10月30日，富士康董事长刘扬伟释放了两组放在一起看很有意思的信息。

一组是，公司正在讨论与OpenAI开展新的合作，正式内容要等细节敲定后再公布。另一组是，富士康已经在美国得州工厂部署超过10台人形机器人，还在扩建休斯敦的AI服务器制造能力。

这两件事看上去分属不同领域：前者是生成式AI，后者是工业自动化。

可它们最后会落到同一间工厂里。

富士康面对的已经不只是服务器订单

富士康是NVIDIA重要的AI服务器制造伙伴。路透社报道，其包含AI服务器在内的云网络产品，已经占到2026年第二季度收入的51%。

这意味着，公司眼前的问题发生了变化。

过去，代工厂最擅长的是把成熟产品快速复制出来。产线、工艺、测试、采购和交付，每一项都追求稳定。

AI服务器却没有那么“安静”。

机柜功率继续提高，液冷、供电、高速网络和现场测试互相牵连。客户芯片更新一次，结构、散热和装配流程都可能跟着调整。制造商不能等所有设计冻结之后再开始生产，它需要更早进入工程阶段。

富士康与OpenAI在2025年已经围绕数据中心机架、组件和硬件设计建立合作。此次所说的新合作尚未公开，但方向值得关注：富士康希望离AI系统的需求定义更近，而不只是等一份零部件清单。

这不是简单地“提高代工毛利”。更关键的是获取设计反馈、产线数据以及下一代基础设施的工程经验。

工业机器人的瓶颈，也恰好藏在工厂数据里

人形机器人发布会很容易拍出漂亮视频。工厂不看这个。

工厂要问得具体得多：

它能否连续工作八小时？换一个料箱要重新训练多久？手臂误差超过几毫米时谁来处理？停机一分钟会不会影响整条线？维护人员需要学多少新东西？

富士康现在部署的数量只有十余台，远谈不上规模化替代工人。可这批机器有一个普通机器人创业公司很难复制的条件：它们就在真实产线上工作。

抓取失败、螺丝角度、物料形变、灯光变化、人员经过，这些细节会不断产生新数据。机器人模型通过仿真学习基本动作，再回到现场修正。

这套过程的价值不在某一次演示，而在更新速度。

有工厂的人可以每天获得问题。有算力的人可以快速重训。有制造能力的人可以修改夹具、关节、传感器和产线设计。三者集中在同一家公司，机器人就有机会由“设备销售”变成持续改善的生产系统。

“制造AI服务器”和“用AI制造服务器”正在相遇

把富士康的两项业务画在一张图上，会看到一个相互加强的结构。

AI服务器订单带来工厂扩张，也带来算力、供电、散热和网络能力。机器人进入工厂，产生操作数据和异常案例。数据又可以用来训练机器人、检查质量、预测设备维护，并重新安排生产步骤。

于是，服务器不只是富士康交给客户的产品，也可能成为它改造自身工厂的工具。

这件事对NVIDIA同样有价值。它销售的不再只是GPU，还包括机器人基础模型、仿真软件、边缘计算设备和工厂开发工具。

OpenAI可能扮演什么角色，目前不能下结论。合理的观察方向包括数据中心硬件协同、工程设计辅助和生产流程中的智能体，但这些都需要等待正式公告。

把推测写成合同内容，会误导读者。

真正的商业价值在哪里

从投资角度看，工业AI的价值不能只用“替代多少工人”计算。

更现实的回报可能来自四处：

• 新产品换线时间缩短；
• 质量异常被更早发现；
• 高危、重复或高精度工序减少人工暴露；
• 工艺经验能在不同工厂之间复制。

这些收益散落在生产周期里，不会像聊天机器人订阅收入那样整齐。因此，制造企业需要先建立基线：当前换线花多少小时、缺陷发生在哪里、设备因何停机、人工返工占多少成本。

没有基线，机器人上线后只能展示“它会动”，无法证明它值多少钱。

中国制造企业可以先做什么

不用一开始就买人形机器人。

更稳妥的方式，是选择一个数据已经较完整、动作相对重复、错误成本可计算的工序。例如物料搬运、外观检查、标准化装箱或设备巡检。

接着做三件小事：

把正常动作和异常情况录下来；把停机、返工和人工耗时算清楚；要求供应商用同一组指标完成小范围测试。

评估周期可以设为四到八周。判断依据不是演示视频，而是干预次数、成功率、停机时间和维护工作量。

如果一套系统只能在供应商工程师驻场时运行，它还不是可复制的生产能力。

边界也很清楚

富士康当前部署的人形机器人数量有限。即使试验成功，也不代表所有制造岗位都会快速被机器人替代。

工厂环境比宣传片复杂。安全认证、设备寿命、维护成本、不同物料的适应能力，都会拖慢推广速度。

另一个风险来自客户集中度。AI服务器需求很旺，但基础设施投资具有周期性。客户放缓建设、改变芯片平台或调整数据中心架构，都会传到制造端。

因此，更准确的判断不是“富士康变成机器人公司”。

它正在利用自己拥有的服务器订单、真实工厂和供应链，把AI基础设施与工业机器人接到一起。若这套组合跑通，制造能力本身会成为AI模型的一部分。

这比发布一台会走路的机器人难得多。也更有商业价值。

本文依据截至2026年10月31日10:00公开可查资料撰写。富士康与OpenAI的新合作尚未正式公布，文中对可能商业方向的讨论属于分析，不代表双方已签署对应合同。文中不构成证券、投资或采购建议。

English Version

Foxconn Is Putting AI Servers and Robots in the Same Factory

On October 30, Foxconn Chairman Young Liu disclosed two developments that are best understood together.

The company is discussing a new collaboration with OpenAI, although the details have not been finalized or formally announced. At the same time, Foxconn has deployed more than ten humanoid robots at its Texas facilities and is expanding its AI-server manufacturing capacity in Houston.

One development belongs to generative AI. The other appears to belong to industrial automation.

In practice, they may converge inside the same factory.

Foxconn is dealing with more than server orders

Foxconn is a major manufacturing partner in NVIDIA’s AI infrastructure supply chain. According to Reuters, cloud and networking products—including AI servers—accounted for 51% of Foxconn’s revenue in the second quarter of 2026.

AI servers are not static products. Higher rack density affects power delivery, cooling, networking and testing. A change in the customer’s chip platform can force changes in mechanical design and assembly.

The manufacturer therefore needs to participate earlier in the engineering process. Waiting for a frozen bill of materials is no longer enough.

Foxconn and OpenAI had already agreed in 2025 to cooperate on data-center racks, components and related hardware. The scope of the newly discussed collaboration remains unknown. Still, Foxconn’s strategic intent is visible: it wants to move closer to the point where AI infrastructure requirements are defined.

The benefit is not limited to manufacturing margin. Early participation provides engineering feedback, production data and knowledge that can be reused in the next generation of systems.

The hardest robotics problems live in factory data

A humanoid robot can look impressive in a product demonstration. A factory uses a less forgiving test.

Can it work through a full shift? How often does a person need to intervene? How long does it take to adapt to a different container or component? What happens when lighting, material shape or line speed changes?

Foxconn’s current deployment is small. More than ten robots do not amount to mass adoption.

The interesting part is where those robots operate: real production lines.

Every failed grasp, misaligned screw, deformed package and unexpected human movement can become training material. A robot may learn basic skills in simulation, then improve through carefully controlled deployment on the shop floor.

Foxconn has three resources that many robotics startups struggle to combine: factories, computing infrastructure and manufacturing engineering. It can update the model, change the gripper and modify the production process without waiting for three separate organizations to coordinate.

That may shorten the distance between a prototype and a maintainable production system.

Manufacturing AI servers and manufacturing with AI are converging

AI-server orders support the expansion of factories, power systems, cooling capacity and high-speed networks. Robots deployed in those factories produce operating data. That data can be used to improve robotic control, quality inspection, maintenance and production planning.

The server is therefore more than a product shipped to a customer. It can also become part of the infrastructure Foxconn uses to redesign its own production.

NVIDIA benefits from the same shift. Its commercial position expands beyond GPUs into simulation software, robotic foundation models, edge computers and factory-development tools.

OpenAI’s role should not be overstated. Potential areas could include infrastructure engineering, design assistance or workflow agents, but none of these possibilities should be treated as confirmed contract terms.

Where the economic value may appear

Industrial AI should not be judged only by the number of workers a machine might replace.

The more practical sources of value include shorter changeover times, earlier defect detection, less human exposure to hazardous or repetitive work, and faster transfer of manufacturing knowledge between sites.

These gains are distributed across production operations. A company needs a reliable baseline before it can measure them.

How many hours does a line change currently take? Where do defects occur? Why does equipment stop? How much labor is spent on rework?

Without those numbers, a robot can demonstrate motion but cannot demonstrate economic value.

A reasonable starting point for manufacturers

A company does not need to begin with a humanoid robot.

It can select one repetitive process with measurable errors and relatively complete operating data. Material handling, standardized packing, visual inspection and equipment patrols are common examples.

Record normal work and exceptions. Calculate current labor, downtime and rework costs. Then require a supplier to run a limited test using the same measurements.

A four-to-eight-week trial can track intervention frequency, task success, downtime and maintenance effort.

If the system works only while the supplier’s engineers remain on site, it is not yet a scalable operating capability.

The boundary remains clear

Foxconn’s robot deployment is still early. Success in a few facilities would not mean that every manufacturing role is about to be automated.

Safety certification, equipment life, maintenance cost and adaptation to changing materials can slow deployment. AI-server demand is also cyclical. Changes in customer spending or chip architecture can quickly reach the manufacturing layer.

The precise conclusion is not that Foxconn has become a robotics company.

It is using its server business, factory network and supply chain to connect AI infrastructure with physical production. If that combination works, manufacturing capability itself may become part of the AI system.

This article is based on publicly available information as of 10:00 a.m. Beijing time on October 31, 2026. The new OpenAI–Foxconn collaboration has not been formally announced. Possible business directions discussed above are analysis, not confirmed contract terms. This article is not investment or procurement advice.