On 11 September 2026, China's Ministry of Industry and Information Technology (MIIT) used a press conference in Beijing to put numbers on something its industrial policy had until then only gestured at. The message was not that China wants to make software smarter; it was that China wants software that acts on its own. The vehicle is the "AI + Software" special action plan, presented by MIIT's Department of Information Technology Development, and for a Chinese policy document its targets are unusually concrete - dated, counted, and therefore auditable.
Five years, five numbers
By 2028, MIIT wants to raise the intelligence level of the software and IT-services sector significantly, spread adoption across 20,000 above-scale software enterprises, organise 100 enterprise-level intelligent technical-upgrade projects, build 100 agent-software benchmark applications in key industries, and incubate at least five key open-source projects. By 2030 it wants a further leap that moves China's software sector toward the high end of the global value chain. Read as a sequence, the list is a diffusion plan: a small number of showcase projects at the top, a benchmark set in the middle, and a mass of adopting firms at the base. It is a familiar Chinese shape - pick winners, publish reference cases, then let adoption do the rest - applied this time to agents rather than to solar panels or EVs.
From tools to intelligent partners
The framing matters as much as the figures. Officials described the plan as software moving from passive tools to intelligent partners that understand intent and complete tasks; development moving from human-written to human-machine collaboration; and delivery moving from selling products to providing capability. The last clause is the one to sit with. Selling capability rather than a product is a business-model statement, not a feature description. It implies metered, outcome-oriented software in which the agent - not the seat licence - is the unit of value, and it points to a services-heavy, subscription-and-throughput economy rather than the one-off licences that built the previous generation of Chinese software firms.
The baseline the plan is built on
Deputy Director Wang Weiwei supplied the numbers behind the ambition. China's software industry revenue reached 15.48 trillion yuan in 2025 - 6.2 times the 2012 level, a 15.1% compound annual growth rate. The multiple is impressive but backward-looking; the growth rate is the forward-looking part, because a sector already compounding at double digits does not need rescuing, it needs a new frontier to expand into. That frontier, in MIIT's telling, is intelligence. But the more revealing number is the second one.
Why tokens, not users, are the signal
By June 2026, China's daily token calls had passed 500 trillion. That figure measures machine consumption, not human use, and the distinction is the whole point. If the industry's centre of gravity is shifting from people writing lines of code to agents invoking models millions of times a day, then the metric worth steering is token throughput, and the sectors worth building are the ones that feed it. MIIT's plan is explicit about which those are: computing power, datasets and standards, listed among the plan's six areas alongside software production, product intelligence, agent software and services. The ordering is telling. A plan that leads with agents but funds the substrate is quietly admitting where the real constraints sit - not in cleverness, but in the physical and organisational capacity to serve machine traffic at scale.
What "providing capability" looks like
The shift from products to capability is easiest to see in the numbers MIIT chose to highlight. If firms buy outcomes rather than licences, then revenue decouples from the number of installed seats and tracks the volume of work an agent performs. That is why the plan's authors can speak of a software sector compounding at double digits and still want to re-found it: the old metric - seats sold - grows roughly with headcount, while the new one - tasks completed - can grow with machine hours. A sector reorganised around that second metric would look less like a packaged-software industry and more like a utility: metered by usage, priced by throughput, and judged by reliability.
The state's hand on demand
What is distinctive here is that the state is setting adoption targets, not only R&D goals. "20,000 above-scale enterprises" is a diffusion bet: it assumes agent software will spread through ordinary Chinese firms, not merely sit inside a handful of national champions. "100 agent-software benchmark applications in key industries" is an attempt to seed reference deployments that other firms can copy - a procurement-led strategy with a mixed record. When benchmarks are chosen for their usefulness, they pull an industry forward; when they are chosen for their countability, they produce demonstrations that satisfy a quota and then go quiet.
Two questions the plan leaves open
The first is durability. Measurable targets are a virtue - you can audit them - but measurability cuts both ways, because a number that can be hit can also be hit without producing anything durable. Whether "100 benchmarks" yields tools that survive contact with real workflows is the test, and the plan itself cannot guarantee the answer. The second is openness. A target of "at least five key open-source projects" is modest against the scale of everything around it, and it leaves unresolved whether China's agent stack will rest on open foundations or on a small number of proprietary platforms that the state can steer. Both questions are about the same thing: whether this plan builds capability or merely counts it.
Our take
Three things are worth carrying away. First, Beijing is no longer talking about software "with AI features"; it is talking about agents as the product and human-machine teams as the default way software gets made. Second, the demand signal it is optimising for is machine consumption, measured in tokens rather than human seats, which reframes the software market around throughput and the infrastructure that supports it. Third, the plan's real bet is infrastructural - computing power, datasets and standards first - which quietly concedes that the binding constraints on AI are physical and organisational rather than algorithmic. The targets are concrete enough to audit and broad enough to survive a change of minister. That combination is what makes them a signal rather than a wish list.
Sources
Ministry of Industry and Information Technology (MIIT), China - "AI + Software" special action plan press conference, Beijing, 11 September 2026: https://www.miit.gov.cn/xwfb/xwfbh/bxwfbh/art/2026/art_95c0b992336a4cb1b87ec3ccc576a669.html