Z.ai Completes 1GW Data Center Running Entirely on Chinese Chips

Domestic Silicon Powers Massive AI Facility

Chinese artificial intelligence developer Z.ai has completed construction of a 1-gigawatt data center that operates entirely on chips manufactured within China, representing a concrete advance in the country's campaign to circumvent Western export controls on high-end semiconductors.

The facility's scale and its reliance on domestic processors signal that Chinese AI companies are moving beyond pilot projects and into production-grade infrastructure built without access to restricted silicon from Nvidia and other US chipmakers. The 1GW power envelope places the installation among the largest AI-focused data centers globally, comparable to hyperscale facilities operated by major cloud providers.

Z.ai has not disclosed which specific Chinese chip architectures power the data center, though the domestic AI accelerator market includes offerings from Huawei's Ascend line, Biren Technology, and Enflame Technology. Each has developed processors aimed at training and inference workloads traditionally dominated by Nvidia's A100 and H100 GPUs.

Beijing's Self-Sufficiency Drive Takes Physical Form

The completion comes as Washington has tightened export controls on advanced semiconductors and chipmaking equipment, blocking Chinese firms from purchasing cutting-edge AI accelerators and the tools needed to manufacture them domestically at leading-edge nodes. Those restrictions, expanded in October 2022 and again in October 2023, aim to prevent China from using US technology to advance military AI capabilities.

Chinese companies and research institutions have responded by accelerating development of alternative architectures and scaling up production on older process nodes that remain accessible. While these chips typically lag Nvidia's latest offerings in raw performance and energy efficiency, deploying them at scale allows Chinese AI developers to continue training large models and running inference services.

The data center's 1GW power capacity suggests Z.ai has prioritized buildout speed and volume over waiting for more advanced domestic chips. Industry analysts estimate that power budget could support tens of thousands of AI accelerators, depending on the specific processors deployed and cooling architecture.

Regional Implications for Asia's AI Race

The project underscores a broader shift in Asia's AI infrastructure landscape. As US export controls segment the global semiconductor market, Chinese firms are building parallel supply chains that exclude American components wherever possible. This bifurcation creates opportunities for suppliers in Japan, South Korea, and Taiwan who can navigate both ecosystems, while also spurring investment in domestic alternatives across the region.

Singapore and other Southeast Asian nations are watching closely, balancing their roles as neutral data center hubs with pressure from both Washington and Beijing. The viability of large-scale Chinese AI infrastructure built on domestic chips will influence how regional governments and investors allocate capital in the coming years.

Z.ai's facility also arrives as Chinese AI companies race to keep pace with OpenAI, Google, and Anthropic in model capabilities. Access to sufficient compute remains a binding constraint, making the ability to deploy and operate gigawatt-scale data centers a strategic imperative regardless of the chips inside.

What Comes Next

The data center's operational performance will be closely monitored by competitors and policymakers alike. If Z.ai can demonstrate that domestic chips deliver acceptable training speeds and inference throughput at scale, other Chinese AI developers are likely to accelerate similar buildouts, further entrenching the bifurcated semiconductor landscape.

For now, the facility stands as evidence that export controls have not halted China's AI infrastructure expansion, but have instead redirected it toward domestic supply chains. Whether those supply chains can sustain the pace of innovation required to remain competitive in foundation model development remains an open question, one that will be answered in the training runs and product releases that follow.