Analysis: Huawei's Horinger data center, among China's largest for AI, has ~7% of the compute of SpaceX's Colossus 2 and may stay far behind until at least 2028
First reported by Epochai.substack ·
Your AI compute capacity is about 7% of the global leader, and this gap is unlikely to close soon.
Huawei's Horinger data center in Inner Mongolia, one of China's largest for AI, has an estimated IT power capacity of 286 megawatts across four operational buildings. Utilizing its domestically produced Ascend 910C and 910B chips, the facility currently offers approximately 81,000 H100-equivalents in AI compute capacity. This positions Horinger with about 7% of the compute power of SpaceX's Colossus 2, the world's largest AI data center. While Huawei plans to expand Horinger with at least three new buildings, projecting a 70% increase in compute power by early 2028, it is still expected to trail Colossus 2 significantly. US export controls on advanced chip technology prevent Huawei from matching the performance and efficiency of competitors like Nvidia. The Horinger facility is a case study illustrating China's substantial investment in AI infrastructure, contrasted with the performance gap in chip technology and overall compute capacity compared to leading US facilities.
Despite China's substantial investment in AI data center infrastructure, including large facilities like Huawei's Horinger campus, a significant performance gap persists compared to US counterparts due to US export controls on advanced chip technology. Huawei's Horinger data center, while substantial by Chinese standards and leveraging domestic chip production, offers only about 7% of the compute power of the leading US facility, Colossus 2. This disparity highlights how geopolitical restrictions continue to limit China's ability to achieve parity in cutting-edge AI hardware, impacting its potential for AI development and deployment scalability.
The projected build-out of the Horinger facility by 2028, aiming for a 70% increase in compute power, will still leave it far behind the projected growth of US data centers. This suggests that while China can amass significant physical infrastructure, its access to the most advanced and efficient AI chips remains a critical bottleneck. Consequently, the frontier of AI model training and inference capabilities in China is likely to continue lagging behind the US, influencing global AI research and commercialization trends for the foreseeable future.
AI-written summary. May contain errors.