8 月 13 日晚间,京东集团发布 2026 年第二季度及中期业绩,几个数字放在一起看,比任何单一指标都更能说明具身智能在中国第一次有了「打地基」的动作。
The Financial Backdrop
- Q2 revenue: RMB 346.4 billion, -2.9% YoY
- Q2 net profit attributable to shareholders: RMB 7.1 billion, +14.5% YoY
- Service revenue: RMB 79.3 billion, rising to 22.9% of total
- H1 system-wide R&D investment: +53.2% (accelerating for three consecutive quarters)
- JD Logistics revenue: +26.5%, JD Health: +17.2%, JD Industrials: +16.3%
- JoyAI-Echo: long-form audio/video generation framework
- JoyAI-VL-Interaction: real-time video vision-language interaction model
- JoyAI-Video-Edit: real-time streaming video editing model
- The ZhiLang goods-to-person solution is now in 60+ warehouses globally (including the UK and Germany)
- The YiLang robotic arm has been upgraded from single-arm to dual-arm, completing recognition, grasping, and stacking in 10 seconds
- Thousands of autonomous delivery vehicles operate routinely across 20+ provinces; Shenzhen opened the first night-time autonomous delivery routes
- A drone delivery network in Zizhong, Sichuan covers 78 administrative villages, crossing mountainous terrain in as fast as 7 minutes
- JD's digital human JoyAvatar has served over 80,000 merchants
- Accounts livestreaming with digital humans grew 3x YoY in Q2; daily active livestreaming accounts surged 6x
- During the 618 shopping festival, digital human livestream duration and transaction volume both grew more than 4x YoY
Product revenue is declining while R&D and service revenue are rising — a traditional e-commerce company trading profit for the next technology cycle.
Three Infrastructure-Level Commitments
1. RoboBase robotics bases — a 5-year plan for 80+ sites
The first RoboBase project broke ground in Guangzhou in Q2. Per the plan, these bases will provide robot data collection, testing and validation, scenario deployment, and maintenance services. JD intends to build 80+ RoboBase bases nationwide over five years, aiming to close the loop of "R&D — manufacturing — application — services" for the physical AI era.
2. EgoLive first-person dataset — 100+ universities applying
JD launched China's first embodied data collection community in Suqian, Jiangsu, and open-sourced EgoLive, the industry's largest human first-person-view dataset. The data comes from real business scenarios in retail, logistics, healthcare, and industrial operations — every frame corresponds to real physical actions and industrial needs. Over 100 universities and research institutes have applied for access.
3. Three open-source models — a physical-world model matrix
JD claims this physical-world model matrix has reached a globally leading level overall.
Three Sets of Deployed Numbers
Warehousing
Last-mile delivery
Merchant services
A Path Different From Other Players
Tesla Optimus takes a full-stack route of "in-house robot + in-house data + in-house scenarios"; Figure AI pursues single-point breakthroughs with a general-purpose humanoid deployed at a BMW factory; Unitree scales via cost-effective hardware across education and inspection use cases.
JD's path is distinct: it is not selling robots — it is building robots as infrastructure. The 80+ RoboBase bases are essentially a fusion of "robot 4S dealership + repair station + data center + training ground," with JD opening up all the real scenarios where robots can operate across logistics, retail, industry, and healthcare as training grounds. That differs from every American player.
The Real Question
JD management explicitly acknowledged on the earnings call that these metrics do not yet directly demonstrate how much new revenue, profit, or cost reduction the AI investment has produced. Three validation points for the next phase:
1. Whether ad conversion rates improve as digital human livestreaming scales 2. Whether YiLang dual-arm robots actually reduce per-unit warehousing fulfillment costs 3. Whether external technology service revenue reaches a quantifiable share
If any two of the three show positive signals within six months, JD's five-year, 80-RoboBase pledge stops being "laying a foundation" and becomes "paving a runway." The embodied AI numbers in the Aug 13 report are just the beginning — the Nov 13 report will confirm whether this line of business truly holds up.