Z.AI 1GW Data Center Without Nvidia Chips

A Paradigm Shift in Hardware Architecture

The creation of massive computing infrastructure for artificial intelligence has traditionally been associated with the products of a single American company. However, the Chinese research laboratory Z.AI, also known as Zhipu AI, has officially completed the deployment of a new data center with a total capacity of 1 GW. The main technical feature of this facility is the complete absence of hardware accelerators from Nvidia. The entire architecture is built exclusively on domestic Chinese semiconductor solutions, marking an important precedent for the entire technology industry. This step indicates serious shifts in the global supply chain of high-tech equipment.

The implementation of such a project demonstrates the ability of Chinese engineers to design, build, and operate ultra-high-density computing clusters under severe U.S. export restrictions. The data center integrates several autonomous segments, each containing over 10,000 specialized processors. This allows the company to train advanced models with hundreds of billions of parameters, ensuring the necessary bandwidth and a stable level of data synchronization at the hardware level. The successful launch of such a complex proves that alternative silicon ecosystems are capable of meeting the highest demands of the artificial intelligence industry.

Export Controls as a Catalyst for Innovation

In recent years, the U.S. government has implemented a series of strict export controls that have prevented the supply of high-performance graphics processing units, such as the A100 or H100, to China. Initially, these measures were intended to restrain the development of the Chinese machine learning sector. Instead, the restrictions acted as a powerful stimulus for internal investment in the development of alternative architectures. Companies were forced to reconsider their strategies and turn to internal resources, which significantly accelerated the development of local technology companies.

Rather than attempting to bypass sanctions through gray supply chains, the Z.AI laboratory invested resources in creating an infrastructure that relies entirely on local suppliers. This decision allowed them not only to avoid risks associated with the logistics of foreign components but also to optimize the hardware for the company’s specific tasks related to training large language models. This approach ensures greater control over the hardware lifecycle and independence from geopolitical factors.

Engineering a Gigawatt-Scale Facility

Designing a data center with a 1 GW capacity requires solving unprecedented engineering challenges. Such energy consumption is comparable to the needs of a small city. To ensure the stable operation of the infrastructure, Z.AI integrated advanced power distribution and thermal management systems. The scale of the project required the development of new standards for routing cable paths and organizing uninterrupted power supply at the level of each individual server rack.

Power Infrastructure and Liquid Cooling

Traditional air cooling systems are unable to efficiently dissipate heat from high-density AI accelerator racks. Therefore, the new data center massively employs direct liquid cooling technologies. A special coolant circulates directly through radiators mounted on the processors, making it possible to maintain optimal temperature levels even during peak workloads. This not only prevents overheating but also significantly extends the lifespan of the silicon dies.

The power supply system is built on the principle of deep redundancy. The use of intelligent power distribution units allows for the dynamic redistribution of loads between different clusters depending on current tasks, minimizing electricity losses and improving the facility’s overall Power Usage Effectiveness (PUE) rating to record-low values.

Hardware Cluster Architecture

Chinese neural network accelerators form the core of the system’s computing power. Market analysts point to the massive use of Huawei Ascend 910B series processors as the most likely underlying component. These processors are designed specifically for deep machine learning tasks, feature custom-architecture tensor cores, and are direct competitors to Western equivalents in the segment of training large datasets.

To effectively connect tens of thousands of chips into a single computing network, engineers abandoned the closed InfiniBand protocol in favor of an optimized version of RoCEv2 (RDMA over Converged Ethernet). This provides high-speed data transfer with minimal latency directly to the RAM of adjacent nodes, bypassing the central processor. This technology is critical for minimizing latency during distributed neural network training.

Building an Independent Software Ecosystem

Hardware is merely the foundation of the infrastructure. Historically, the main advantage of Nvidia’s solutions has been the CUDA ecosystem, which became an industry standard. To effectively utilize Chinese chips, Z.AI engineers had to develop and adapt a robust software stack capable of translating complex mathematical operations onto the architecture of domestic tensor cores without sacrificing performance.

Adapting the CANN Environment

The laboratory’s specialists integrated their own solutions with open frameworks and the CANN (Compute Architecture for Neural Networks) environment. This software stack includes specialized tools for automatic load balancing across thousands of nodes, innovative gradient compression mechanisms during synchronization, and intelligent systems for isolating faulty services.

If one of the chips fails during training, the system can isolate it, redistribute the load to neighboring nodes, and continue the process without entirely stopping the cluster. The results of extensive stress tests confirmed that the infrastructure can operate stably for many weeks under continuous extreme workloads, which is necessary for training modern LLMs.

Training the GLM Series Models

This entire gigantic infrastructure was built with a clear purpose – accelerating the development and training of the General Language Model (GLM) series. The Z.AI laboratory has already proven the competitiveness of its algorithms by releasing models that, in many benchmarks, approach the capabilities of leading Western counterparts. Having its own computing power will allow the company to significantly speed up the release cycle for new versions.

Pipeline Parallelism

Using clusters of over 10,000 nodes drastically reduces the training time for models with over 500 billion parameters. The increased bandwidth of the internal network allows for the use of larger batch sizes and more efficient parallelization of computations, both at the tensor level and at the pipeline level (pipeline parallelism). The algorithms automatically determine the optimal way to partition the computation graph to maximize the utilization of each available core.

The successful launch of training sessions in the new data center proves that the hardware lag in lithography is partially compensated by massive horizontal scaling, sophisticated network architecture, and deep optimization of software code for specific workloads.

Global Economic Implications

The launch of a computing facility of this scale without components from the dominant American supplier has far-reaching consequences for the entire global semiconductor industry. It is a signal to other countries about the possibility of building sovereign digital infrastructure.

• Market Diversification The success of the project encourages other large tech companies to invest in their own processor designs, reducing reliance on monopolies.
• Stimulating Competition The emergence of strong players in the AI hardware sector will force existing leaders to lower prices and accelerate innovation.
• New Industry Standards The experience of building gigawatt systems with liquid cooling and RoCE networks sets new industry standards for designing the data centers of the future.

In the long run, the emergence of powerful alternative ecosystems will help lower the cost of computing power and further improve AI processor architectures beyond traditional platforms. The Z.AI project is just the first step in forming a bipolar world in the artificial intelligence industry, where the dominance of a single architecture is gradually becoming a thing of the past.