China plans 39 AI data centers equipped with approximately 115,000 restricted Nvidia Hopper GPUs, raising questions about sourcing despite US export bans.

Chinese companies are preparing to equip 39 new AI data centers — mostly in Xinjiang and Qinghai — with over 115,000 high-performance Nvidia Hopper GPUs, whose shipments to China are restricted by U.S. export rules. According to Bloomberg's review of official investment documents, tenders, and filings, 70% of the processing capacity — enabled by around 80,500 Nvidia H100 and H200 GPUs — is expected to be concentrated in a single state-owned data center located in Yiwu County, Xinjiang. The remaining 30% will be spread across at least 38 additional data center projects, largely in Xinjiang outside of Yiwu and in Qinghai province, including a larger project run by Nyocor that plans to install 625 H100 DGX servers with around 5,000 H100 accelerators in multiple phases, starting with 250 8-way machines containing 2,000 H100 GPUs.

To contextualize the scale of this deployment, it took Elon Musk's xAI around 100,000 H100 processors to train the Grok 3 AI model, one of the most advanced currently available, while DeepSeek trained its R1 model using a GPU cluster of 50,000 Nvidia Hopper GPUs consisting of 30,000 H20 HGX units, 10,000 H800s, and 10,000 H100s. If completed as planned, the Yiwu data center with 80,500 GPUs could rank among the most powerful AI clusters in the People's Republic, significantly strengthening China's AI infrastructure to train advanced large language models and large reasoning models. For reference, Xinjiang has already built a data center providing "24,000 PetaFLOPS of processing power," said to be equivalent to around 12,000 Nvidia H100s, available to other cities like Chongqing. Local officials offer a 20% discount on electricity, along with financial and housing incentives for experts in AI and green technologies.

Chinese operators and their customers indicated in permit documents what they plan to use, suggesting they require the high performance of H100 and high memory capacity of H200 for their projects rather than lower-cost alternatives. H100 can be 3.3 to 6.69 times faster than cut-down H20 with AI data formats and 1.52 to 64 times faster with HPC data formats. Theoretically, satisfying the computing needs of 115,000 H100 GPUs would require 380,000 to 770,000 H20 units as substitutes. Bloomberg estimates that completing all 39 projects as envisioned would require procurement of more than 14,000 servers using either H100 or H200 processors, worth billions of dollars on China's black market.

While sources with knowledge of U.S. government investigations were unaware of the specific Xinjiang projects, they confirmed the existence of some unauthorized Nvidia hardware in China but expressed doubt that any organized network could supply over 100,000 restricted processors to one country, let alone in a single region. Estimates on the total number of such chips in China vary significantly; two senior officials in the Biden administration mentioned a figure closer to 25,000, far less than what the Chinese projects require. To date, there is no direct proof that China has accumulated or will soon receive the more than 115,000 restricted GPUs outlined in these construction plans. Nonetheless, work on the facilities continues, with a large solar power tower erected in Yiwu to provide consistent electricity. The location was chosen for its access to solar and wind energy, inexpensive land, and high elevation to cool hardware.