China Races to Build Massive AI Computing Clusters

Cities and technology companies across China are building massive artificial intelligence (AI) computing clusters, aiming to accelerate model training, reduce costs, and broaden AI adoption. These facilities, often linking 10,000 or more AI accelerator chips, have emerged as a new form of infrastructure over the past two years.

Credit: UNSPLASH

These powerful clusters function as supercomputers, integrating high-performance graphics processing units (GPUs) and advanced storage into a single system. They significantly reduce model training times, enabling faster iteration of AI capabilities.

Chinese domestic champions, including tech giants such as Huawei Technologies and Alibaba Group Holding, along with GPU specialists like Moore Threads, are actively competing to supply chips for these systems. Only a limited number of domestic providers can deliver chips for such large-scale systems.

Orient Securities identified Huawei, Hygon Information Technology, and Moore Threads as first-tier players in this segment. An Orient Securities report stated, "Having the capability to deliver clusters with tens of thousands of cards signifies a firm foothold in the most intensely competitive area of Sino-US tech competition."

Cities throughout China are vying to host these significant projects. Many are led by the country’s three state-backed telecoms operators, often collaborating with partners across the value chain.

ByteDance and Alibaba are also constructing private clusters. These are intended to support their internal AI development efforts and provide cloud services to clients.

For example, Alibaba recently deployed a 10,000-card cluster in Shaoguan, Guangdong province, powered by its Zhenwu AI chips. This deployment was a partnership with China Telecom.

This followed the launch in March of a 10,000-card intelligent computing cluster, built with Huawei’s Ascend 910C AI chips in Shenzhen. Elsewhere, Harbin has become a northern hub for AI infrastructure.

China Mobile is set to launch the country’s largest single-operator intelligent computing centre there in late 2024, which will feature more than 18,000 domestic AI accelerator cards. In central China, Zhengzhou activated a 60,000-card cluster in April, built by state-backed supercomputer maker Sugon for scientific AI applications.

A key trend in this development is the shift towards home-grown chips. This aligns with China’s broader push for semiconductor self-sufficiency, a response to US export controls.

The Huawei-powered cluster in Shenzhen is reportedly the first in China to be fully equipped with domestically developed advanced chips. Commercialisation of these clusters is taking shape along two distinct tracks.

These tracks are state-driven demand and private-sector use. Chief Economist Charlie Zheng of Samoyed Cloud Technology Group, stated that adoption was likely to follow a "multi-tier system."

Zheng predicted that government, finance, and state-owned enterprises would lead uptake due to data sovereignty requirements and robust budgets. He added that the next wave of adoption was expected to come from healthcare and education.

Universities, for instance, are anticipated to rent computing power from these clusters to support their research initiatives. Companies like ByteDance and Alibaba employ a different model.

They are building and operating their own clusters to reduce reliance on external computing resources. This approach aims to lower long-term operational costs for these technology companies.

As AI evolves towards multimodal systems, model sizes are expanding rapidly, with parameter counts growing tenfold each year. Analysts anticipate cluster sizes will scale beyond 100,000 cards, potentially reaching up to one million cards.

This rapid expansion has raised concerns about overinvestment and excessive competition within the industry. However, analysts believe some inefficiency is unavoidable in the context of the US-China technology race.

Beijing-based computing consultant Li Yangwei said, "To win the ultimate competition with the US, there will certainly be some waste. It’s just a matter of how much."

Li compared the current buildout to China’s early development of solar power and electric vehicles. He added, "Without external ‘catfish’ like Nvidia, domestic players across model development and chip manufacturing will need to compete fiercely."

Li explained that "catfish" is an industry term describing external rivals that spur competition. He also warned, "The first wave of companies is likely to be washed up on the beach."

---

• China is rapidly constructing massive AI computing clusters with 10,000 or more accelerator chips to enhance model training and reduce costs.

• These clusters function as supercomputers, integrating high-performance GPUs and advanced storage for faster AI capability iteration.

• Domestic champions like Huawei, Alibaba, and Moore Threads, alongside state-backed telecoms, are key players in this drive for semiconductor self-sufficiency.

Source: SCMP