More than enough: China’s AI data centers overextend as facilities sit empty and GPU prices plummet

China‘s AI data center boom is rapidly deflating as newly constructed facilities sit empty and GPU prices plummet. This unraveling infrastructure bubble illustrates the risks of speculative tech investment, as government-backed and private companies poured billions into AI computing resources that now face weak demand. The situation has been further complicated by the rise of more efficient AI models like DeepSeek, which require less computational power and have fundamentally shifted the economics of the AI industry in China.

The big picture: China’s AI infrastructure gold rush has turned into a crisis with up to 80% of newly built computing resources reportedly sitting unused, as data center operators struggle to find customers for their GPU rental services.

• Just months ago, Chinese AI infrastructure was booming with contractors and investors rushing to build data centers filled with high-performance GPUs, many smuggled through overseas channels to Shenzhen.
• At the peak of demand, a single restricted Nvidia H100 chip could fetch up to 200,000 yuan ($28,000) on the black market, but now prices have dropped significantly as the market shifts from scarcity to oversupply.

Behind the numbers: The sudden collapse of demand stems from fundamental miscalculations about the AI computing market and changing technological requirements.

• Many data center projects were speculative investments based on the assumption that GPU rental would be consistently profitable, but this model has proven unsustainable as supply has rapidly outpaced actual industry needs.
• GPU traders who previously boasted about securing shipments of restricted Nvidia chips are now more discreet, with many project managers reporting they’re being forced to sell surplus chips in a market where “everyone is selling, but few are buying.”

Industry disruption: The rise of DeepSeek, a Chinese AI model renowned for its efficiency, has dramatically altered computational requirements and undermined the business case for massive GPU infrastructure investments.

• DeepSeek’s approach demonstrated that powerful AI models could be developed with fewer computational resources, making expensive data centers filled with top-tier Nvidia chips less necessary than previously thought.
• This technological shift has left many data center operators with costly infrastructure investments that may never generate their expected returns.

What they’re saying: “It seems like everyone is selling, but few are buying,” says Xiao Li, a real estate contractor who pivoted to becoming a data center project manager during the 2023 AI boom.

Implications: The data center bubble bursting reveals deeper issues with China’s approach to AI development and highlights the risks of infrastructure-focused technology investment.

• The situation demonstrates how quickly technological developments can render massive infrastructure investments obsolete, particularly in rapidly evolving fields like artificial intelligence.
• This market correction may ultimately lead to a more sustainable AI ecosystem in China that focuses on innovative models and applications rather than raw computing power accumulation.