For the past year, a persistent question has surrounded the rapid advancement of Chinese artificial intelligence models: what hardware is powering them? Z.AI, the company previously known as Zhipu, has now provided a partial answer with the construction of a massive data centre that runs entirely on domestic chips. According to a report by Bloomberg, citing a person familiar with the matter, Z.AI has begun partial operations at a 1-gigawatt facility designed to train its GLM family of models. The company did not respond to requests for comment, but the implications of this buildout are significant for the global AI landscape.
The Scale of the Operation
The sheer size of the data centre is noteworthy. A single gigawatt of power capacity is roughly equivalent to the electricity consumption of 750,000 homes at any given moment. This puts the Z.AI facility among the largest server hubs ever built by a Chinese AI lab, surpassing even some of the largest cloud data centres operated by Alibaba and Tencent. The chip count is equally striking. The facility comprises several computing clusters, each housing more than 10,000 chips. Critically, none of these chips come from Nvidia, the American semiconductor giant that dominates the global AI accelerator market.
Bloomberg did not specify the exact chip suppliers, but China's leading designer of AI accelerators is Huawei, whose Ascend series has become the primary domestic alternative. Other players include Cambricon, a company that develops AI chips for cloud and edge computing, and Alibaba, which produces its own Hanguang 800 processors. These chips, while still lagging behind Nvidia's A100 and H100 in raw performance, have made significant strides in recent years, particularly in terms of software ecosystem and integration with China's domestic supply chain.
Why This Matters for the AI Race
US export controls, first imposed in 2022 and tightened in 2023, have blocked Chinese labs from purchasing Nvidia's most advanced chips, such as the A100, H100, and the subsequent Blackwell architecture. Chinese companies have been limited to older or specially designed versions, such as the A800 and H800, which were later also restricted. This has forced Chinese AI labs to seek alternatives, but there was an open question about whether homegrown parts could carry the load for training frontier models, not just for inference or smaller-scale tasks.
A 1GW cluster built entirely on domestic silicon provides a real-world answer. It demonstrates that Chinese labs can continue to scale their training capabilities even while cut off from the chips that the rest of the industry treats as default. Z.AI's move suggests that the bottleneck created by export controls may be less severe than initially feared, at least for companies with the resources and engineering talent to integrate thousands of domestic chips into a stable, high-performance system.
Timing and Competitive Pressure
The buildout comes at a tense moment for China's AI industry. Just days earlier, Beijing-based Moonshot AI launched its Kimi K3 model, which reportedly matched top US models in benchmarks but then ran short of compute capacity and had to pause new user sign-ups. This highlights the intense demand for training infrastructure, as Chinese labs race to keep pace with OpenAI, Google DeepMind, and Anthropic. Z.AI is betting on owning its hardware to avoid such shortages and to gain a competitive edge.
Z.AI itself has been a prominent player in China's AI scene. Originally founded as a research lab under the name Zhipu, it rebranded to Z.AI as it shifted toward a commercial focus. The company is best known for its GLM (General Language Model) series, which has been positioned as a direct competitor to OpenAI's GPT-4. Z.AI has secured significant funding from investors including Alibaba, Tencent, and state-backed funds, and it has been expanding its enterprise AI offerings, similar to the strategy of Anthropic in the US.
China's Broader Data Centre Buildout
Z.AI is not acting alone in this hardware push. China is preparing to spend approximately 2 trillion yuan (about $295 billion) over the next five years on data centre construction across the country. This investment is part of a national strategy to build a self-sufficient digital infrastructure, reduce reliance on foreign technology, and support the development of cutting-edge AI models. Cloud giants like Alibaba and China Telecom have been the biggest builders so far, but newer entrants like Z.AI are also making significant capital expenditures.
The financial health of Z.AI further supports its aggressive expansion. The company recently listed on the Hong Kong Stock Exchange and conducted a follow-on share sale. It is on track to achieve $1 billion in annual recurring revenue, which would make it the first Chinese AI firm to reach that milestone. The day the data centre news broke, Z.AI's shares jumped nearly 20%, reflecting investor confidence in its hardware-first strategy and its potential to compete globally.
The Hardware Challenge: Domestic Chips in Focus
While the use of domestic chips is a milestone, the details matter. The chips powering Z.AI's clusters are likely Huawei's Ascend 910B or similar accelerators. These chips are designed to handle large-scale AI training and inference, but they still face significant challenges. Raw performance per chip is lower than Nvidia's top products, and the software stack, including libraries and frameworks, is less mature. Stitching together thousands of chips into a reliable, low-latency network is a complex engineering feat, especially when the interconnect technology (such as Huawei's HCCS) differs from Nvidia's NVLink.
To compensate, Chinese chipmakers have focused on improving chip-to-chip communication and memory bandwidth, but early benchmarks suggest that training efficiency on domestic hardware can be lower. However, the gap is narrowing. Huawei's latest Ascend 910C is rumoured to offer performance closer to Nvidia's H100, though it has not yet been officially launched. Z.AI's willingness to commit to a 1GW cluster indicates that it believes the performance shortfall can be managed through scale and software optimizations.
Potential Limitations and Caveats
It is important to approach this news with a degree of caution. The chip details come from an anonymous source, and Z.AI has not confirmed the specifics of the hardware. Building a cluster is not the same as proving it can train a frontier model as efficiently as an Nvidia-based system. The real test will come when Z.AI releases its next GLM model, trained entirely on this domestic infrastructure. If that model can match the capabilities of GPT-4 or Claude 3.5, it will validate the viability of a non-Nvidia training stack.
Until then, the industry will watch closely. The Chinese government's support for domestic chip production and the massive investment in data centre infrastructure suggest that the direction is clear. The effort to build a Chinese alternative to Nvidia has moved from slideware and prototypes to a working, gigawatt-scale data centre. This changes the debate about how far export controls can hold back China's AI ambitions. Even if the chips are not yet as powerful individually, the aggregate capacity and the rapid iteration cycle of Chinese hardware manufacturers could close the gap faster than many observers expect.
In the coming months, analysts will be scrutinizing energy efficiency benchmarks, training time, and model quality from Z.AI. For now, the company has made a bold statement: it is willing to bet on domestic silicon, and it has the resources to build at scale. The rest of the industry will have to take note.