Beijing completes a 1-gigawatt data center running entirely on Chinese-made processors, demonstrating domestic semiconductor self-sufficiency.

A Beijing-based AI company has launched a 1-gigawatt data center running entirely on domestically produced semiconductors. Z.AI's facility is now operational and purpose-built to train the company's GLM (General Language Model) AI platforms without a single Nvidia chip.

Since 2025, US export controls have forced Chinese companies to pursue alternative sourcing strategies for AI compute. Washington's restrictions specifically targeted high-end Nvidia GPUs such as the H100 and its successors—chips that have become the backbone of AI training worldwide.

China's domestic semiconductor efforts are accelerating. In April 2026, Alibaba launched its own data center equipped with 10,000 Zhenwu AI chips, proprietary silicon designed specifically for AI workloads. Z.AI's facility significantly exceeds it in both scale and ambition.

The Chinese government has unveiled a five-year plan with an estimated budget of approximately $295 billion to create a network of interconnected AI data centers, targeting 80% domestic technology sourcing across the entire network. New policies now mandate the use of domestic chips in state-funded projects, effectively creating a guaranteed market floor for Chinese chip manufacturers. This assured demand allows them to scale production, iterate on designs, and drive down costs without needing to compete directly with Nvidia for commercial customers.

For Nvidia, Z.AI's data center represents proof that China can build AI infrastructure at massive scale without its products. Every gigawatt of Chinese compute capacity running on domestic silicon represents a gigawatt that will never generate Nvidia revenue, even if export controls are eventually relaxed.

For the broader AI infrastructure industry, the critical metric is not whether Chinese chips match Nvidia's benchmarks on paper, but whether Chinese AI models trained on domestic hardware can compete with models trained on Nvidia silicon. DeepSeek's earlier breakthroughs have already demonstrated that sophisticated engineering can partially compensate for hardware limitations.