A founder in Zhongguancun logs on to a government portal and finds a new questionnaire about computing supply. The questions are not about grants. They are about whether her team can get stable access to domestic accelerators, and whether the models she builds can be tested inside the city's own hospitals and schools.
That small form is a useful window into a much larger document. At the start of 2026, Beijing Municipal Development and Reform Commission released the Beijing Artificial Intelligence Innovation Highland Construction Action Plan (北京人工智能创新高地建设行动计划) at the city's annual artificial-intelligence kickoff meeting. The plan is not a research wish-list. It is an operating plan with deadlines, named campaigns, and one headline target that the city now repeats at every artificial-intelligence event.
What the plan commits to
The plan's central promise is quantitative and time-bound. Within about two years, Beijing aims for its artificial-intelligence core industry to surpass 1 trillion yuan (about US$140.8 billion, HK$1.1 trillion). Alongside that, it targets more than 10 newly listed artificial-intelligence companies and more than 20 unicorns, and it commits to building a 100,000-card-class domestic intelligent-computing cluster (国产智算集群).
Those are not abstract ambitions. The 100,000-card cluster is a direct answer to the compute (算力) bottleneck that Chinese model developers cite repeatedly. The listing and unicorn targets signal that the city wants a financing ladder, not just a research base. The plan also names four artificial-intelligence innovation districts (创新街区) launched at the same meeting: Haidian Origin Community (海淀·原点社区), the Economic-Technological Development Area's Model-and-Data World (经开·模数世界), Chaoyang Light-Intelligence Space (朝阳·光智空间), and Shijingshan Culture-Intelligence Realm (石景山·文化智境).
Nine campaigns, not slogans
The document organizes its work into nine campaigns (九大行动). They are worth listing because they show where the city is choosing to spend political capital:
- Technology breakthrough — foundational theory, new model architectures, group-based, system-level攻关.
- Autonomous computing ecosystem — domestic chips, cluster key components, super-nodes, optical interconnect, and an open software stack.
- High-quality data — embodied-intelligence datasets, scientific-intelligence datasets, high-end data-labeling bases.
- All-domain application — "one committee, one sector; one district, one product" coordination across medicine, manufacturing and education.
- Scientific-intelligence reform — intelligent instruments, intelligent labs, and a research-paradigm shift.
- Embodied-intelligence chain — embodied models, motion control, and an autonomous operating system.
- Innovation ecosystem — talent, patient capital, and open-source organizations.
- Open-source cohesion — open-source technology systems and communities.
- Safety governance — deep security stress-testing services for models.
The list is notable for what it pairs: chips and datasets on the supply side, with medicine, manufacturing and education on the demand side. The plan is explicitly trying to close the loop between building capability and using it inside the city's own institutions.
Compute, data and talent
Three inputs dominate the plan's logic. The first is compute (算力). A 100,000-card domestic cluster only matters if researchers can actually reserve it, which is why the founder's questionnaire about accelerator access is the real implementation surface. The second is data. Beijing's emphasis on high-end labeling bases and sector datasets reflects a widely shared complaint that clean, usable training data is scarcer than raw model capacity. The third is talent and patient capital, captured in the unicorn and listing targets and in the explicit call to make better use of long-horizon investment.
The plan also leans on a mechanism Beijing has used before: open calls (揭榜挂帅) to select roughly ten key projects for directed攻关. That turns a broad vision into a short, fundable list — a pattern that is easier to track than a vague industrial strategy.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), the real test of an "innovation highland" is not how many models are released in a year but whether a self-reinforcing ecosystem of chips, models and applications actually forms and stays inside the jurisdiction.
Honest limitations
This is a municipal action plan with announced targets, not a legally binding statute, and the 1 trillion yuan figure is a directional goal rather than a guaranteed outcome. The plan's success depends on factors outside any single city's control, including domestic chip supply, national energy and data policy, and the willingness of hospitals and schools to open real scenarios. The nine campaigns are broad, and breadth can dilute accountability if progress is not reported against specific milestones. Some targets, such as the unicorn and listing counts, track venture outcomes that policy can only influence indirectly. Readers should treat the document as a signal of where Beijing is concentrating effort, not as a forecast of market size.
What readers can do now
- For operators building in Beijing: map your roadmap to the nine campaigns and the four innovation districts, and engage the open-call (揭榜挂帅) mechanism early rather than waiting for a generic grant cycle.
- For compute-dependent teams: treat the 100,000-card domestic cluster as a planning assumption and build reserve and scheduling logic now, because access policy — not raw capacity — will likely be the binding constraint.
- For investors and partners: track the plan's named milestones (cluster build-out, listing and unicorn counts, dataset bases) as the real indicators of execution, and discount headline ambition that is not tied to a reported metric.