Walk into almost any Chinese software firm this autumn and the busiest "engineer" is no longer a person. It is a copilot that writes, tests, and ships code. On 11 September 2026, the Ministry of Industry and Information Technology (MIIT) made that shift national policy.
The document is the "Artificial Intelligence + Software" Special Action Plan (《"人工智能+软件"专项行动实施方案》). It is the clearest signal yet that Beijing wants AI woven into the country's software industry from the compiler up—not bolted on as a chatbot afterthought.
What the plan actually promises
The plan runs on a two-stage clock. The first stage lands in 2028; the second in 2030.
- By 2028: intelligent coding tools and platforms reach 20,000 above-scale software enterprises.
- By 2028: 100 software-enterprise intelligent transformation projects rolled out.
- By 2028: 100 benchmark "agent software" (智能体软件) applications built in key industries.
- By 2028: 5+ high-quality open-source projects incubated.
- By 2030: critical software achieves a full intelligent upgrade, and AI + software becomes a new growth pole, with internationally influential open-source communities and clusters.
The framing is blunt. MIIT wants development to move from "human writing" to "human–machine collaboration," products from "passive response" to "active execution," and services from "product delivery" to "value delivery."
Why this is happening now
China's software sector is large but uneven. In 2025 the industry booked 15.48 trillion yuan in revenue—about US$2.2 trillion (HK$17 trillion)—which is 6.2 times its 2012 figure and a 15.1% compound annual growth rate. The base is real; the fear is that the base is also stagnant.
The more telling number is usage. As of June 2026, China's daily token (词元) call volume surpassed 500 trillion. That is not a curiosity—it is the meter reading that convinced policymakers AI load has left the pilot phase and entered scaled production. When model inference is measured in hundreds of trillions of tokens a day, the software that calls those models becomes infrastructure worth a national plan.
Agents, not just models
The plan leans hard on agent software (智能体软件)—systems that perceive, remember, decide, and act rather than merely answer. It calls for an agent-software app store and a skills (Skills) resource library, and it promotes "model-as-a-service" (模型即服务) and "agent-as-a-service" (智能体即服务) as new business shapes.
For industry, the plan pushes computer-aided design (CAD), engineering (CAE), and electronic-design-automation (EDA) tools to absorb AI, and tells industrial firms to package hard-won process knowledge into reusable model libraries. It also backs the domestic large model (大模型) open-source ecosystem, including the AtomGit community, which by August 2026 hosted over 80,000 models and 370,000 datasets with more than 12 million registered users.
The humanoid side note
The plan is not only about screens. It also tasks builders with operating a humanoid (人形机器人) open-source community—robot operating systems, middleware, and the "brain and body" models—so that the same software stack feeds physical machines. The line between "software" and "embodied AI" is deliberately thin.
Six work streams, one reset
The plan organizes itself into six work streams: reshape software production, upgrade software products, grow agent-software formats, expand intelligent services, build the base (compute, data, standards), and improve the environment (talent, employment, safety). The "employment-friendly" line is explicit—MIIT expects repetitive coding roles to shrink while demand rises for people who can express intent to machines and verify their output. New national job titles such as "generative-AI system application engineer" already exist.
On the ground, the plan promises practical levers: compute vouchers (算力券) to cut inference costs for smaller firms, high-quality software-engineering datasets to fight "data islands," and adaptation centers—MIIT mentions building 15+ adaptation centers linking basic software to domestic chips and servers. The message to foreign toolchains is implicit: local stacks get policy tailwinds.
Safety is written into the same page
MIIT does not treat safety as a footnote. The plan requires behavioral controllability and full traceability for agents, especially industrial ones that touch production lines. It flags new attack surfaces—prompt injection, poisoned training data, autonomous tool calls—and asks for identity tagging and trusted interconnection of agents. "Develop fast and stable" is the stated balance.
Honest limitations
This article is built on MIIT's 11 September 2026 press conference and the plan text carried by Xinhua and Guangming Daily. It reports targets and intentions, not outcomes—every figure above is a goal the government set, not a result audited today.
- The 2028 and 2030 numbers are policy commitments; none have been independently verified as achieved. Treat them as direction, not fact.
- The "US$2.2 trillion / HK$17 trillion" software-revenue conversion uses an approximate 7.1 RMB/USD and 0.91 RMB/HKD rate and is rounded; the underlying 15.48 trillion yuan figure is MIIT's own statistic.
- The 500-trillion-token daily volume is a MIIT-stated metric; we could not locate a public methodology document, so "token (词元)" counting across vendors is not independently audited here.
- We did not access the full plan PDF; quotes and targets come from the official press conference transcript and state-media summaries, which may omit implementation detail.
- Forecasts that China's AI-related sectors will top 10 trillion yuan by 2030 are Xinhua editorial projections, single-sourced, and excluded from the hardened facts above.
What readers can do now
- If you build software in or for China, map your 2026–2028 roadmap against the four 2028 benchmarks—above all the agent-software app store and skills library, which will define distribution.
- If you watch China tech policy, track the "model-as-a-service / agent-as-a-service" standards MIIT says it will write; those specifications, not the press release, will decide who can sell into the public sector.
- If you invest, separate plan targets from execution risk—re-read this piece in late 2028 against the actual count of benchmark agent apps shipped.