2026年的具身智能，热和冷是同一天发生的。热的一面，上半年国内融资约935亿元，同比涨了约5倍，估值过百亿的企业增加到20家；IDC的数据更猛，全球人形机器人出货接近2.5万台，其中中国超过1.9万台，吃掉全球约77.9%的份额。钱还在往里涌，只是涌得比前两年更挑。
Embodied intelligence in 2026 ran hot and cold on the same day. On the hot side, domestic financing hit about 93.5 billion yuan in the first half, up roughly fivefold, with 20-plus companies above the 10-billion valuation mark; IDC's data is sharper—global humanoid shipments near 25,000, over 19,000 of them in China, about 77.9% of the world. Money still flows in, just more selectively than before.
冷的那面，信通院在《具身智能发展报告（2026年）》里给了一句刺眼的话：出货不等于上岗，相当比例设备月均开机天数不足10天，订单含金量低。文娱商演加科教数采两项就占了应用结构的六成多，导览导购约19%，真正落在智能制造和仓储物流的加起来才约18%。
On the cold side, CAICT's "Embodied Intelligence Development Report (2026)" delivered a stinging line: shipment is not employment, a sizable share of devices boot fewer than 10 days a month, and order quality is low. Entertainment shows plus education-data collection alone take over 60% of the application mix, guiding and retail about 19%, and real manufacturing and warehousing together just 18%.
投资人的问题也变朴素了。以前比谁的Demo走得稳，今年比谁能干得久。10月22日科大讯飞1024开发者节上，聆动通用没有再放一段演示视频，而是拿出一条能被订单、出货和回本周期校验的全栈路径：LDU环境式数据采集、iFlyBot-WAM工业动作模型、LindenBot-Harness执行组织系统，外加LDB具身智能通用机器人。
Investors' questions got plainer. Last year it was whose demo walked steadier; this year it is who can work longer. At iFlytek's 1024 Developer Festival on October 22, Lingdong General did not show another demo video—it put forward a full-stack path verifiable by orders, shipments, and payback: the LDU ambient data-collection kit, the iFlyBot-WAM industrial action model, the LindenBot-Harness execution system, and the LDB general-purpose embodied robot.
回本周期是这个路径最硬的证据。聆动说部分场景能让客户两年内收回成本，行业普遍要两到三年，非标产线三到五年。码货和供件方案已经在头部物流企业跑通，仓储分拣还在POC，今年LDT和LDB两款产品合计出货大约200台，合同体量以亿级和数亿级为主。能持续交付，账本才立得住。
Payback is the hardest proof of that path. Lingdong says some scenarios let customers recover cost in two years, versus the industry's two-to-three and three-to-five for non-standard lines. Its palletizing and feeding solutions are already live at top logistics firms; warehousing sorting is still in POC; this year LDT and LDB together ship about 200 units, with contracts in the hundreds-of-millions range. Only steady delivery makes the ledger stand.
四层架构把量产账本拆到了工程层级。数据层用背包式、可换电的LDU走到哪采到哪，压低复制成本；模型层让WAM在动作生成前先做安全仲裁，把"敢不敢让机器人无人值守"前置；执行层用Harness智能体统一调度，换产线不重编；本体层靠全国首个CR认证工规级机器人LDB和CE-MD资质，拿到投标门票。
The four-layer architecture breaks the mass-production ledger down to engineering tiers. The data layer uses a backpack-style, swappable-battery LDU that collects wherever it walks, cutting replication cost; the model layer lets WAM arbitrate safety before action, front-loading "dare we run it unattended"; the execution layer uses the Harness agent to schedule uniformly, so a line change needs no recode; the body layer rides the nation's first CR-certified industrial-grade robot LDB and CE-MD, buying the bid ticket.
信通院把现在定义为期望膨胀期，预计未来两三年完成首轮行业出清。出清的标准不是融资额、不是估值、甚至不是出货量，而是设备在真实场景持续创造价值的能力。站在分水岭两侧的名单正在变短，决定去留的，是谁的账本更早算得平。
Capital turned pragmatic too. In the first half Sequoia China joined 20 rounds here across 14 firms, Hillhouse 18; by Q3 Sequoia fell to 8 and Hillhouse to 4. Money did not shrink—it moved from casting a wide net to precisely backing top hardcore teams; Demos and slides no longer win approval.
资本也跟着变务实。上半年红杉中国在这一赛道参与20起融资、投了14家，高瓴创投参与18起；到了三季度，红杉降到8起、高瓴4起。钱没有变少，只是从广泛撒网转向精准打头部硬核团队，Demo和PPT换不来过会了。
CAICT calls this the inflation-of-expectations phase and expects the first industry shakeout within two to three years. The yardstick is not funding, valuation, or even shipments, but a device's ability to keep creating value in real scenes. The roster on either side of the watershed is shrinking; what decides stay or go is whose ledger balances first.
