China just coded its AI strategy into law. The rest of the world should be paying attention.
The 15th Five-Year Plan, approved by the National People's Congress in March, is not a policy suggestion. It is a procurement directive. Buried in its 141 pages is a sentence that should make every closed-model lab in the United States reconsider its assumptions: open-source AI is now a named flagship strategy, explicitly positioned as the structural differentiator from America's approach of keeping frontier weights in private hands.
The number that tells you how seriously to take this: the word "AI" appears 52 times in the document, according to the South China Morning Post. The previous five-year plan, released in 2021, mentioned it 11 times.
"Open source wasn't mentioned in previous reports, and this is also a key difference between the Chinese and American AI approaches," Tilly Zhang, a technology and industrial policy analyst at Gavekal Dragonomics, told Reuters. "I believe China has studied this very carefully and decided to make open-source AI a flagship strategy and a competitive advantage against the United States."
That is a policy statement, not a prediction. It describes a deliberate bifurcation now baked into law on both sides of the Pacific: China betting on open-source diffusion and hardware self-sufficiency; the United States betting that frontier capability stays concentrated in closed labs it can control.
The plan does not merely mention open-source AI. Chapter 5 designates robotics as one of eight strategic emerging industries, a step up from the 14th Five-Year Plan which addressed it through a subordinate MIIT sub-plan. Section 2 of the same chapter lists embodied intelligence alongside quantum technology, biomanufacturing, brain-computer interfaces, and 6G as six designated future industry tracks. Each receives mandatory coordination requirements across central ministries, provincial governments, and state financial institutions.
The money is real. Bloomberg reported that Beijing plans up to EUR 60 billion in additional semiconductor subsidies, targeting breakthroughs at 3-5nm and 7-10nm process nodes. The goal is explicit: shift data centers away from US chips toward domestic AI accelerators, primarily from Huawei and Cambricon.
The deployment targets are specific and staged. AI devices, agents, and applications should reach 70 percent penetration by 2027, 90 percent by 2030, and ubiquitous deployment by 2035, according to MERICS. The AI Plus initiative separately targets AI integration into 90 percent of China's economy by 2030.
The embodied intelligence designation is particularly revealing. It describes a training architecture that Western labs like Physical Intelligence, Figure, and Sanctuary AI are building ad hoc: coordinating physical training grounds, promoting virtual-real fusion collaborative training, developing integrated big-brain and small-brain models, and accelerating humanoid robot deployment. Beijing is turning that architecture into a named procurement category with state-directed capital behind it.
There is an existing proof of concept in heavy industry. CITIC Pacific Special Steel, a state-owned manufacturer, has developed more than 100 AI vertical models for its operations and converted one of its plants into what government-aligned media calls the first "lighthouse factory" in the global special-steel industry. The plan scales that model across every provincial manufacturing zone.
Whether the targets hold is another question. MERICS analyst Rebecca Arcesati notes the penetration figures function as signals to CCP cadres and provincial officials about where to direct investment, not as enforceable KPIs. Execution across 23 provinces and dozens of ministries consistently is harder than drafting a document. And China's semiconductor ambitions face real constraints: Huawei's Ascend chips remain behind the performance curve of Nvidia's latest offerings, and the extreme ultraviolet lithography equipment needed for leading-edge production is manufactured exclusively by ASML in the Netherlands, subject to export controls.
But the strategic direction is legible. Beijing appears to have read DeepSeek's R1 release — which demonstrated competitive reasoning performance at a fraction of the training cost of comparable Western models — as confirmation of a bet it was already making. Open-source models diffuse capability faster, train on more diverse data, and are harder to sanction. Closed models concentrate advantage in whoever holds the weights.
The question for builders and investors is not whether China's plan executes cleanly. It will not, not all of it. The question is what the open-source bet implies for the global distribution of AI capability over the next decade, and whether the US closed-model approach can match the diffusion speed that Beijing is now treating as a national objective.
† Add source attribution (e.g., 'according to MERICS' or other registered source) or footnote with 'Source-reported; not independently verified.'
† Add source attribution (e.g., 'according to MERICS' or other registered source) or footnote with 'Source-reported; not independently verified.'
