.1 per cent over the same period; and Sullivan statistics, 36,000 in the same period, generated $540 million, which was more than 2025.
The account book of the robotics company is another picture, with the top income share being the top one, with a loss of $339 million in the first half of the year, and another head company, Woo Tree Technology, with more than 5,500 outgoings in 2025, with a net interest rate of about 34.8 per cent, including profits from four full-sized robotic operations, whose main strength is the G1 type of tens of thousands of dollar-sized half machine.
Your machine has raised 80 per cent of the revenue of the industry, but has not yet made profits for anyone.
Most of the 25,000 deliveries in the first half of the year consisted of “cheap” items with a unit price of less than $30,000, with a much smaller number of fully capable and expensive machines. This explains the continuing surge in the size of the industry and the fact that businesses do not earn money.
I'm losing money.
In the first half of this year, Sullivan placed income-based seats on the global human robotics enterprise, the top five being Chinese companies: Euphoria, Ji-won, Woo-ki, Galactica, and Convergence. The first two were selected, with 16.5 per cent of the total human robot income and 26 per cent of the total smart track income.
Full-sized intelligence is the highest threshold in the Sullivan definition: non-controlled, non-pre-programmed, above 160 cm height, with a physical capacity usually above 200 TOOPS. According to Sullivan, in 2025, the total number of human robots selected was 13838, of which only 1,079 were fully intelligent, or less than 10 per cent. In the first half of 2026, 921 full-size smarts were sold, while the total sale of corporate human robots during the same period was 16123, further reducing the share of full-sized smarts to 5.71 per cent. According to financial data, there have been continuing losses in the last few years and no profit in the first half of the year.
The situation is slightly better, with more than 5,500 deliveries in 2025, approximately 1.7 billion yuan in revenue, and about 34.8 per cent net interest rate withheld. The main force is a half-size machine such as G1, with clients based on scientific education.
In fact, the two families follow the same path, which is done on a non-full-size machine. The difference is that Woo-hee has earned money and has not yet made his choice. The difference is not in the size of the product, but in the mode - Woo-tree sells a standard product for a single shipment, with a quick return; a fully-sized machine of choice enters the factory, enters the scene, has a long delivery cycle and has high service costs. The most “value” track in the full size industry has not run out of profitable sample.
It's cheap and it's expensive.
Based on IDC data, the average unit price dropped from $37,000 in the same period last year to $296 million in the first half of the year, but this change was mainly due to structural, cheap machine-type discharges, which brought down the weighted average price, not every machine by 20 per cent. The growth in sales of 432.1 per cent was much higher than the growth in income of 322.7 per cent, so the increase in the discharge was concentrated in the low-price machine type.
According to Salivin, the smart-type (machines capable of self-perception, decision-making, work) sold 0.88 million units in 2025, earning $375 million, or 75 per cent of industry revenues; in the first half of 2026, it sold 105,000 units, earning $448 million, or 83 per cent. During the same period, the pre-programming and remote control models were sold at a steady level and income was still going down.
The price is set at two feet. By size, in the first half of 2026, 0.84 million full-size machines were sold — only about 20 per cent of the sales volume — while $408 million was taken from them, representing more than 75 per cent of industry revenues; 197,000 small-size machines were sold, accounting for more than half of the deliveries and only $19 million.
In terms of intelligence, the highest threshold is full-size smart (non-controlled, non-pre-programmed) with only 0.58 million units sold for six months, with an income of $320 million; compared with the figures for the first half of the year and the year as a whole (0.53 million, $260 million), less than 10 per cent more sold, while income is almost one quarter higher and unit prices are moving up.
The quantitative contrast is more direct. Recent studies by Morgan Stanley show that in 2026, half-size machines (e.g., Wootree G1) accounted for about 70 per cent of deliveries, and only 30 per cent of full-size operating models. The output is sustained by cheap machines, the income by expensive machines, and the product at different unit prices is counted into the same “output”.
Work, less than 15%.
Where is the industry? The scene is clearer than the number.
Sullivan divided the applications into four categories: entertainment performances, scientific education, data collection, and productive work. In the first half of 2026, recreational performances continued to be one of the largest in volume, with a small fall in scientific education, with data collection generally flat, with only productive operations earning a significant income of $78 million in six months, or 65.9 per cent more than the $470 million recorded throughout 2025. However, US$ 78 million, or less than 15 per cent of the US$ 540 million in semi-annual income, was released.
In accordance with the five stages of the commercial maturity of Sullivan, the production process is to be technically demonstrated, the client pilot, the paid deployment, the normalization operation, and the scaling up and replication. Most orders are also between client pilots and paid deployments, and very few go to normal operations.
A few friends in September took this topic to the surface. Mercamand, listed on September 1st, whose founders, Shao Tian Lan, issued several circles of friends calling for criticism of the “salvation” enterprises in the industry: Selling the robots to the local government's digital centers and buying back the data, “The money is in the hands of his own people and the book money is `made'”. He directly named Galaxy in the comment area with the qualifier “Beijing, Shanghai, high reputation, high valuation, spring night”. In response to the phrase “not to engage in unnecessary saliva warfare”, Galaxy stated two days later that the contents of the network were false and reported to the public security authorities.
There are no audit findings at this time. But the issues raised by this debate and the IDC, Sullivan's reports are very tight: 60 percent of the products are from scientific research, performances and measurements; and less than 15 percent of the work is actually done on the basis of income.
The Digital Centre itself is not a bad thing. Smart training requires real time data of tens of thousands of hours, while high-quality data are available globally for hundreds of thousands of hours, with a gap of about 95 per cent, and the site fills the pit. It is believed that by mid-2026, more than 70 smart training sites had been built and 46 were under construction or planning.
The problem is the structure of the transaction. The centre procures robotics, but the buyer may have an equity interest in the manufacturer and the transaction is completed between the related parties rather than from the real end demand. In an interview with the media in September, the Smart CEO Kingo stated that the smart TA series of robots sold between $119.9 million and $199 million, and that under the trans-regional remote operating scenario, a single station could generate approximately $4000 per month of human arbitrage, double double the double shift, which would allow customers to return the book in a single year; the combined cost of single-year data collection could be contained within $300,000 and equal to about $100 to $150 per hour.
Kingo calculates a client-side return account, which is established on the assumption that the end-user has a real and ongoing demand for use. If the buyer was a related party to the shareholders, the question mark would have to be given as to whether the product actually entered the scene.
According to publicly available information, the number of human and robotic winnings in the first half of 2026 was 218 with a total value of RMB 1,723 million, of which only 21.1 per cent were related to industrial and technological enterprises (58.5 per cent for educational research, 20.6 per cent for government and public platforms and 2.8 per cent for others). The statistics do not include K12 small desktop equipment, engineering construction type projects and small lease items of less than $100,000, and the “success mark” does not equal the delivery of acceptance check or the recognition of revenue.
It should be noted that the Sullivan report, in the context of its statistical revenues, excludes transactions within the manufacturer and related transactions that do not represent the real needs of third parties. That is also why the income from production operations, as measured in Sullivan, is only US$ 78 million, or less than 15 per cent — the ceiling for real third-party demand is here when water is squeezed out.
Different calibres, different stories.
Even without looking at the issues on the stage, statistics on human robots are available for “article”.
According to IDC, global deliveries in the first half of 2026 were close to 25,000 units, an increase of 432.1 per cent over the same period; the market size was $740 million, an increase of 322.7 per cent over the same period, and the average price per station was approximately $296 million. In the first half of the year, according to Sullivan's data, 36,000 units were sold globally, generating $540 million and about $15,000 per station. The same “global human robotic market”, with 40 per cent of the output and nearly double the unit price gap. The main difference is calibration, i.e., the IDC is a “human robot” complex, and Sullivan has also included medium-sized, small-sized, remote-controlled, excluding related transactions. As soon as the low price machine came in, the average price was lowered.
The projections for the entire year are even more confusing. The Ministry of Industry and Communications is talking about the whole machine production, “promising to exceed 100,000”, while Morgan Stanley is talking about the delivery of goods, bringing the domestic forecast up from 28,000 to 50,000. One statement of production and another of goods should not be directly compared, but these figures are often quoted in parallel to the same conclusion.
In the same half-year and in the same market, the average price differential can be doubled in two reports; by year-round projections, the gap between institutions is large. This is not a matter of statistical technology, but of industry, which does not have a commonly accepted set of standards, and where data are defined by individual countries, the difference in calibre is the space for storytelling.
Who's gonna take the lead?
According to the White Horse, “outload” is the most easily deceptive indicator of the human robot industry at this stage. This figure, which combines entry products and full-size high-end machines, shows that the growth rate of 432.1 per cent is “hot”, but does not show who is making money or who is most likely to make it.
So, depending on the human robotics industry, there are three more solid calibres.
Number one, what's the full size? Morgan Stanley concluded that the full-size operating-type output would rise from 30 per cent in 2026 to 50 per cent in 2027 and 70 per cent in 2028. And that's what makes robots work. Sullivan's figures are already pointing in this direction, with smart machines, which account for only about 30 per cent of deliveries, taking 83 per cent of revenues.
Second, how far did the scene go? In accordance with the five stages of the commercial maturity of Sullivan, most of the orders are now in the client pilot and paid deployment; commercialization will be a real landing when a group of businesses can be seen to stabilize to normal operations.
Third, where's the single price? At present, the change in average prices is mainly structural, not 20 per cent cheaper than the machine itself, and what is really worth looking at in the future is whether the price of the same type of machine has gone down. If even the high-price class of full size is beginning to fall, it suggests that the focus of competition has shifted from “can or not” to “how much to one”. However, price fighting in an industry that has not yet run out of profit-making sample is a sign of vigilance.
In its report, IDC moved its 2030 projections upwards to “over 750,000 global deliveries”, about 50 per cent higher than the previous one. This figure depends on the way forward for the three above indicators. Returning to the present time, the volume of deliveries has increased fourfold and profitability remains difficult, which is the key issue in determining the direction of the industry.
Articles and opinions are for reference purposes only and do not constitute investment recommendations. Investment is risky and decision-making requires caution.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.