China is building a network of super data centers with NVIDIA chips despite US bans | Hardware catalog | Nvidia hardware | What is hardware | Turtles AI

Chinese companies are planning a network of 36 data centers in western regions, already under construction in Yiwu and Xinjiang, with the goal of hosting over 115,000 NVIDIA H100/H200 GPUs. Monetization is under close monitoring, with routes in Southeast Asia affected.

Key points:

• 36 data center project, with over 115,000 NVIDIA H100/H200 GPUs
• Locations chosen for climate and energy sources (wind, solar, coal)
• GPU supply hampered by US controls; use of alternative routes
• Transit reported via Singapore, Malaysia, Thailand, Vietnam

China has launched an ambitious project in the heart of its western regions, planning the construction of 36 new energy-intensive computing hubs, particularly in the Yiwu area (Xinjiang) and similar desert regions, where climate and renewable energy sources (wind, solar, coal) offer natural advantages. Official documents and tender notices report a projected 115,000 units of NVIDIA H100 and H200 GPUs, flagship chips designed for advanced AI training and inference.

However, access to such a large number of GPUs is complicated by restrictions imposed by the United States since 2022, requiring special licenses to export such high-end hardware to China—licenses that have so far not been granted. Bloomberg and other sources indicate that Beijing is exploring unofficial channels: routes via Singapore, Malaysia, Thailand, and Vietnam, using local companies as fronts to obtain or traffic GPUs to China.

In particular, the US government has launched investigations into illegal transit operations of servers containing NVIDIA chips from Southeast Asian countries to China, including a court investigation in Singapore involving Dell and Supermicro. Similar suspicions are targeting Malaysia and Thailand, with controls that could soon extend to exports to southeastern ports to block shipments.

Despite these scenarios, China is openly discussing projects like Nyocor’s, which involves the initial deployment of 250 H100 servers (approximately 2,000 GPUs) in partnership with AI organizations like Infinigence. Although experts estimate that the volume of prohibited chips currently arriving in China amounts to approximately 25,000 units, a figure far from the announced plans, it is clear that the black market continues to operate, albeit not on a large enough scale.

Domestically, Beijing continues to invest in domestic solutions: chips like the Huawei Ascend 910C and A800/H800/H20 are becoming widespread, although they still have performance gaps compared to the most advanced NVIDIA GPUs. Chinese authorities are also encouraging large tech companies to gradually reduce their dependence on foreign hardware and adopt domestic alternatives.

Meanwhile, NVIDIA is developing chips compliant with US regulations, such as the B30, B40, and detuned versions of the H20 series specifically designed for the Chinese market, thus seeking to maintain market share while adhering to restrictions.

The "hyperscale" data center project in China is in an advanced planning stage, and infrastructure is taking shape in strategic locations; however, the real issue remains the procurement of NVIDIA H100/H200 GPUs at the necessary scale. Routes through Southeast Asia, trade transit patterns, and growing technological autonomy suggest that Beijing is adopting a patchwork of hybrid strategies to circumvent barriers, but it remains to be seen whether this will be sufficient to guarantee the envisioned computing power.

The success of these projects will depend on China’s ability to balance technological development and circumventing international restrictions in an increasingly complex and highly controlled geopolitical environment.