Beijing Published the Blueprint for a National AI Agent Operating System. Nobody in the English-Language Press Covered It.

What a buried MIIT document reveals about China’s real AI infrastructure strategy, and why it matters more than any model benchmark.

Everyone tracking the US-China AI race is watching the wrong scoreboard. Model benchmarks. Parameter counts. Training compute. The United States leads on all of them. Earlier this month, Beijing published a document suggesting it knows, and doesn’t care.

China’s Ministry of Industry and Information Technology just laid out a three-year plan organized entirely around AI agents, not AI models. Not because China lost the model race, but because it left.

Document No. 121, signed June 3 and cascaded to county-level government bureaus within fifteen days, isn’t a telecom upgrade plan. It’s the architectural specification for a nationally controlled AI agent ecosystem: proprietary Chinese protocol stack, domestically manufactured edge hardware, state-supervised cloud infrastructure. I read all eight pages in the original Mandarin. Here’s what it says that nobody else is reporting. The full text has been publicly available on MIIT’s website since June 11, in Mandarin, with no access restrictions.

The document, 《“人工智能+信息通信”创新发展实施意见（2026–2028年）》 (“Implementation Opinion on Innovative Development of ‘AI + Information Communications,’ 2026–2028”), looks like boilerplate. Seventeen numbered tasks. Six sections. Standard ministerial formatting. But three things buried inside it are not standard at all.

First: China is building its own AI agent communication protocol. Section 2 mandates developing a 自主智能体通信协议, an autonomous agent communication protocol. In the West, this space is converging around two open standards under Linux Foundation governance: Anthropic’s Model Context Protocol and Google’s Agent-to-Agent protocol. MCP alone has crossed 97 million monthly SDK downloads. China’s response, embedded in state industrial policy, is to build a parallel stack from scratch. The document doesn’t reference MCP. It doesn’t reference A2A. It references nothing from the Western ecosystem. This is the Great Firewall logic applied to the agent layer. Not blocking content, but building a separate infrastructure so Chinese AI agents operate on Chinese protocols, inside a Chinese-controlled network topology.

Second: it introduces a concept that doesn’t exist in Western policy vocabulary. 智能体互联网: the “Internet of Agents.” A defined network architecture where AI agents are first-class endpoints, connecting through a Chinese-standard communication protocol, hosted on domestic cloud platforms, coordinated through 多智能体协同 (multi-agent collaboration). Cloud-based agent deployment. Model-as-a-service delivery. Everything an agent needs to operate, all of it domestic, all of it controllable at the border.

Third: foundation models barely appear. This is the anomaly that reframes everything else. Every major Chinese AI policy document from 2023 to 2025 was organized around 大模型, large models. Model scale. Model competition. Model benchmarks. Document 121 treats models as infrastructure to be hosted, not goals to be chased. From models to agents. That’s how completely the organizing concept has shifted, from the things that think to the systems that use thinking to act in the world. 智能体 (agent) appears over twenty-five times. 大模型 appears a handful of times, always as a supporting element.

When I first noticed the imbalance, I went back through every MIIT AI document I could find from the last three years. It’s real. Someone designed an architecture vocabulary (Internet of Agents, agent communication protocol, network agents, industrial agents, large-small model collaboration) and threaded it through a state-level policy instrument as a coherent system. This wasn’t assembled by committee. It was built by people who knew exactly what they were specifying.

The protocol fork matters most. Right now, the Western agent interoperability stack is consolidating under neutral governance. MCP and A2A are both Linux Foundation projects. OpenAI, Anthropic, Google, Microsoft, and AWS all participate. Silicon Valley’s implicit assumption is that AI agent protocols will be global and interoperable, the same way HTTP became the universal transport layer for the web.

China just told you it disagrees. Document 121 mandates an indigenous protocol developed under state direction. If this protocol matures and gets embedded in Chinese 5G-A and 6G networks, which is exactly what the document prescribes, then within the 2028–2030 target window AI agents operating inside China and AI agents operating outside China will speak different languages. They won’t interoperate. What happened with the internet is about to happen with the agent layer, except this time it’s being engineered from the start rather than bolted on after the fact.

Hardware follows protocol. Document 121 mandates edge compute infrastructure across 75% of Chinese metro areas at 1-millisecond latency by 2028. Those are performance specifications that match autonomous vehicle control and military command-and-control requirements. It calls for 400 and 800 gigabit-per-second backbone networks. These are not aspirational numbers dropped into a policy document for show. They are engineering targets with a two-year deadline.

Tying it together is what the document calls 大小模型协同, “large-small model collaboration.” Lightweight models handling routine inference at the network edge. Large models handling complex queries in the cloud. This is the infrastructure spec for distributed AI inference that reduces dependence on concentrated GPU clusters, the exact capability constrained by U.S. semiconductor export controls. Huawei is already building the hardware. Its Ascend 950PR inference chip is projected to generate $12 billion in revenue this year. DeepSeek is running inference on Huawei silicon. MIIT didn’t invent Huawei’s strategy. It built the state-mandated market that makes the strategy viable at national scale.

Read the document’s R&D priorities carefully. High-speed optoelectronic chips. Switching ASICs. Co-packaged optics. AI compute interconnects. That list maps precisely to the technologies restricted by BIS Entity List actions and the October 2022 and 2023 semiconductor rules. Nowhere does the document mention export controls. It doesn’t need to. Its R&D priorities are a publicly readable capability gap register. Every item listed is something China cannot yet produce domestically at the required performance tier. If you want to evaluate whether export controls are working, don’t wait for a BIS enforcement report. Read what MIIT is still investing in.

Who has the best frontier model? By that metric, the United States leads. But Document 121 suggests Beijing is playing a different game: not who builds the smartest model, but who controls the infrastructure that models run on, the protocols agents speak, and the edge hardware that puts inference within a millisecond of every user. If that’s the right scoreboard, the United States doesn’t have a score yet. It hasn’t started playing.

China’s three state-owned telecom carriers, China Mobile, China Telecom, and China Unicom, are expected to publish their own implementation plans within three to six months. When they do, the investment figures and procurement specifications will reveal whether Document 121 is aspirational policy or funded mandate. That’s the next document to read. I’ll be reading it.