As September begins, many airports across the country are suddenly dominated by a striking billboard - "Push AI into the Physical World," with few words and a very direct meaning, credited to JD Cloud.
At the same time this billboard was launched, the theme of the JDD (JD Release Conference/Developer Conference) held on September 9 was clearly focused on "Physical AI."
"Physical AI" has undoubtedly been the most important hot topic in the Chinese context over the past six months, with the physical world considered the next destination for AI.
The industry has gone through two years of technological revelry, with a craze over reasoning speed, increasing context length, and even which chatbot resembles a real person more. But now a very realistic and even somewhat cruel question has been put on the table: How should these tokens, which have consumed astronomical computational power, and the massive traffic accumulated from consumers be transformed into real commercial value?
From the perspective of industry observers, this shift reflects two core propositions: Why must AI move out of the digital world? And why are companies like JD seizing this high ground?
When AI begins to leave the chat box
In recent years, AI has primarily been racking up points in the virtual digital world. However, the diminishing marginal returns of technological iteration and high operating costs have forced the entire industry to face reality - how much work can these capabilities ultimately complete, and how much actual value can they create?
Agents are the most obvious signal.
In the past, Copilots mainly participated in human work processes, helping programmers write code, sales people organize materials, and finance analysts analyze data; Agents are now beginning to take over entire tasks, reading corporate data, invoking various tools, and writing results back into business systems. A study released by OpenAI in June of this year summarized this change as the basic unit of knowledge work shifting from one-off interactions to "delegable long-term tasks."
Companies' ways of measuring AI have also changed. If Tokens measure how much intelligence the model produces, and Tasks measure how much work AI can undertake, then what companies ultimately need to see is still reduced costs, increased efficiency, and a closed business loop - in other words, Result.
As the general capabilities of foundational large models gradually level out, "whose parameters are larger" no longer holds an absolute moat. The battleground for AI competition will inevitably shift to "who can deeply connect with industrial processes and solve concrete problems."
This is also the backdrop for JD placing "Physical AI" in a core position at this year's JDD.
JD's judgment is that the AI competition is shifting from a "parameter competition" to a "productivity competition." At JDD, JD is putting cloud, data, models, terminals, scenarios, and supply chains into the same system, hoping to let AI further enter real scenarios like retail, logistics, industry, health, and home.
Compared to the industry’s generic use of "Physical AI" to broadly refer to robots, the logic that JD wants to explore is clearly broader.
Robots are a typical form in which AI directly impacts the real world, but more extensive changes have already quietly occurred across different industries. Retail AI ultimately needs to influence transactions and fulfillment, logistics AI is starting to engage in warehousing and delivery, industrial AI is entering production processes, while robots convert model judgments into physical actions. AI is beginning to move from producing information to further participating in the real-world production and service processes.
This judgment is also becoming the underlying consensus of the global AI industry. This year, Nvidia has continuously reinforced its layout in Physical AI, from Cosmos world models and Isaac simulation frameworks to GR00T robot models, teaming up with giants like ABB, FANUC, and Figure to advance technology into electronic assembly and industrial automation scenarios.
Physical AI is extending from a technical branch in the robotics field to the main channel for AI to penetrate real production systems and realize productivity value.
Why is it harder for AI to enter the physical world?
The rapid progress of large models over the past few years has largely been built on the vast amounts of digital content accumulated on the internet. Web pages, books, images, videos, and code can continuously be used for training; once the model is trained, it can be quickly deployed to various products and users through APIs.
However, when it comes to Physical AI, the Scaling Law begins to encounter the famous "Moravec's Paradox" - for AI, abstract logical reasoning has become simpler, but grasping an object and identifying a small obstacle in the complex real physical world is extremely difficult.
An AI entering a warehouse needs to know where the goods are, how inventory changes, and how different devices coordinate; entering a factory requires understanding materials, machines, and production processes; entering a household requires adapting to different spaces and hardware; entering robotics further involves actions, forces, environmental feedback, and differences between different entities.
This is also why the development of Physical AI is difficult to replicate the scaling path of large models in the past few years. Nvidia has summarized the core problem faced in robotics development as the "data gap": the internet has provided abundant pre-training data for large language models, but the amount of robotic data in the real world is limited and costly to obtain, and a lot of extreme scenarios are difficult to cover through genuine collection.
Specifically, in industrial applications, the problems that Physical AI needs to solve can be further summed up into three layers: Is there enough data to learn? Is there enough scene validation? Can it be scaled after successful validation?
This corresponds exactly to the logic JD proposed at JDD: Data allows AI to evolve, scenarios allow AI to validate, and the supply chain allows AI to scale.
First, there is data.
The data in the digital world mainly records knowledge already expressed by humans, while Physical AI needs to further understand how people and objects move, how actions change the environment, and how tasks are completed in real space. The production cost of this type of data is also higher. A piece of web text can be directly read by the model, while a robotic task may involve cameras, sensors, action trajectories, force feedback, and other modalities, and needs to go through collection, labeling, training, and real environment validation.
JD has chosen to extend its data infrastructure into the real world. Currently, its embodied data system covers the entire link of "collection, storage, labeling, training, evaluation, simulation, testing" and plans to accumulate 10 million hours of real scene video data within two years. The data sources have also expanded from robotic body data to tactile data, simulation data, and human-first perspective operation videos.
Beyond data, the second barrier is scenarios.
Physical AI ultimately faces continuously changing real environments, with a very low tolerance for error. A capability that works in benchmarking or laboratories does not mean it will still be effective after entering warehouses, factories, and homes.
This also leads to a fundamental difference between Physical AI and general digital models: in Digital AI, models typically complete training before being deployed to scenarios such as search and office; while in Physical AI, the relationship between the scene and the model is becoming deeper, with the scene itself being part of the continued iteration of AI's capabilities.
Real business generates data, data enters model training, and models re-enter to execute business tasks, with execution results continuing to form new feedback - thus forming a continuous circular feedback loop between data, models, and scenarios.
From this perspective, JD's accumulation of businesses in retail, logistics, industry, health, and home adds a unique value for the AI era. Taking industrial scenarios as an example, JD disclosed that its industrial large model JoyIndustrial has cumulatively been called over 1 billion times, with applications deepening from industrial knowledge understanding into Agent task planning, tool invocation, and execution of corporate business systems, and then feedbacking the execution results back into the system.
Scenarios are not only the places where AI capabilities are ultimately applied but also the physical closed-loop where AI gains real feedback and accomplishes self-evolution.
The third barrier is scaling.
Completing a task in the lab and ensuring that numerous devices operate stably for a long time in different environments presents a significant engineering and industrial chain gap. After AI enters the real world, model deployment is just one step; hardware production, components, channels, logistics, installation, maintenance, and after-sales will all impact whether a commercial closed loop can ultimately be formed.
Therefore, among the six words "cloud, data, model, terminal, scenario, chain," the last "chain" is particularly worth attention.
The model decides whether a capability can be realized, the scenario tests whether it can operate in a real environment, and the supply chain determines whether this capability can enter a sufficiently large real world.
This also forms a more complete set of industrial barriers for Physical AI: Data determines how AI understands reality, scenarios determine how AI adapts to reality, and supply chains determine how AI scales into reality.
As the resources needed for competition change, the original ranking of AI advantages among large companies may also change accordingly.
Physical AI may redefine the AI advantages of large companies
As Physical AI brings competitive variables into the real world, the physical infrastructure that JD has accumulated over the past twenty years gains a new interpretative space.
Compared to many major online internet platforms, JD's business has long been deeply integrated into the actual flow processes of goods and services. From procurement, inventory, warehousing to distribution and after-sales, as well as industrial procurement, health services, home appliances, and smart hardware, a large amount of business needs to connect online systems and offline fulfillment simultaneously.
These capabilities primarily served e-commerce and supply chain efficiency in the past. As we enter the Physical AI phase, they may also become the foundation for AI to obtain data, train and validate capabilities, and ultimately achieve large-scale deployment.
JD's proposal of a "Super AI Supply Chain" at JDD focuses on deeply connecting this physical network with AI. From the foundational cloud, data, and models, to the middle layer of Agents, development platforms, and simulation tools, and finally connecting terminals such as robots, home appliances, medical devices, and unmanned vehicles, ultimately entering real businesses like retail, logistics, industry, health, and home.
The robotics industry can more directly reflect the implementation of this "Super AI Supply Chain."
JD currently does not limit itself to developing a certain type of robot product. According to the plans disclosed at this JDD, JD hopes to simultaneously enter various links including embodied data collection, robotics manufacturing bases, core component supply, retail channels, and maintenance services.
In the next five years, JD plans to establish over 80 RoboBases; on the supply chain side, it hopes to reduce the BOM cost of robotic entities by over 50% through bulk purchasing, C2M customization, and supply chain collaboration; on the sales and service side, it aims to help 100 robot brands achieve sales exceeding 1 billion yuan in the next three years and build a robot after-sales service network covering over 100 countries and regions worldwide.
These plans correspond to questions that remain uncertain about the future of the Physical AI industry: If robots ultimately become a sufficiently large category of intelligent terminals, how will the industry value be distributed?
After the maturity of the PC industry, the value is dispersed across different segments such as chips, operating systems, complete machines, software, and channels; further, the smartphone industry has formed a complete division of labor between chips, operating systems, terminal brands, application ecosystems, and supply chains. If robots truly enter factories, shopping malls, and homes, they will similarly need models, data, core components, complete machines, channels, maintenance, and service systems.
Currently, this industry is still in its early stages, and the final industrial pattern has yet to be formed. But for JD, participating in Physical AI signifies more than just launching a certain robot or model.
It is trying to answer: When a large number of AI terminals truly enter the real world, what position can a platform with a supply chain and physical operation network occupy within it?
This also provides an entry point for understanding JD's proposal of "the world's largest physical world operation center."
JD has long operated the flow of goods in the real world. Once an order is generated, it connects procurement, inventory, warehousing, transportation, distribution, and after-sales; the digital system ultimately needs to fulfill in the real world.
Physical AI opens up further extension possibilities for this capability. In the future, what needs to be organized at scale may also include how AI obtains real data, enters different terminals, completes tasks in specific scenarios, and continuously operates through supply chains and service systems.
The competitive dimensions in the second half of AI are being redefined by the physical world, which happens to be JD's deepest and most adept battlefield.
Image Source｜AI Generated
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
