AI Reconstructs the Supply Chain, JD Aims at a Trillion-Dollar Artificial Intelligence Ecosystem

Recently, JD.com, which has been highly active, has "set its sights" on AI again.

At the JDD Conference held today, JD.com unveiled three AI products, three upgraded in - depth application platforms, and the application results of JD's self - developed AI in four major scenarios. Meanwhile, Xu Ran, the Vice Chairperson of the SEC of JD Group and the CEO of JD Group, announced that JD will continue to invest in the development of artificial intelligence in the next three years, drive an AI ecosystem worth trillions of yuan, and open up scenarios and training data to AI enterprises such as those focusing on embodied intelligence.

JD.com's concentrated display of its self - developed AI capabilities actually provides a differentiated sample for observing the implementation of AI applications in the current Internet industry. JD.com's "scenario - driven" strategy has achieved results: it first verifies the practical value of AI applications in its own supply chain system, and then systematically outputs the capabilities tested in actual combat.

Behind this path choice, it not only represents JD.com's essential understanding of the strategic significance of AI but also echoes the current industry trend of AI shifting from emphasizing technology R & D to focusing on practical application value. JD.com's practice may provide key elements and possible paths for the in - depth application of AI.

Self - use and Polishing: AI Transforms the Supply Chain

Judging from the announcements at the JDD Conference, during the period when JD.com was "working in isolation" on self - developed and self - used AI, it has achieved the result of running its self - developed AI capabilities through core business scenarios.

Taking the logistics scenario, a key link in the supply chain, as an example, JD.com has deeply embedded large models in its self - built logistics system. After the upgrade, the Logistics Superbrain 2.0 is no longer just a single - point automation tool. Through the triple mechanisms of full - domain perception, model evolution, and human - machine collaboration, it has transformed the operation processes that originally relied on manual experience into data - driven dynamic decision - making.

Data released by JD.com shows that the Logistics Superbrain 2.0 has made the full - link operations "perceivable, predictable, and collaborative": in the monitoring link, the standardization level of operations has increased by 15%; in the scheduling of people, vehicles, goods, and sites, the front - line efficiency has nearly increased by 20%; in human - machine collaboration, the efficiency has increased by more than 20%.

This achievement has further achieved full - scenario coverage of the logistics from warehouse storage to retail terminals through self - developed products such as Zhilang, Tianlang, Dilang, Feilang, Dulang, Yilang, and the automatic sorting wall. Currently, the "Wolf Clan" products have covered the world, and the number of warehouses with large - scale deployment and application has exceeded 500.

In the industrial scenario, JD.com's JoyIndustrial has been deeply embedded in the entire supply chain process. This is an industrial supply chain large model. Relying on 57.1 million industrial product SKUs and data from more than 40 sub - industries, it not only covers links such as procurement and fulfillment but also gives rise to more than 40 types of intelligent agent applications such as product governance, supply - demand matching, and cross - border compliance.

According to JD.com, in product governance, the AI virtual governance team has compressed the time for hundreds of thousands of governance tasks from "months" to "hours"; in supply - demand matching, clarifying thousands of business opportunities no longer takes 5 hours but is shortened to 15 minutes; in cross - border business, the customs clearance agent intelligent agent has helped enterprises reduce labor costs by more than half, and customs clearance evaluation has been improved from T + 3 to near - real - time, achieving a leap in supply chain efficiency.

A key point here is that in addition to improving efficiency, the JoyIndustrial industrial large model also lies in enhancing governance standardization and compliance credibility. This also echoes another unique narrative of JD.com in AI application - trust first.

Especially in the health scenario, the Jingyi Qianxun 2.0 medical large model launched by JD.com has made "credibility" its core priority. It can not only simulate doctors' clinical diagnosis and treatment thinking but also actively introduce evidence - based medical evidence in the reasoning process and align with medical consensus to ensure the clinical credibility of diagnosis and treatment results. At the same time, Jingyi Qianxun 2.0 can fuse and analyze multi - modal medical data such as text, images, and test reports, establish a cross - modal "thinking chain", and improve the integrity and accuracy of diagnosis.

The same trust logic is also clearly reflected in the retail scenario, which is the end of the supply chain and also one of JD.com's core businesses. If in the medical scenario, "credibility" means the guarantee of evidence - based medicine and clinical logic, then in the retail scenario, "credibility" is reflected in the reliability of consumer decision - making and fulfillment: through applications such as AI shopping and AI try - on, JD.com enables users to make quick decisions based on personalized recommendations and real - visible try - on effects instead of repeatedly comparing prices and searching in a vast amount of information; for merchants, tools such as the Jingmai AI Assistant and the Jingdiandian Design Intelligent Agent make marketing and fulfillment more compliant and efficient, reduce return rates and losses, and enhance the trust in transactions.

It can be seen that from logistics, industry to health and retail, JD.com has achieved a paradigm reconstruction and efficiency re - creation of the entire supply chain with AI. In this process, the supply chain also serves as the best reinforcement learning scenario for JD.com's AI. Due to the large number of links, large amount of data, and fast feedback loop, JD.com can polish more robust AI capabilities in its own scenarios. This is why when the industry is generally still exploring the implementation path, JD.com has taken the lead in presenting a verifiable application report card.

In - depth Application after In - depth Trials

In addition to the display of application results in core business scenarios, another highlight of the JDD Conference is the three AI products and three in - depth application platforms. Their launch marks that JD.com is transforming the results of self - use and polishing into capabilities that can be called by the industry. From the focus of empowerment of these products and platforms, we can also see JD.com's strategic stratification in promoting the in - depth application of AI.

The three products launched based on the JoyAI large model are targeted at different business directions.

For consumers, JD.com is expanding AI from "shopping recommendations" to full - scenario life services. Among them, Jingxi is an AI - native application. It can not only complete actions such as shopping, ordering food, and booking hotels with a single command but also has the biggest difference in the online - offline linkage: it is directly connected to services such as JD.com's retail, food delivery, and hotel and travel services, truly embedding AI in real life.

"HeSheIt", positioned as a digital human assistant, integrates voice, expressions, and actions. It can not only meet the needs of knowledge Q&A, medical consultations, financial services, etc. in one - stop but also, through the "same soul" connection, implant intelligent agent capabilities into hardware terminals, extending the companionship and interaction of digital humans to a tangible form in reality, and attempting to create a symbiotic AI experience.

It can be seen that JD.com is not only optimizing its own retail experience through AI but also transforming its consumer - oriented interaction and service capabilities into a replicable AI application model to empower user touchpoints in more scenarios.

For the hardware ecosystem, JD.com officially launched the Embodied Intelligence Platform JoyInside 2.0. It is equivalent to a "brain" for robot and toy manufacturers. By opening up this platform, JD.com enables partners to quickly access mature intelligent agent capabilities and embed intelligence in hardware terminals. Currently, JoyInside 2.0 has been connected to robot and AI toy brands such as Unitree and Fuzozo, and relevant products have been launched on JD.com for sale. According to statistics, the average number of dialogue rounds of smart hardware connected to JoyInside has increased by more than 120%. From the rapid access of partners to the gradual formation of the user perception of "buying robots on JD.com", JoyInside 2.0 is no longer just a platform but an accelerator for the intelligentization of the hardware ecosystem.

The three in - depth application platforms generally aim to lower the threshold for in - depth AI application in various industries.

Among them, as the industry's first 100% open - source enterprise - level intelligent agent, JoyAgent 3.0 not only integrates multi - modal RAG, data governance and other capabilities accumulated within JD.com but also can seamlessly connect with the enterprise's own knowledge base and database, becoming the "intelligent decision - making center". Since its open - source in July 2025, JoyAgent has received more than 10K Stars on GitHub. At the JDD Conference, JD.com said that JoyAgent has further open - sourced the DataAgent and DCP data governance modules, fully supporting multi - modal data processing, and enabling intelligent agents to be truly used in enterprise production scenarios.

JoyCode 2.0 combines the "intelligent agent + code platform" and is positioned as an enterprise - level intelligent development platform. It supports fully automated programming with "zero hand - written code" and has served tens of thousands of R & D personnel at JD.com, with an adoption rate of generated code exceeding 40%. As an open - source tool, JoyCode 2.0 productizes JD.com's internal R & D efficiency - improvement experience to help external developers and enterprises build complex applications with low thresholds.

The Digital Human 4.0 has moved from "substitute execution" to "personalized creation" and demonstrated new value in scenarios such as marketing, cultural tourism, and endorsement. It can not only outperform human anchors in live - streaming e - commerce at 1/10 of the cost but also become a long - term digital asset for brands. Relying on the JoyAI LiveHuman model, through the open interface and toolchain, JD.com's digital human practice experience can be expanded on a large scale to more industries.

Open - source is not only a common feature of the three platforms but also a key path for JD.com to release AI capabilities to all industries.

In addition to the above three in - depth application platforms, JD.com has also open - sourced a number of supporting capabilities, including: the OxyGent multi - intelligent agent collaboration framework, the xLLM inference framework optimized for domestic chips, the JoySafety large - model security, and the medical large model Jingyi Qianxun 2.0 that breaks through trusted reasoning and full - modal capabilities. These are all specific efforts of JD.com to lower the threshold of AI application across the industry, transforming its polished experience and methodology into industry public resources and accelerating the popularization of in - depth application of large models.

The release of this entire set of AI products + platforms is essentially that JD.com is not simply packaging AI capabilities into new products and launching them into the market. Instead, along the strategic paths of consumer experience, hardware ecosystem, and industrial application, it is outputting the capabilities systematically polished in long - term scenarios such as retail, logistics, health, and industry to external enterprises and users. This is a manifestation of the long - accumulated and suddenly - released AI capabilities of JD.com: they come from actual - combat precipitation rather than the empty rotation of computing power, parameters, and concepts; then, the deeply - refined capabilities are packaged into complete solutions that can be directly called by enterprises and even C - end users.

JD.com Provides Answers in the ROI Era

In fact, not only as comprehensively shown at this JDD Conference, JD.com's AI strategy has deviated from the level of flaunting parameters and skills from the very beginning.

Since around 2020 when JD.com started to layout the Yanxi large model, it has insisted on rooting AI in the supply chain and physical business scenarios it is best at, emphasizing that large models themselves cannot directly create value and need to be placed in scenarios to truly play a role.

In 2021, JD.com launched the K - PLUG model with a scale of 1 billion parameters, which generated a total of 3 billion characters of product copywriting for the retail business; in 2022, it released the Vega model with a scale of tens of billions of parameters, which served the one - stop digital and intelligent supply chain data platform of JD.com's logistics and helped enterprises achieve cost reduction and efficiency improvement. It can be seen that these large models were deeply embedded in JD.com's core scenarios such as retail and logistics at the beginning of R & D and polished their capabilities through large - scale trials in high - complexity scenarios.

Then, in 2023, JD.com officially released a large model with hundreds of billions of parameters. At that time, the press conference did not overly focus on parameters but emphasized that the model integrated "70% general data + 30% JD.com's digital and intelligent supply chain native data", had higher industrial attributes and generalization ability, and would deeply cultivate knowledge - intensive and task - oriented scenarios such as retail, finance, logistics, and health to solve practical problems.

By this year, before the JDD Conference, He Xiaodong, the senior vice - president of JD Group and the deputy dean of the JD Exploration Research Institute, also mentioned in an interview with the media that "JD.com's large model has emphasized industrial attributes from the beginning. We not only use general Internet data but also construct data for specific industries and tasks to make the model more deeply used in vertical fields." This again highlights JD.com's strategic orientation of using industrial depth to polish AI depth.

It can be seen that the emphasis on in - depth application at the JDD Conference is actually a re - affirmation of JD.com's consistent AI strategy. The only difference is that after self - use and polishing, JD.com has begun to release its AI capabilities externally to promote the actual application of AI in all industries. However, whether it was only for self - use at the beginning or open - source now, JD.com's AI ambition has always been rooted in scenarios.

Since this year, both the industrial circle and the capital market have been emphasizing the ROI index of AI applications, which actually also emphasizes that AI should move from the stacking of computing power and parameter indicators to the value improvement of real application scenarios.

Taking Microsoft, which is deeply tied to OpenAI, as an example, in its latest financial report and earnings conference call, it hardly mentions the parameters of the GPT series models as it used to, but instead emphasizes more on the performance optimization, actual application effects, and relevant product capabilities of the models, especially the usage of Copilot. It is reported that the monthly active users of the Copilot series have reached 100 million, and the penetration rate of paying users has reached 35%. That is to say, Microsoft's large - scale investment in AI has not only brought actual efficiency improvement to users but also begun to bring business and investment returns to itself.

Domestically, the most intuitive change is that at this year's WAIC, a major stage for the concentrated display of artificial intelligence, exhibitor enterprises have also begun to show more of their actual application capabilities to better answer the questions of investors visiting the exhibition about the ROI index.

JD.com's concentrated output of AI capabilities at this time echoes the industry trend and can even be said to be a leading example of the industry leader in the in - depth application of AI. Its underlying methodology is: first, root in scenarios and polish self - use in complex supply chain links such as retail, logistics, industry, and health; then, rely on the feedback loop to iteratively optimize through high - frequency data such as transactions, fulfillment, and diagnosis and treatment to make the model more and more robust; finally, move towards openness, productize and platformize the precipitated capabilities, and more importantly, open - source them, contributing the technical capabilities of intelligent agents, inference engines, and security gateways, as well as the business capabilities in vertical fields such as medical diagnosis and treatment to the industry, enabling enterprises and users to directly call and deploy JD.com's experience.

JD.com is building an in - depth AI application link in the way of "self - use precipitation to external empowerment", and judging from the existing results, this may be a highly feasible and replicable path for industrial AI.