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。
At the end of September, market research firms Frost & Sullivan and IDC released humanoid robot reports in succession. Both judged the industry is scaling, but their figures differed by 40%: IDC counted nearly 25,000 units shipped globally in the first half, up 432.1% year-over-year; Frost & Sullivan counted 36,000 units for the same period, with revenue of $540 million, exceeding all of 2025 in half a year. The books of robot companies tell a different story. Unitree, which ranks first in revenue share, lost 339 million yuan in the first half; another leading company, Unitree Robotics, shipped over 5,500 units in 2025 with a non-GAAP net margin of about 34.8%, including quadruped robot business revenue and profit, with humanoid robot主力 being half-size models like the G1 priced in the tens of thousands of yuan. Expensive machines支撑 over 80% of industry revenue but have not yet支撑 profit for anyone. Of the 25,000 units shipped in the first half, most are "cheap goods" priced under $30,000; truly capable, expensive full-size machines are far fewer. This explains why industry scale keeps surging while companies cannot make money.
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。
In the first half of this year, Frost & Sullivan ranked global humanoid robot companies by revenue, with the top five all Chinese: UBTech, AgiBot, Unitree, Galbot, and Leju. UBTech took two firsts: 16.5% overall humanoid robot revenue market share, and 26% in the full-size embodied intelligence track. Full-size embodied intelligence is the highest-threshold category in Frost & Sullivan's definition: non-teleoperated, non-preprogrammed, over 160 cm tall, with onboard compute typically above 200 TOPS. According to Frost & Sullivan, UBTech's total humanoid robot sales in 2025 were 13,838 units, of which full-size embodied intelligence was only 1,079 units—less than 10%. In the first half of 2026, full-size embodied intelligence sold 921 units, while the company's total humanoid robot sales were 16,123 units, with full-size embodied intelligence share further dropping to 5.71%. From financial data, UBTech has持续 lost money over the past few years and remained unprofitable in the first half of this year.
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。
Unitree's situation is somewhat better: over 5,500 units shipped in 2025, revenue of about 1.7 billion yuan, non-GAAP net margin of about 34.8%, with half-size models like the G1 as主力 and customers mainly in research and education. In fact, both take the same path: using non-full-size models to build volume. The difference is that Unitree has already made money; UBTech has not. The difference is not in product size but in model—Unitree sells standard products shipped on order, with fast payment collection; UBTech's full-size models must enter factories and scenarios, with long delivery cycles and high service costs. The full-size track, the most "valuable" in the industry, has not yet produced a profitable sample.
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。
According to IDC data, average unit price dropped from $37,000 in the same period last year to $29,600 in the first half of this year, but this change mainly comes from structure—cheap models scaling pulled down the weighted average, not that every machine became 20% cheaper. Sales growth of 432.1% far exceeds revenue growth of 322.7%, so this round's incremental volume is concentrated in low-priced models. According to Frost & Sullivan data, embodied intelligence models (machines that can perceive, decide, and work on their own) sold 8,800 units in 2025 with $375 million in revenue, accounting for 75% of industry revenue; in the first half of 2026 they sold 10,500 units with $448 million in revenue, with share rising to 83%. Meanwhile, preprogrammed and teleoperated models had flat sales, with revenue still declining.
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。摩根士丹利近期的调研显示，2026年的出货里，半尺寸机型约占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。
The volume-price inversion holds under both measures. By size, full-size models sold 8,400 units in the first half of 2026—only about 20% of sales but taking $408 million in revenue, over 75% of industry revenue; small-size models sold 19,700 units—over half of shipments—with only $19 million in revenue. By intelligence level, the highest-threshold full-size embodied intelligence sold only 5,800 units in half a year with $320 million in revenue; comparing first-half 2026 data with full-year 2025 (5,300 units, $260 million), volume rose less than 10% while revenue rose nearly 25%, with unit prices trending up. Recent Morgan Stanley research shows that in 2026 shipments, half-size models account for about 70%, full-size operational models only 30%. Shipment volume is支撑 by cheap machines, revenue by expensive machines, and products at different price points are counted into the same "shipment" figure.
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。
Where the industry stands is clearer from scenarios than from numbers. Frost & Sullivan divides application scenarios into four categories: entertainment performance, research and education, data collection, and production operations. In the first half of 2026, entertainment performance remained the largest segment, research and education slightly declined, data collection was roughly flat, and only production operations showed notable revenue growth—half-year revenue of $78 million, 65.9% more than full-year 2025's $47 million. But $78 million in the industry's $540 million half-year revenue is less than 15%. According to Frost & Sullivan's five stages of commercialization maturity, production operations must pass through technology demonstration, customer pilot, paid deployment, normalized operation, and scaled replication. Most orders are still between customer pilot and paid deployment, with few reaching normalized operation.
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。
A few social media posts in September brought this topic to the surface. Mech-Mind, which listed on September 1, saw founder Shao Tianlan post multiple messages criticizing "assemblage-type" companies in the industry: selling robots to data collection centers built by local governments, then buying the data back, with money circulating among "one's own people" to "create" book revenue. He directly named Galbot in the comments, with qualifiers "Beijing, Shanghai, very famous, very high valuation, been on Spring Festival Gala." Galbot replied it would not participate in pointless口水 wars, and two days later issued a statement saying related content circulating online was untrue and it had reported to public security. Who is right or wrong has no audit conclusion yet. But the issue this debate raised is tightly tied to IDC and Frost & Sullivan reports: on one hand, counting shipments, 69% come from research, performance, and data collection; on the other hand, counting revenue, truly working operation scenarios are less than 15%.
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。
Data collection centers themselves are not bad. Embodied intelligence training needs tens of millions of hours of real-machine data, while global high-quality data is only in the hundreds of thousands of hours, a gap of about 95%. Training grounds fill exactly this hole. According to CAICT statistics, as of mid-2026, over 70 embodied intelligence training grounds have been built and put into use in China, with 46 more under construction or planned. The problem lies in the transaction structure. Data collection centers procure robots, but the buyer may have equity ties with the manufacturer, with transactions completed between related parties rather than from real end demand. Lingyu Intelligent CEO Jin Ge told media in September that Lingyu's TA series robots are priced at 119,900 to 199,900 yuan, and in cross-regional remote operation scenarios, a single unit can generate about 4,000 yuan per month in labor arbitrage space, doubling under two-shift operation, by which calculation customers can recoup costs in a single year; single-year data collection comprehensive cost can be controlled within 300,000 yuan, equivalent to about 100 to 150 yuan per hour.
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。
Jin Ge calculated the customer-side payback account, and the premise for this account is that end customers have real, sustained usage needs. If the buyer is already a related party of shareholders, whether the product truly enters the scenario and plays a role is questionable. According to public information, 218 humanoid and embodied robot bidding projects were disclosed in the first half of 2026, totaling 1.723 billion yuan, of which industrial and technical enterprise-related projects accounted for only 21.1% (education and research 55.5%, government and public platforms 20.6%, others 2.8%). This statistic excludes K12 small desktop devices, engineering construction projects, and small leasing projects under 100,000 yuan, and "winning the bid" does not equal delivery acceptance or confirmed revenue. It should be noted that Frost & Sullivan's report, when counting revenue, excludes internal manufacturer transactions and related-party transactions that do not represent real third-party demand, with standards aligned to IPO auditing. Precisely because of this, Frost & Sullivan's production operation revenue is only $78 million, less than 15%—after squeezing out水分, the ceiling of real third-party demand is right here.
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。
Even without looking at these surface issues, humanoid robot statistics have plenty of "room for interpretation." According to IDC data, global shipments in the first half of 2026 were nearly 25,000 units, up 432.1% year-over-year; market size $740 million, up 322.7%, average unit price about $29,600. Frost & Sullivan data shows 36,000 units sold globally in the first half, revenue $540 million, about $15,000 per unit. The same global humanoid robot market has a 40% difference in shipments and nearly double the unit price. The main difference is scope: IDC counts humanoid robot complete machines, while Frost & Sullivan includes medium-size, small-size, and teleoperated models, while excluding related-party transactions. Once low-priced models are included, the average price is immediately pulled down. Full-year forecast scopes are even messier. MIIT talks about complete machine output, "expected to exceed 100,000 units"; Morgan Stanley talks about shipments, raising its China forecast from 28,000 to 50,000 units. One says output, one says shipments—they should not be directly compared, but these numbers are often cited side by side to argue the same conclusion. Same half-year, same market, two reports can produce a doubling in average price difference; for full-year forecasts, different institutions differ enormously. This is not a statistical technique problem, but that the industry still lacks a commonly accepted standard. Data is defined by each party, and the difference in scope is the space for storytelling.
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。要看清这个行业，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。
At this stage, looking at the humanoid robot industry, "shipment volume" is the most deceptive metric. This number adds entry-level products and full-size high-end models together, producing a lively 432.1% growth rate, but it does not show who is making money or who is most likely to. To see the industry clearly, there are three more substantive measures. First, what share is full-size? Morgan Stanley judges that full-size operational models' shipment share will rise from 30% in 2026 to 50% in 2027 and 70% in 2028. Only when this ratio truly rises can it show robots can really work. Frost & Sullivan's numbers already point this way: embodied intelligence models account for only about 30% of shipments but take 83% of revenue. Second, what stage has scenario application reached? According to Frost & Sullivan's five stages of commercialization maturity, most orders are still stuck between customer pilot and paid deployment; when a batch of companies can be seen steadily reaching normalized operation, commercialization will have truly landed. Third, where are unit prices heading? Currently, average price changes mainly come from structure, not that machines themselves became 20% cheaper. What is truly worth watching in the future is whether prices for the same model type are declining. If even the high-priced full-size category starts dropping, it indicates the focus of competition has shifted from "can it be made" to "how much per unit." However, a price war in an industry that has not yet produced a profitable sample is the signal to be wary of. IDC raised its 2030 forecast to "over 750,000 units shipped globally," about 50% higher than last time. Whether this number materializes depends on how the above three indicators move. Back to the present: shipments up fourfold, profitability still difficult—and this is precisely the key question determining the industry's direction.
