At the Dinghao Building in Beijing's Haidian District, an embodied intelligent robot named Xiaogai is now working at a convenience store, handling tasks like grabbing sausages and retrieving items for customers. Twenty kilometers away in Guomao, an office worker tells the "Xiaotuan" app on her phone five minutes before leaving work: "Find us a Sichuan restaurant near the office for six people at 7 PM." By the time she exits the elevator, the reservation is confirmed and a taxi voucher has arrived in her account. The foundational strength of China's AI capital is being demonstrated across countless everyday consumer scenarios in Beijing.
In June 2026, eight government departments including the Ministry of Commerce issued implementation guidelines on accelerating the development of "AI + Consumer," and in July, the State Council's approval of the "15th Five-Year Plan for Expanding Consumption" explicitly called for deepening this initiative. With such policy tailwinds, who is best positioned to capitalize? The answer lies in Beijing's AI infrastructure. The city didn't start building its AI capabilities only after policies were announced. As of July 2026, Beijing had filed 259 large models for record, with daily calls exceeding 1.4 billion times.
Here, models like GLM-5.2 from Zhipu and Kimi K3 from Moonshot AI have topped global open-source rankings, while the latest FlagOS platform developed by the Beijing Academy of Artificial Intelligence (BAAI) supports multi-chip deployment for over 90% of mainstream open-source models. As early as 2024, Beijing was the first city in China to release an "AI+" action plan, establishing itself as an AI hub with complete computing power, algorithms, and application scenarios. By 2025, Beijing's core AI industry is estimated to reach 450 billion yuan in scale.
In the first half of 2026, retail sales of wearable smart devices, high-efficiency home appliances, and audio-video equipment in Beijing all grew by more than 60%, while production of service robots and industrial robots surged by 230% and 75.5% respectively. Beijing is proving that the true value of an AI capital is ultimately realized across industries in both the digital and physical worlds.
The Foundation of an AI Capital
At the BAAI conference held in June 2026 at the Zhongguancun International Innovation Center, visitors stopped at the Tsingmicro booth to examine its AI chips, which use 3.5D heterogeneous stacking and Chiplet architecture to vertically stack reconfigurable computing chiplets with DRAM storage chiplets—transforming a traditional 2D "single-lane" into a three-dimensional "four-lane highway." That same month, Tsingmicro completed Day0 adaptation for Zhipu's GLM-5.2 model, and by the following month, it adapted Tencent's Hunyuan Hy3 through FlagOS. Moore Threads has also moved quickly to collaborate with large models: between July 21 and August 3, its full-function GPU completed training and inference validation with Extreme Vision's embodied world model, then announced Day0 adaptation for Kimi K3 and MiniMax H3.
As chips form the foundation and computing power serves as the energy source, Beijing added 22,000 P of intelligent computing power in the first half of 2026, bringing its total supply to 82,000 P. At current levels, 30,000 P of computing power can support parallel training for over a thousand large models. The impact of infrastructure on model performance is already tangible. In the early hours of July 17, Moonshot AI released Kimi K3, featuring 2.8 trillion parameters and a 1-million-token context window—enough to fully load the entire epic novel "Fengshen Yanyi" in one go, with room to spare for additional materials and conversation history. The model topped the Arena leaderboard, voted on by global users, on its release day, drawing a comment from Tesla CEO Elon Musk: "Impressive."
Musk had previously praised Moonshot's attention residual technology report released in March, which was also highlighted by Moonshot founder and CEO Yang Zhilin during his speech at the Zhongguancun Forum annual conference in Haidian District that month. In late July, at a press conference for Beijing's measures to support agent-led development, Lin Jianhua, deputy director of the Beijing Development and Reform Commission, also referenced attention residuals. That same press conference mentioned foundational model Zhipu GLM-5.2, general agent development platforms like ByteDance's Coze and Baidu's Wenxin, vertical agent platforms and frameworks from FanShi Intelligence and ModelBest, as well as inference engines from Tile RT, Zhongke Jiahe's SigInfer, and Qingcheng Jizhi's Chitu. This list reveals how Beijing's AI infrastructure and industry chain are interconnected.
Among the 2026 Forbes China Top 50 AI Technology Companies, 21 are based in Beijing—the highest of any city—spanning large models, AI chips, and enterprise AI. Ten of these companies are located in Haidian, home to the "Origin Community," a roughly 3-square-kilometer area that is one of Beijing's first four AI innovation districts, bringing together over 1,000 AI scientists, more than 13,000 AI developers, and over 100,000 AI students. As a national hub for AI industry innovation, Beijing's 259 registered large models cover general-purpose, industry-specific vertical, and mobile edge models, with leading models boasting over 2 billion registered users and 1.4 billion daily calls.
Making Technology Work Through Policy
With infrastructure—chips, models, and frameworks—plus computing power in place, AI applications in consumer and service scenarios can truly take off. How does Beijing convert technical capabilities into practical use? The latest answer is through agents, which transform AI capabilities into digital labor that can work, transact, and save time, embedding directly into consumer scenarios. Lu Hang, a post-90s fashion graduate, runs a sewing workshop offering "zero-basics, 3-day sewing crash courses" with students from across the country. For him, teaching and pattern-making aren't the hard part—the most time-consuming task is turning students' vague ideas into executable clothing plans. Previously, Lu relied on repeated communication, sketching, and revisions, often spending a full day on a single design mockup.
Now, using Baidu's Dazi general agent, he has created an "AI Sewing Study Room." Through daily conversations, Lu describes students' outfit inspirations, and the agent generates visual garment mockups. After adding fabric, craftsmanship, pattern suggestions, and production cost details, it organizes everything into course steps, lecture outlines, and short video scripts. This transformation in the garment industry is now being replicated across other manufacturing sectors. With agents, merchants can turn customer ideas into visual designs on the spot, provide production recommendations, and connect with source factories for manufacturing. Yuan Foyu, vice president of Baidu, articulates the convergence of AI and consumption clearly: AI development is an "infinite game," and so is consumption. "AI's value is shifting from efficiency gains to growth drivers. Consumer competition has moved from 'meeting needs' to 'discovering and generating needs.' AI can understand deeper intentions, continuously stimulate demand, and serve the needs that consumers themselves may not have clearly articulated. A deeper change is the transfer of consumption decision-making power: algorithms used to stand on the platform's side, deciding what users see; in the future, consumer agents will stand on the user's side, helping consumers filter, compare, and even execute transactions."
This agent-based transformation of AI technology is already showing up in data. In June, Volcano Engine reported that Doubao's large model daily token calls surpassed 180 trillion, a more than tenfold increase year-over-year, with business focus shifting toward coding, agents, and production scenarios. In specific applications, Beijing's medical AI pilot base formed eight expert agents in the first half of 2026, including systemic chronic disease retinal early screening, while the manufacturing AI pilot base accelerated full-lifecycle intelligent automotive R&D. The digital economy added value grew 7.8% in the first half of the year. This trend is also being systematically advanced at the policy level. In late July, the Beijing Development and Reform Commission and relevant municipal departments jointly released measures to accelerate agent-led development, incorporating concepts from the tech community such as Agentic AI, Harness Engineering, AIP (Agent Interconnection Protocol), and AI OS. The policy promotes integrating agent capabilities into terminal hardware, developing next-generation smartphones, and supporting innovative entrepreneurship models like OPC (One Person Company), with plans to create open, shared OPC communities in eligible districts.
The Supply Side of AI + Consumer
Agents have restructured the supply side of AI + Consumer, but the real goal is integrating AI into everyday scenarios. "At noon I watch the food preparation, in the afternoon I check inventory and promotions, and at night I review data and schedule the next day's shifts. Nothing is particularly difficult—it's just time-consuming," Lao Zhou, who runs a restaurant in Beijing, told a reporter. Now, using an AI assistant built with Meituan's CatPaw, he can automatically compile sales, review, and marketing data into daily reports, with anomalies and priority issues flagged upfront. Lao Zhou only needs to adjust prices and make judgment calls.
On the other side of the dining experience, "Five people, near Qianmen, 8 PM, Sichuan food, no waiting" was how Ye Qing, a tourist visiting Beijing during the May Day holiday, used Meituan's "Xiaotuan" to solve dinner plans. During the holiday period, Xiaotuan-related services received over 3.9 billion impressions. The recently upgraded version 2.0 of Xiaotuan can now not only search and compare options but also assist with placing orders, hailing taxis, and making reservations in real time. The 700 million merchant verifications and 1.3 billion user reviews stored in the background have moved beyond spreadsheets—they're now simply a matter of "help me arrange it" from users.
Government policy is also pointing toward deeper applications. The implementation guidelines issued in June by eight departments, including the Ministry of Commerce, outline 17 specific directions, from increasing new AI product supply to promoting AI robot consumption. It explicitly calls for accelerating demonstration applications of elderly-care robots, companion robots, and AI life assistants for key populations like the elderly and children, enhancing capabilities in emotional companionship, health monitoring, mobility assistance, and smart housekeeping. In June, Songyan Dynamics and Kidswant agreed to jointly promote the large-scale deployment of embodied intelligence in parent-child family scenarios. Beijing-based XinYan Robotics is also focused on people: at the recent WAIC (World Artificial Intelligence Conference), its proactive intelligent robot "Babu" demonstrated conversation, photography, and head-turning/eye-blinking functions, positioned as a family companion. Unlike industrial robots that precisely grip cups, companion robots serve a different purpose—yet both reflect the value of robotics applied across various scenarios.
In the first half of 2026, Beijing's industrial robot production grew 75.5%, while service robot output surged 230%. Over the same period, AI-related financing in the city exceeded 95 billion yuan, accounting for over 30% of the national total, with the number of new listings already matching the full-year figure for 2025. "We expect to open user experience registration at our Beijing headquarters in September," Zhang Xinhua, co-founder and co-president of XinYan Group, told a reporter. Against the backdrop of policy supporting "first launches and first exhibitions of AI products" and "AI companies opening branded stores and experience centers," such initiatives are already underway. Zeng Hui, senior vice president of Galaxy General Robotics and president of its cultural tourism division, told a reporter, "Galaxy General has already replicated the Beijing-launched space capsule retail store concept to nearly 170 service locations across more than 40 cities nationwide."
Looking at Beijing's ever-improving AI track record, a self-reliant and controllable AI technology foundation has formed in the northwest. Unique exchange platforms like Haidian Night School, daytime coffee/nighttime drinking spots, hotpot gatherings, and Party Nights foster vibrant intellectual collisions, with nearly 180 OPC companies concentrated mainly among 25-to-30-year-olds. In the southeast, the city's first token factory has a daily production capacity of 1.4 trillion, fully supporting model training and applications, while the nation's first AI data training base has created industry-specific data zones for aerospace, autonomous driving, and more, strengthening the full-chain AI supply. Along the central axis, whether traveling clockwise or counterclockwise, the "AI + Consumer" vision outlined in the State Council's July approval of the "15th Five-Year Plan for Expanding Consumption" is unfolding before your eyes.
Comments