Zhejiang Builds AI Vertical: Why China Needs a Province-Supercomputer

While we're tracking Sam Altman's latest tweet or debating whether the generative AI bubble will burst, China is quietly laying plans that will define the industry for a decade to come. Zhejiang Province, which happens to be home to Alibaba and a whole constellation of tech giants, has released a draft of its strategy for the "15th Five-Year Plan." And this is no mere collection of boring bureaucratic slogans about progress.

Before us lies a detailed engineering blueprint for how to turn an entire region into a giant, smoothly operating computer. If you thought cloud computing was simply renting servers, Zhejiang is about to prove it's a matter of national security and deep industrial integration. The central idea of the document is total coordination of computational resources.

The provincial authorities are clearly not pleased that every startup or factory tries to build its own small and often inefficient data center. Instead, they propose creating a multi-layered system of services where capacity is distributed as logically as electricity in a city grid. This is a direct continuation of China's ambitious "computing from east to west" policy, but now applied at the micro level of a single, albeit very wealthy, province.

Zhejiang has always been a testing ground for economic experiments, and now it's becoming a testing ground for AI centralization. The most interesting and perhaps most crucial for understanding the situation is hidden in the term "chip-model binding." Under conditions where access to top Western accelerators like H100 is restricted by sanctions, the Chinese are betting on deep synergy.

This means that the development of new chip architectures and the training of neural networks will go hand in hand. If you don't have the world's fastest processor, you must write software so it squeezes the maximum out of the hardware you have on hand. This is a forced, but extremely effective transition from extensive growth to meticulous optimization, which in the long term could give China an advantage in computing costs.

Another important point in the plan is the creation of a system of "base and vertical" models. While Silicon Valley chases universal superintelligence capable of writing poetry and coding simultaneously, Zhejiang pragmatically looks toward industrial application. Base models (LLM) will serve merely as a foundation upon which hundreds of narrowly specialized solutions for the real sector will grow: from optimizing factory machinery to logistics chains in the ports of Ningbo.

This is a living embodiment of the "AI+" strategy, aimed at embedding algorithms into every element of the production chain, rather than simply leaving them to entertain users in chatbots. But technologies don't exist in a vacuum, and China's leadership understands this perfectly. The plan emphasizes the importance of an "open ecosystem" and the creation of national experimental zones for data work.

In today's world, data is oil, and high-quality labeled data for training is high-octane fuel. Zhejiang intends to build the world's best "refueling station," where legal and technical barriers to information exchange between state and private business are minimized. It's an ambitious challenge, given the traditional opacity of corporate data, but within the framework of a "five-year plan," such matters are usually resolved directively.

Why is all this needed in a global sense? Officially—to achieve "common prosperity," the mantra of China's current leadership. In reality, it's a large-scale attempt to create complete technological sovereignty and prove that a planned economy can effectively manage such a chaotic and rapidly growing sphere as artificial intelligence.

While Western companies compete with each other under free market conditions, Zhejiang is trying to make all system elements—from silicon to cloud services and end applications—work as a single, perfectly tuned mechanism. The bottom line: China is definitively transitioning from purchasing computing power to its total state management and hardware optimization. If the "chip-model binding" strategy works under conditions of hardware scarcity, we'll see a new type of AI economy, where software efficiency and integration density become more important than the sheer number of teraflops.

Whether bureaucracy will manage not to stifle innovation in the process of this rigid "coordination" remains to be seen, but the scale of the ambition is impressive.