Zhipu Completes 1-Gigawatt AI Data Center Built Entirely on Domestic Silicon

A Rebound Built on Domestic Infrastructure

Shares of Zhipu, the Beijing-based AI developer also known as Z.ai, climbed 37 percent in Hong Kong trading on Tuesday, closing at HK$1,219. The rally followed confirmation that the company has finished building a 1-gigawatt AI computing center relying solely on domestically manufactured processors. The single-session gain reversed more than a week of declines that had erased over 40 percent of the stock's value, underscoring how quickly investor sentiment can shift when infrastructure milestones land in an environment of constrained chip supply.

At DailyTechWire, we've tracked the rising stakes around compute capacity across Chinese AI labs. Zhipu's move to bring a gigawatt-scale facility online using only local silicon represents one of the most concrete tests yet of whether China's semiconductor ecosystem can support frontier model training without access to the latest Nvidia accelerators. The facility will serve as the primary training and inference backbone for the company's GLM family of large language models, which compete directly with offerings from Baidu, Alibaba, and a crowded field of domestic players.

Scale and Sourcing in the Shadow of Export Controls

A 1-gigawatt data center is substantial by any measure. For context, that power envelope is roughly equivalent to the electricity consumption of a mid-sized city, and it places Zhipu's new facility among the largest AI-specific compute clusters disclosed in Asia to date. The decision to rely entirely on Chinese chips reflects both necessity and strategic positioning. U.S. export restrictions, tightened repeatedly since 2022, have effectively cut off access to Nvidia's H100 and A100 GPUs for many Chinese entities. Zhipu has responded by designing its infrastructure around alternatives from Huawei's Ascend series, along with smaller contributions from startups such as Biren and Moore Threads, according to people familiar with the deployment.

The performance gap between these domestic accelerators and cutting-edge U.S. hardware remains real. Huawei's Ascend 910B, the most capable chip widely available to Chinese labs, delivers roughly 60 to 70 percent of the training throughput of an H100 in mixed-precision workloads. Closing that gap at scale requires more silicon, more power, and tighter orchestration across thousands of nodes. Zhipu's engineering team has spent the past eighteen months optimizing distributed training frameworks to squeeze maximum efficiency from the hardware at hand, work that includes custom kernel libraries and revised communication protocols to reduce inter-chip latency.

GLM Models and the Economics of Homegrown Compute

Zhipu's GLM series has carved out a niche in the Chinese market by focusing on bilingual capabilities and relatively aggressive pricing for API access. The company positions its models as enterprise-ready alternatives to OpenAI's GPT family, with particular strength in code generation and long-context understanding. Bringing the 1-gigawatt facility online gives Zhipu the headroom to train larger parameter counts and run more extensive fine-tuning experiments without competing for scarce cloud GPU allocations.

The economics, however, are less forgiving than they would be with Nvidia silicon. Higher chip counts mean higher capital expenditure per unit of compute, and the power bill for a gigawatt facility running at even modest utilization can exceed tens of millions of dollars per quarter. Zhipu will need to demonstrate that the quality and differentiation of its models justify that cost structure, especially as rivals with access to legacy Nvidia inventory or smuggled chips continue to operate. The firm raised approximately $800 million in its most recent funding round, and a significant portion of that capital has flowed directly into data center buildout and power purchase agreements.

Investor Reaction and the Volatility of AI Infrastructure Bets

Tuesday's 37 percent surge was as much about relief as conviction. The stock had fallen sharply in the preceding sessions on concerns that construction delays and chip yield issues might push the facility's completion into the fourth quarter. Confirmation that the center is operational and beginning to onboard workloads removed a major overhang. Yet the volatility itself illustrates how tightly coupled equity performance has become to infrastructure delivery in the AI sector. Investors are pricing not just model performance or revenue growth, but the ability to secure and deploy compute at scale in a fragmented and politically contested supply chain.

Zhipu's valuation now embeds an expectation that homegrown chip ecosystems will continue to mature. If Huawei and others can close the performance gap further over the next twelve to eighteen months, the company's all-domestic strategy will look prescient. If the gap widens or power costs spiral, the capital intensity of the approach may weigh on margins and limit reinvestment in research. The Hong Kong listing gives Zhipu access to international capital, but it also exposes the firm to shareholder scrutiny that domestic peers trading in Shenzhen or Shanghai face less directly.

What This Means for China's AI Compute Landscape

Zhipu's 1-gigawatt facility is unlikely to remain an outlier for long. Several other large Chinese AI labs are pursuing similar strategies, driven by the same mix of necessity and nationalist industrial policy. The central government has signaled that self-sufficiency in AI infrastructure is a strategic priority, and subsidies for domestic chip purchases and favorable power rates for data centers are part of that push. At the same time, the technical challenges of running frontier training workloads on non-Nvidia hardware are forcing labs to innovate in areas such as model architecture, quantization, and distributed systems design.

For the broader AI industry, Zhipu's milestone offers a data point on how quickly alternative chip ecosystems can scale when capital and policy support align. The performance trade-offs are real, but they are not insurmountable, especially for workloads that can tolerate longer training windows or accept slightly lower throughput in exchange for supply-chain security. As export controls remain in place and geopolitical fragmentation deepens, the ability to build and operate large-scale AI infrastructure without relying on U.S. silicon will increasingly define competitive dynamics in the region. Zhipu's share price may have rebounded on news of a single data center, but the longer-term question is whether that facility can deliver the model quality and cost efficiency needed to sustain growth in a market that remains intensely competitive and capital-hungry.