China is rapidly expanding and reorganizing its data center infrastructure in a bid to close the gap with the United States in artificial intelligence, as part of what officials and industry executives are calling the country’s “Stargate of China” plan, the Financial Times reported.
On a 760-acre island on the Yangtze River in Wuhu, rice fields are being transformed into a “mega-cluster” of server farms. The initiative forms a cornerstone of Beijing’s strategy to consolidate fragmented data centers, boost efficiency, and maximize limited access to advanced processors as AI demand surges.
Although the Wuhu project will not rival the scale of the $500 billion “Stargate” AI data center complex being developed in Texas by OpenAI, Oracle, and SoftBank, it is a critical piece of China’s national effort to manage scarce computing resources. The Financial Times noted that Wuhu’s so-called “Data Island” already hosts four major AI facilities operated by Huawei, China Telecom, China Unicom, and China Mobile, with total investment across the city reaching Rmb270 billion ($37 billion).
Closing the US gap
China faces a steep challenge: US researchers hold about three-quarters of global AI computing capacity, compared with just 15% in China, according to Epoch AI. Export restrictions imposed by Washington have barred Chinese groups from buying Nvidia’s most advanced AI chips, while also preventing TSMC and Samsung from manufacturing them for Chinese clients.
American tech giants — including Meta, Google, and Elon Musk’s X.ai — are deploying tens of thousands of Nvidia’s latest processors, leaving China scrambling to optimize what it has. Local chipmakers like Huawei and Cambricon are struggling to fill the void.
Networking fragmented resources
Beijing’s approach involves reconfiguring existing data centers. Remote facilities in western provinces such as Gansu and Inner Mongolia are being directed to focus on training large language models, while new clusters near cities like Shanghai, Hangzhou, and Guangzhou will handle inference — the process of generating AI responses — to reduce latency for end-users.
To overcome inefficiencies, Chinese telecom groups including Huawei and China Telecom are networking disparate processors into unified clusters, using advanced transponders, switches, and routing software. Huawei is also developing UB-Mesh, a new networking technology it claims can double the efficiency of training large models across multiple clusters.
“The move to network the chips together is an important method of consolidation in this fragmented market,” said Edison Lee, an analyst at Jefferies.
Subsidies and smuggling risks
Local governments are sweetening the effort with subsidies, covering up to 30% of AI chip procurement costs in Wuhu — more generous than in other regions, according to officials cited by the Financial Times.
At the same time, a shadow network of intermediaries has emerged to circumvent US restrictions by smuggling Nvidia GPUs into China. One such Wuhu-based supplier, Gate of the Era, has reportedly acquired large quantities of banned Nvidia Blackwell servers, though it declined comment. Nvidia has dismissed such efforts as technically unviable, saying: “Trying to cobble together data centres from smuggled products is a non-starter.”
A fragmented path to scale
Analysts warn that using multiple smaller, older data centers is less efficient than building large modern ones. “It’s about economies of scale,” said Edward Galvin of consultancy DC Byte. Still, Beijing’s strategy reflects both necessity and ambition: making the best use of limited hardware, while laying the foundation for longer-term AI dominance.
“China is starting to triage scarce compute for maximum economic output,” said Ryan Fedasiuk of the American Enterprise Institute. “Beijing is now planning data centre infrastructure with this in mind.”
