- Joined
- Aug 19, 2017
- Messages
- 3,251 (1.12/day)
China's leading AI companies have unveiled plans to construct a massive network of 36 data centers across China's Western deserts, which will house over 115,000 NVIDIA AI processors. According to a Bloomberg analysis, which gained insight into investment approvals and tender documents, the critical location for this effort is a complex facility situated near Yiwu in Xinjiang Province, selected for its suitable wind and solar resources, coal reserves, and cooler high-altitude climate. Chinese AI labs aim to deploy NVIDIA's flagship H100 and H200 GPUs, which are primarily sourced from third-party suppliers. One Chinese firm proposes an initial phase of 625 H100 servers, equivalent to roughly 2,000 chips, with additional phases to follow.
A key obstacle to this great vision is US export controls that prohibit the sale of NVIDIA's most advanced processors to China without special licenses, which have not been granted. None of the firms or government spokespersons Bloomberg contacted have explained how they will acquire these embargoed GPUs, and experts familiar with trade enforcement and underground AI chip markets express skepticism that such a large volume of cutting-edge chips could be smuggled undetected. If direct access to NVIDIA hardware remains limited, Chinese chipmakers such as Huawei may fill part of the demand. Construction in Yiwu presses on regardless, where a slogan painted on a nearby hillside proclaims "data‑electricity fusion shows great promise," reflecting China's determined push to take a leading role in the next generation of AI innovation, despite the hardware slowdown.
Huawei-made systems, such as the CloudMatrix 384, are viable alternatives. However, manufacturing these chips still proves to be complex, even for SMIC. Huawei has expressed frustration with its manufacturing workflows and slow turnaround times. The company has built its own complete semiconductor supply chain as a response. Chinese AI labs rush to acquire more NVIDIA chips, Huawei is gradually working on replacing them; however, the demand for NVIDIA hardware remains massive, as the entire training and inference operation is carried out most effectively on CUDA-accelerated hardware.
View at TechPowerUp Main Site | Source
A key obstacle to this great vision is US export controls that prohibit the sale of NVIDIA's most advanced processors to China without special licenses, which have not been granted. None of the firms or government spokespersons Bloomberg contacted have explained how they will acquire these embargoed GPUs, and experts familiar with trade enforcement and underground AI chip markets express skepticism that such a large volume of cutting-edge chips could be smuggled undetected. If direct access to NVIDIA hardware remains limited, Chinese chipmakers such as Huawei may fill part of the demand. Construction in Yiwu presses on regardless, where a slogan painted on a nearby hillside proclaims "data‑electricity fusion shows great promise," reflecting China's determined push to take a leading role in the next generation of AI innovation, despite the hardware slowdown.
Huawei-made systems, such as the CloudMatrix 384, are viable alternatives. However, manufacturing these chips still proves to be complex, even for SMIC. Huawei has expressed frustration with its manufacturing workflows and slow turnaround times. The company has built its own complete semiconductor supply chain as a response. Chinese AI labs rush to acquire more NVIDIA chips, Huawei is gradually working on replacing them; however, the demand for NVIDIA hardware remains massive, as the entire training and inference operation is carried out most effectively on CUDA-accelerated hardware.
View at TechPowerUp Main Site | Source
.