近段时间，DeepSeek被曝拟领投Winverse AI种子轮，该企业成立仅数月，创始人胡瀚此前在腾讯负责混元多模态和具身基础模型。在这之前，DeepSeek已经在8月以1.41亿元参与宇树科技的战略配售，锁定期36个月，双方签署了双向优先采购备忘录。OpenAI首席执行官山姆·奥特曼则在播客对外披露，OpenAI正在布局人形机器人研发赛道，相关项目将把机器人智能系统研发作为核心攻关方向，他预判OpenAI有望在2027年推出效果亮眼的演示样机。智谱、Kimi等大模型厂商也在布局具身智能。
Recently, DeepSeek was reported planning to lead Winverse AI's seed round; the company was founded only months ago, with founder Hu Han previously responsible for Hunyuan multimodal and embodied foundation models at Tencent. Before that, DeepSeek had already participated in Unitree's strategic placement with 141 million yuan in August, with a 36-month lock-up, and the two signed a bidirectional priority procurement memorandum. OpenAI CEO Sam Altman disclosed on a podcast that OpenAI is布局 humanoid robot R&D, with related projects making robot intelligence system R&D a core focus; he predicts OpenAI could launch an impressive demo prototype by 2027. Large model makers including Zhipu and Kimi are also布局 embodied intelligence.
最顺手的解读是大模型厂商正在集体入局具身智能。但是，DeepSeek选择的方式是投资，不碰本体、不建产线；OpenAI从投资别人转向自研，自己造执行器、自建数据采集体系、自己管供应链。入局与否之外，商汤的数据提供了另一个观察视角。商汤把GPT-6 Astra接入RoboDojo测试平台，跑了一组机器人操作任务。结果很有意思，9个此前连续5次全部失败的任务，在允许模型用更多步数去尝试之后，变成了5次全部成功，成功率推到81.9%。但测试还记录了另一组数字，完成这些任务，叠碗约131分钟、19美元，笔装笔筒约263分钟、55美元。放在真实场景里，它想要替代的是一个工人原本几分钟内就能完成的工作。
The most convenient interpretation is that large model makers are collectively entering embodied intelligence. But DeepSeek chose investment—not touching bodies, not building production lines; OpenAI shifted from investing in others to in-house development—building its own actuators, its own data collection system, its own supply chain. Beyond entry itself, SenseTime's data offers another observation angle. SenseTime plugged GPT-6 Astra into the RoboDojo testing platform and ran a set of robot manipulation tasks. The results are interesting: 9 tasks that previously failed 5 consecutive times all became 5-time successes after allowing the model more steps to try, pushing success rate to 81.9%. But the test also recorded another set of numbers: completing these tasks took about 131 minutes and $19 for stacking bowls, and about 263 minutes and $55 for pen-in-holder. In real scenarios, what it wants to replace is work a worker could complete in minutes.
机器人单次执行的成本是人工的数倍，耗时是人工的数十倍，这意味着，现阶段即使模型在测试环境里做到了81.9%的成功率，它在真实场景里每完成一次任务，付出的时间和金钱成本仍然高到几乎没有工厂或仓库愿意买单。那么，物理交互的瓶颈就不在于能力有没有突破，而是以什么价格、什么速度做。由此，成本结构上的数量级差距，也可以解释DeepSeek和OpenAI看似都在入局具身智能，却在起点处背道而驰。
Robot single-execution cost is several times that of manual labor, and time consumption is dozens of times—meaning at this stage, even if the model achieves 81.9% success in test environments, every task it completes in real scenarios costs time and money so high that almost no factory or warehouse is willing to pay. So the physical interaction bottleneck is not whether capability has broken through but at what price and speed. Thus, the order-of-magnitude gap in cost structure can also explain why DeepSeek and OpenAI, seemingly both entering embodied intelligence, diverge at the starting point.
DeepSeek两笔落子具身智能的投资，没有一笔指向机器人本体或真机数据的所有权。1.41亿元战略配售宇树，核心条款是双向优先采购，宇树采购模型训练服务时优先考虑DeepSeek，DeepSeek探索具身业务时优先考虑宇树。拟领投的Winverse AI，多位行业人士明确其更偏底层基础设施与模型加速，并非具身整机公司。DeepSeek用资本锁定了身体的接入权和大脑加速的接口，但没有为任何一条产线、任何一座数据采集工厂付固定成本。
DeepSeek's two embodied intelligence investments, neither points to robot body or real-machine data ownership. The 141 million yuan strategic placement in Unitree has core terms of bidirectional priority procurement—Unitree prioritizes DeepSeek when procuring model training services; DeepSeek prioritizes Unitree when exploring embodied business. The planned lead investment in Winverse AI—multiple industry figures clearly state it leans more toward底层 infrastructure and model acceleration, not an embodied complete machine company. DeepSeek used capital to lock in body access rights and brain acceleration interfaces but paid no fixed costs for any production line or any data collection factory.
OpenAI走的是另一条路。截至2026年10月，官网以robotics为关键词可搜出27个职位，今年5月这个数字还是11个，新增岗位横跨执行器、电机设计、制造、供应链和数据采集。其中，执行器岗位明确要求设计custom robotic actuators，覆盖电机、传动、传感和热管理，数据采集端则在招聘能把机器人、传感器、操作员界面和数据管线组合成production-ready系统的技术项目经理，跟踪利用率、节拍时间、吞吐量和停机时间。这几乎是一套工厂的生产指标。从招聘信息看，OpenAI在造为AI控制而设计的身体，同时想建一座为身体生产数据的工厂。
OpenAI takes another path. As of October 2026, searching "robotics" on its website returns 27 positions; in May this year the number was 11. New positions span actuators, motor design, manufacturing, supply chain, and data collection. Among them, actuator positions explicitly require designing custom robotic actuators covering motors, transmission, sensing, and thermal management; the data collection side is hiring technical project managers who can combine robots, sensors, operator interfaces, and data pipelines into "production-ready" systems, tracking utilization, cycle time, throughput, and downtime. This is almost a set of factory production metrics. From the recruitment information, OpenAI is building a body designed for AI control while wanting to build a factory that produces data for that body.
DeepSeek在语言模型上的优势是算法效率和推理成本控制，但这个优势在物理世界无法直接兑换，因为物理交互成本的瓶颈并非模型参数，是真机数据的采集和训练场的固定投入。根据行业公开估算，真机数据的单小时有效数据的采集成本在500至1000元，一座数千平方米的训练场投入数千万元至上亿元，设备、人力、场景建设和维护等固定成本，不会因为模型效率提升而自动摊薄。所以DeepSeek投宇树和Winverse，是用投资和采购协议这类可进可退的支出，替代产线和数据工厂的固定投入。
DeepSeek's advantage in language models is algorithm efficiency and inference cost control, but this advantage cannot be directly exchanged in the physical world because the physical interaction cost bottleneck is not model parameters but real-machine data collection and training ground fixed investment. According to public industry estimates, effective data collection cost for real-machine data is 500 to 1,000 yuan per hour; a training ground of thousands of square meters requires tens of millions to hundreds of millions in investment, with equipment, manpower, scenario construction, and maintenance fixed costs not automatically amortized by model efficiency improvements. So DeepSeek investing in Unitree and Winverse uses investment and procurement agreements—reversible expenditures—to replace production line and data factory fixed investment.
OpenAI的重资产投入同样是被成本结构推出来的。早前，OpenAI也曾经走轻资产路线，投资过Figure，签协议为Figure的人形机器人开发下一代AI模型。合作于2025年2月终止，Figure转向自研Helix，阿德科克公开说的理由是必须垂直整合机器人AI，不能外包AI。但更根本的教训来自数据，如果身体属于别人，模型就拿不到机器人在真实环境中产生的感知和操作数据。OpenAI多年前解散机器人团队时，给出的理由正是缺乏训练数据。Figure把门关上之后，OpenAI的位置变得很窄，继续做只出大脑的供应商，数据的接入权始终在身体方手里，要拿到物理数据的所有权，就需要自己亲自做。
OpenAI's heavy-asset investment was also pushed out by cost structure. Earlier, OpenAI也曾 took a light-asset route, investing in Figure and signing an agreement to develop next-generation AI models for Figure's humanoid robots. The cooperation terminated in February 2025; Figure turned to in-house Helix, with Adcock publicly stating the reason: "must vertically integrate robot AI, cannot outsource AI." But the more fundamental lesson comes from data—if the body belongs to someone else, the model cannot obtain perception and manipulation data generated by robots in real environments. When OpenAI disbanded its robotics team years ago, the stated reason was lack of training data. After Figure closed the door, OpenAI's position became narrow—continuing as a brain-only supplier leaves data access in the body party's hands; to obtain physical data ownership, it must do it itself.
所以，理解两家公司的战略差异，要看物理交互的固定成本这道门槛。DeepSeek没有选择承担自建身体的固定成本，进行了关系控制；OpenAI在关系控制里被切断数据接入，选择了承担固定成本、转向资产控制。二者具体实施的方向相反，但受到的约束同源。放眼整个市场，2026年上半年国内具身智能赛道融资约438亿元，其中大脑层算法与控制软件公司吸纳222.53亿元，占比50.8%，本体整机公司仅占12.8%。虽然资本在往大脑走，但物理交互的固定成本并没有因为资本流向而消失，不同能力结构的玩家，面对成本门槛各自选择了自己的站位。
So understanding the two companies' strategic differences requires looking at the physical interaction fixed cost threshold. DeepSeek chose not to bear the fixed cost of building its own body, opting for relational control; OpenAI, having data access cut off in relational control, chose to bear fixed costs and shift to asset control. Their specific directions are opposite, but the constraint is the same source. Across the market, 2026 first-half domestic embodied intelligence track financing was about 43.8 billion yuan, of which brain-layer algorithm and control software companies absorbed 22.253 billion, accounting for 50.8%, while body complete machine companies accounted for only 12.8%. Although capital flows toward brains, physical interaction fixed costs have not disappeared because of capital flow; players with different capability structures each chose their positions facing the cost threshold.
按照谁承担了物理交互的固定成本、承担到什么程度，市场上存在四类位置。第一类几乎不承担物理交互固定成本，做模型与接口的浅层卡位。阿里Qwen-Robot、腾讯Tairos开放平台、字节投资自变量机器人，都属于这一类。第二类不拥有完整闭环，但以数据和场景为切口分担部分固定成本。京东做数据采集，美团投资加自建场景。第三类承担物理交互的全部固定成本，并在闭环内部积累壁垒。宇树、智元、优必选属于此类。第四类不拥有任何一个闭环，但控制闭环必须经过的底层基础设施。英伟达GR00T/Jetson体系、华为CloudRobo平台，以及阿里云和腾讯云在算力、仿真和训练框架层的布局，属于这一类。
By who bears physical interaction fixed costs and to what degree, there are four types of positions in the market. The first type barely bears physical interaction fixed costs, doing shallow positioning in models and interfaces—Alibaba Qwen-Robot, Tencent Tairos open platform, and ByteDance investing in AgiBot all belong here. The second type does not own a complete closed loop but shares部分 fixed costs through data and scenario entry points—JD.com doing data collection, Meituan investing plus building scenarios. The third type bears all physical interaction fixed costs and accumulates barriers within the closed loop—Unitree, AgiBot, and UBTECH belong here. The fourth type owns no closed loop but controls underlying infrastructure that closed loops must pass through—NVIDIA's GR00T/Jetson system, Huawei CloudRobo platform, and Alibaba Cloud and Tencent Cloud's布局 in compute, simulation, and training framework layers belong here.
当物理交互的成本曲线发生变化时，四类位置各自的优势能维持多久，判断标准不在于谁离物理闭环更近。第一类浅层卡位的优势，来自本体厂商仍然需要外部模型能力，但这个优势正在被头部本体厂商把大脑能力内化的趋势削弱。第二类数据与场景切口的优势，来自真机数据仍然不可替代，但这个优势正在被硬件成本下探部分缓解，合成数据的标注成本已从一帧十几元压到几毛甚至几分钱，触觉传感器单片价格从数万元至十万元级降至199元。第三类物理闭环内部的优势，来自宇树的真机数据、本体迭代经验和交付能力，但它的风险也最高，按部分报告口径，单台轮臂或双足人形机器人年综合运维成本约50万至70万元，高于多数岗位的人力成本。第四类底层基础设施的优势，建立在既有生态引力是否足够强的前提上，CUDA的历史已经证明，接口一旦成为默认路径，移除成本较高。
When the physical interaction cost curve changes, how long each of the four positions' advantages can last—the criterion is not who is closer to the physical closed loop. The first type's shallow positioning advantage comes from body manufacturers still needing external model capabilities, but this advantage is being weakened by leading body manufacturers internalizing brain capabilities. The second type's data and scenario entry advantage comes from real-machine data still being irreplaceable, but this advantage is being partially alleviated by hardware cost declines—synthetic data annotation cost has dropped from over ten yuan per frame to a few cents or even less; tactile sensor unit prices have dropped from tens of thousands to hundreds of thousands of yuan to 199 yuan. The third type's physical closed-loop internal advantage comes from Unitree's real-machine data, body iteration experience, and delivery capability, but its risk is also highest—by some report metrics, single wheeled-arm or bipedal humanoid robot annual comprehensive operations cost is about 500,000 to 700,000 yuan, higher than most positions' labor costs. The fourth type's underlying infrastructure advantage is built on whether existing ecosystem gravity is strong enough—CUDA's history has proven that once an interface becomes the default path, removal costs are high.
把当前产业阶段放进来，无论强如DeepSeek、OpenAI，还是快速发展的大模型厂商，都面临着共同的限定条件，物理交互的成本仍然高到难以支撑大规模商业化。在终局之前，大模型厂商在具身智能赛道上的位置，甚至要比它们自己以为得要窄些，也比任何控制权叙事所能描述的，更受制于物理世界本身的节奏。
Placing the current industry stage in context, whether strong as DeepSeek, OpenAI, or rapidly developing large model makers, all face a common constraint—physical interaction costs remain too high to support large-scale commercialization. Before the endgame, large model makers' positions in the embodied intelligence track are even narrower than they themselves think, and more constrained by the physical world's own rhythm than any "control" narrative can describe.
