← Back to all posts

Hangzhou AI City

Hangzhou treats AI as municipal infrastructure. DeepSeek, the Six Little Dragons, 70,000 petaflops of public compute, a City Brain running traffic and security, robots on the streets. But most of the world cannot replicate this. The dependency stack — compute, weights, silicon — means AI sovereignty has no third way. Hangzhou is the exception. The question is whether it is the only one.

aihangzhouchinasmart-citydeepseekalibabacity-brainautonomous-systemsai-sovereigntyexport-controlsmistraltsmc

Hangzhou was Alibaba's company town. Now it is the first city on earth to treat AI as municipal infrastructure — not a sector, not a policy document, not a pilot programme. A layer of the urban stack. The traffic lights, the hospital triage, the railway station, the robots — all running on shared public compute, all deployed now.

A city is a stack. Hangzhou is replacing every layer with AI. The transformation is not a strategy document. It is physical infrastructure.

The anchor is DeepSeek — the open-source model that triggered a global recalibration of AI economics in early 2025. It was built here. So was Unitree Robotics, whose quadrupeds captured 19% of the global market in seven years. So was Manycore Tech, whose spatial AI models reconstruct interior worlds. So was BrainCo, which relocated from Boston to Hangzhou's AI Town. Together with Game Science (Black Myth: Wukong) and Deep Robotics, they form the Six Little Dragons — the core of a trillion-yuan AI industry cluster growing at 26% annually.

Six companies. One city. A trillion-yuan target. Hangzhou bets on density, not scale. The towns are 3 km². The compute is public. The robots are deployed.

The density is engineered. Hangzhou is organized into technology towns — compact 3 km² zones where living, social, industrial, and entrepreneurial infrastructure are compressed into walkable districts. Yuhang's AI Town hosts BrainCo. Xihu's Yunqi Town birthed Alibaba Cloud. Binjiang's Embodied Intelligent Robot Town hosts Unitree. Each town is a micro-ecosystem. The governance philosophy is serve without interfering. Permits are fast. Costs are low. The talent pipeline runs through Zhejiang University — birthplace of DeepSeek, Deep Robotics, and Manycore's founders.

The compute is real. By the end of 2026, Hangzhou will have 70,000 petaflops of dispatchable intelligent computing — the largest municipal AI compute pool in China. The Zhejiang New Computing Center already houses a fully liquid-cooled 10,000-card cluster. The city runs an intelligent computing scheduling platform that allocates resources across the towns. Compute is a public utility.

Most cities talk about smart cities. Hangzhou built the compute first, then made it a utility, then let companies build on it. The order is the argument.

The applications are deployed, not demoed. The City Brain unifies public security, transportation, weather, and urban management into a single AI governance layer. Traffic lights adjust in real time. The West Railway Station identifies ride-hailing vehicles via AI cameras. Sir Run Run Shaw Hospital runs AI triage at 98.7% accuracy — mis-registration down 60%. Udeer.AI cleaning robots identify dirty areas rather than sweeping entire surfaces. Deep Robotics quadrupeds patrol power tunnels in Singapore. A TCM diagnosis AI analyzes tongue images and biological data, validated across 157,000 clinical studies. Agents for public security, sleep health, and AR tourism are live.

When the traffic lights, the hospital triage, the railway station, the robots, and the medical diagnosis all run on AI, you are not talking about adoption. You are living in an AI city.

Hangzhou treats AI as a layer of the urban stack, not a subsidized sector. Compute is a public utility. Data trades on an exchange. Models are open-source. Robots are deployed. The transformation is running now.

The dependency stack

Hangzhou built its own stack. Most of the world cannot. In June, the US ordered Anthropic to cut foreign access to its newest model, Fable. The company couldn't separate American users from the rest, so the world lost access overnight. OpenAI's GPT-5.6 Sol followed weeks later. Two American labs, both switched off abroad by decree.

The instinct is to go open. Moonshot AI released Kimi K3 — open-weight, beats American leaders at code. DeepSeek followed. Free of Washington. Except you cannot. Three walls. Compute: 2.8 trillion parameters. Almost no one can run it. Transparency: open-weight is not open source. No training data. No methodology. Reciprocity: Beijing is now weighing the same export limits.

Fully open alternatives — Switzerland's Apertus, Germany's Soofi — sit a step behind the frontier. Honest and sovereign is not the same as competitive. Mistral, Europe's contender, does everything right: independent ownership, its own data centers. Yet its chips are designed in America, built by TSMC in Taiwan — the one factory on earth that can make them. If the company that did everything right cannot reach the bottom of its stack, neither can you.

AI sovereignty is not about the model. It is about the stack beneath it. You are only as sovereign as your weakest layer — and for almost everyone, that layer is silicon.

Hangzhou is the exception that proves the rule. It built compute, talent, and institutions in a single dense corridor. Most cities and most countries cannot. If the AI we depend on can be switched off by a capital we do not vote in, and runs only on machines we cannot build, what is the plan for the next year, the next three, the next ten? Pretending we have already escaped is not one.

References: