Z.AI Launches China-Built AI Data Center Amid US Chip Restrictions

Chinese AI startup Z.AI, previously known as Zhipu, has completed construction of a 1-gigawatt data center designed to operate exclusively with domestically manufactured semiconductors. This development follows the company’s recent launch of its GLM-5.2 large language model and positions Z.AI as a key player in China’s drive for technological self-reliance in artificial intelligence. The move comes as US export controls prevent Z.AI from acquiring chips from companies like Nvidia.

US Export Controls Drive Domestic Strategy

Since January 2025, Z.AI has been on the US Department of Commerce’s Entity List, which restricts its ability to legally purchase Nvidia chips. This restriction has accelerated the company’s strategy to build its future infrastructure using only national suppliers. This aligns with Beijing’s broader objective to reduce its technological dependence on Washington amidst the global AI race.

The New Data Center: Scale and Purpose

The new data center has already begun partial operations and is intended for training and deploying Z.AI’s GLM family of models. At full capacity, the 1-gigawatt facility could consume as much electricity as approximately 750,000 households, making it one of the largest data centers ever constructed by a Chinese AI startup. Currently, the facility houses at least 10,000 China-sourced chips, though the proportion of total planned capacity this represents is not specified.

Key facts:

| Company | Z.AI (formerly Zhipu) |
| Data Center Capacity | 1 Gigawatt (at full operation) |
| Chip Source | Exclusively Chinese-made semiconductors |
| Primary Use | Training and deployment of GLM AI models |

Domestic Chip Landscape and Challenges

While the specific supplier of the chips is not detailed, all indicators point to Huawei. Z.AI’s GLM-5.2 model, unveiled in June, was entirely trained using Huawei’s Ascend accelerators, bypassing Nvidia hardware. This model has reportedly performed well in open-source model rankings like Design Arena.

However, challenges remain. Chinese chips, such as the Ascend series, are reported to offer lower performance per watt compared to Nvidia’s latest Blackwell architecture. This implies that a 1-gigawatt Chinese-powered data center may not yield the same computational capacity as one equipped with Nvidia hardware. Additionally, manufacturing bottlenecks persist. SMIC’s most advanced chip-making process (equivalent to a 7-nanometer node) is operating at over 93% capacity, and a shortage of domestic High Bandwidth Memory (HBM) limits Huawei’s ability to assemble more accelerators. Huawei reportedly shipped about 812,000 AI chips last year and aims to produce 1.6 million Ascend cores in 2026, sufficient for around 600,000 of its flagship 910C chips. Other Chinese rivals like Cambricon target 500,000 accelerators this year. This suggests that China’s capacity to build data centers may outpace its ability to produce the necessary chips to fill them.

Broader National AI Strategy

Z.AI’s data center initiative is part of a larger national plan by the Chinese government. Beijing is reportedly preparing a five-year investment of approximately 2 trillion yuan (around $295 billion USD) to establish a national network of AI data centers, with a target of at least 80% of technology sourced from Chinese providers. This ambitious plan includes Z.AI, Moonshot, and other emerging AI startups, underscoring a concerted effort to achieve AI leadership through domestic technological independence.

Source: Xataka IA – La startup china de IA que lanzó el sorprendente modelo GLM-5.2 acaba de alcanzar otro hito: crear su propio centro de datos (https://www.xataka.com/robotica-e-ia/startup-china-ia-que-lanzo-sorprendente-modelo-glm-5-2-acaba-alcanzar-otro-hito-crear-su-propio-centro-datos)

Datos clave