China has initiated a large-scale project to develop a network of artificial intelligence data centers, starting with a major hub on a 760-acre island in the city of Wuhu. This national strategy aims to close the significant gap in AI computing power with the United States by optimizing its data infrastructure and overcoming hardware limitations imposed by international sanctions.
Key Takeaways
- China is constructing a massive AI data center cluster in Wuhu, part of a Rmb270 billion ($37 billion) city-wide investment.
- The national plan involves building new data centers for AI response generation and networking existing remote facilities for model training.
- This strategy is a direct response to the U.S. holding approximately 75% of global AI computing power, compared to China's 15%.
- U.S. export controls on advanced chips from companies like Nvidia have forced China to find alternative solutions, including developing domestic networking technology.
A New National AI Infrastructure Plan
On a large island in the Yangtze River, former rice fields are being transformed into a high-tech campus of server farms. This development in Wuhu is a central part of Beijing's new, more coordinated approach to managing its AI resources. The goal is to build a powerful and efficient computing infrastructure capable of supporting the country's growing AI ambitions.
The strategy addresses a critical disparity. According to research group Epoch AI, the United States currently controls about three-quarters of the world's AI computing capacity. In contrast, China holds only 15%, a gap that Beijing is determined to narrow.
An executive from a supplier involved in the Wuhu project described the ambition as an effort to "build the Stargate of China," a reference to the massive $500 billion AI data center planned in the U.S. by companies including OpenAI and Oracle. While the Wuhu project is smaller, it represents a key piece of China's broader strategic vision.
Two-Part Strategy for Computing Power
In March, the Chinese government outlined a dual-focus plan. First, new data centers will be built near major urban areas to handle AI "inference," the process of generating answers from trained models. Second, existing data centers in remote western provinces will be repurposed for the intensive task of "training" large language models.
Strategic Hubs Across the Country
The Wuhu "Data Island" is a prime example of the new infrastructure plan. It will house four major data centers operated by state-owned giants: Huawei, China Telecom, China Unicom, and China Mobile. This cluster is strategically located to serve the economically vital Yangtze River Delta, which includes cities like Shanghai, Hangzhou, and Nanjing.
This model is being replicated nationwide to create a network of regional AI hubs:
- North: Ulanqab in Inner Mongolia will provide AI services to Beijing and Tianjin.
- South: Facilities in Guizhou province will support the major city of Guangzhou.
- Central: The city of Qingyang in Gansu province will serve Chengdu and Chongqing.
Ryan Fedasiuk, a fellow at the American Enterprise Institute, noted the shift in planning. "China is starting to triage scarce compute for maximum economic output," he said. "Beijing is now planning data centre infrastructure with this in mind."
Significant Local Investment
According to a local government notice, 15 companies have invested a total of Rmb270 billion ($37 billion) in building data centers across Wuhu. To attract this investment, the local government is offering generous subsidies, covering up to 30% of the cost of procuring AI chips.
Overcoming Technological Hurdles
China's AI ambitions face significant obstacles, primarily due to U.S. export controls. These restrictions prevent Chinese companies from accessing the most advanced processors from leading chipmaker Nvidia. Domestic alternatives from companies like Huawei and Cambricon have not yet been able to match the performance and scale of top-tier international hardware.
Further compounding the issue, the U.S. has prohibited leading semiconductor manufacturers like TSMC and Samsung from producing advanced AI chips for Chinese clients. Meanwhile, American tech giants such as Meta and Google are deploying tens of thousands of Nvidia's latest processors.
To acquire necessary hardware, some Chinese entities have turned to unofficial channels. Reports indicate that a network of intermediaries has emerged to secure banned Nvidia GPUs through the black market. However, Nvidia has stated that building large-scale data centers with smuggled components is not technically or economically viable.
"Trying to cobble together data centres from smuggled products is a non-starter, both technically and economically. Data centres are massive and complex systems, making smuggling extremely difficult and risky, and we do not provide any support or repairs for restricted products." - Nvidia Official Statement
Connecting Disparate Resources
Beyond acquiring new hardware, China is also focused on maximizing the efficiency of its existing resources. A construction boom from 2022 led to numerous data centers being built in energy-rich but remote provinces like Gansu and Inner Mongolia. Many of these facilities have been underutilized due to a lack of local demand and technical expertise, leaving valuable processors idle.
Instead of physically relocating servers, Beijing has mandated a technological solution. State-owned firms like China Telecom and Huawei are implementing advanced networking technology to link these scattered data centers into a single, cohesive computing cluster.
"A technical solution has to be found. That is connecting data centres," said Edison Lee, an analyst at Jefferies. This approach allows computing power from the west to be seamlessly directed to users in the east.
However, this method has drawbacks. Edward Galvin, founder of the data center research firm DC Byte, pointed out the inefficiencies. "Using multiple smaller and older data centres is less efficient than using one bigger and modern one," he explained. "It’s about economies of scale."
To address this, Huawei is developing a new networking technology called UB-Mesh. The company claims this system can double the training efficiency of AI models across multiple clusters by intelligently allocating tasks across the network. This technological innovation is seen as a crucial step in consolidating China's fragmented market and maximizing the utility of every available processor.