While most Chinese tech giants raced to ship a consumer chatbot, JD.com (京东) did something less photogenic and possibly more durable: it grew an AI model out of its warehouse floors, delivery routes, and customer-service logs.
From Yanxi to JoyAI
JD unveiled its Yanxi (言犀) large model on July 13, 2023, at the JD Global Tech Explorer Conference — a hundred-billion-parameter "industrial" model trained on a mix of general data and JD's own digital-intelligence supply-chain data (roughly 70% general, 30% native). In July 2025, at the World Artificial Intelligence Conference (WAIC), JD rebranded the family as JoyAI, with a full-size matrix spanning 3B to 750B parameters across language, speech, image, video, and digital-human modalities. In February 2025, all Yanxi products were connected to DeepSeek; in April 2025, an AI outbound-call service went live.
The stated philosophy, from CEO Xu Ran (许冉) at the September 2025 JDDiscovery conference, was blunt: JD does not chase "performative AI" (运动式AI); it pursues AI that creates real industrial value. The implication is a deliberate contrast with peers who built a general model first and hunted for use cases second.
Where the AI actually runs
The hard proof is in logistics. JD Logistics runs a "wolf" (狼族) robot army: the "Smart Wolf" (智狼) goods-to-person system is deployed in more than 60 warehouses worldwide, including sites in the UK and Germany; the "Lone Wolf" (独狼) unmanned vehicles run routinely across more than 20 Chinese provinces; the "Flying Wolf" (飞狼) drones built China's largest village-delivery drone network in Zizhong, Sichuan. Orchestrating them is JD Logistics' SuperBrain (超脑) 2.0 model, which can complete routing calculation and simulation for a billion packages in about three minutes.
The measurable gains JD cites: inbound put-away efficiency up 6x, storage density 4x that of a traditional warehouse, outbound picking efficiency up 3x, picking accuracy 99.99%, and logistics automation rate 95%. By the end of 2025, AI agents were applied across more than 1,000 supply-chain scenarios, and the count of internal AI agents exceeded 50,000.
The consumer side
AI is not only behind the scenes. JD's AI shopping assistant crossed 150 million annual active users with over 20% penetration, driving several billion yuan (≈ US$ hundreds of millions) in GMV. Across the company, Yanxi/JoyAI supports more than 1,000 business scenarios, and token-call volume grew nearly 100x year on year in 2025. JD's compute comes from its Tianqin Alpha (天琴α) supercomputing center, and it uses a self-developed vector database, Vearch, for model memory and retrieval.
The loop Xu Ran describes is the interesting part: business generates data, data improves the model, the model improves operations, and bigger operations generate more data. Twenty years of self-run e-commerce and self-built logistics gave JD a data soil — real transactions, warehouse moves, deliveries — that a model lab simply does not have.
Why it matters
JD is the clearest example of "AI as operations, not AI as a product." Its model is not trying to beat GPT on a leaderboard; it is trying to shave a percentage point off fulfillment cost across a national logistics network. That is harder to demo at a launch event but harder for competitors to copy, because the moat is twenty years of supply-chain data, not a weights file.
The labor question is unavoidable. JD employs a large blue-collar workforce — unusual among internet companies — and founder Liu Qiangdong (刘强东) publicly said employees displaced by robots "will not be laid off." Whether that promise holds as automation rate hits 95% in some warehouses is the social test of this strategy, and one JD has not yet been forced to answer at scale.
What readers can do now
- If you run operations: copy the loop, not the model. JD's edge is proprietary supply-chain data; the lesson is to point AI at your own highest-volume, most data-rich process first.
- If you are a JD merchant: the Yanxi/JoyAI toolchain (customer service, digital humans, AIGC marketing) is already available — test it on product imagery and copy before buying external AI tools.
- If you invest: judge JD's AI by logistics cost per order and automation rate, not by model benchmarks. Those operational ratios are where the value shows up.
Honest limitations
Scenario counts (1,000+), internal-agent count (50,000+ by end-2025), efficiency multipliers (6x, 4x, 3x, 99.99% accuracy, 95% automation), and GMV figures are JD's own disclosures reported through Chinese industry media (NetEase News, Baidu Baike); they were not independently verified. The "nearly 100x" token-growth and 150M AI-assistant-user figures are company-reported. Liu Qiangdong's no-layoff statement is quoted from internal remarks reported in media and is a commitment, not a contracted guarantee. JoyAI's model-performance claims (3B–750B matrix) come from JD and Baike entries, not third-party benchmarks. Currency conversions use an approximate RMB/USD rate of 7.3.