Zhipu AI Builds 1GW Data Center on Chinese Chips as Stocks Surge 40%

Zhipu AI’s Hong Kong-listed shares surged more than 40% on July 21 after the company completed construction of a 1GW-class AI data center built entirely on domestically produced chips, with no Nvidia hardware in the stack.

The facility will train the company’s GLM family of large language models.

The move is China’s most visible effort yet to prove that frontier AI training can run at scale without American semiconductors, and it marks a turning point in the global race over Chinese Chips.

Inside the 1GW Zhipu AI Data Center Built on Chinese Chips

Zhipu AI data center construction was confirmed by TechNode on July 21, citing reporting from Chinese tech outlets. A gigawatt of installed capacity is a significant threshold.

For comparison, the largest hyperscale data centers run by Google and Microsoft range from 100 megawatts to several hundred megawatts of IT load. A 1GW facility is approximately the size of a small city’s power draw, dedicated entirely to compute.

The facility is not merely large.

Its architecture matters more than its scale. Every chip in the rack is domestically produced, which means Zhipu is training models on domestic accelerators without Nvidia’s H100 or H200 GPUs, the chips that power the majority of frontier AI development globally.

U.S. export controls introduced in 2022 and tightened in 2023 effectively cut off China’s leading AI labs from Nvidia’s most powerful accelerators. Zhipu’s data center is a direct answer to that constraint.

The company also acquired compiler startup Zhongke Jiahe and is separately exploring custom AI chip development.

A compiler sits between software written by AI researchers and the physical hardware that executes it. By owning the compiler, Zhipu gains fine-grained control over how its models run on homegrown hardware, compensating for hardware performance gaps through software optimization.

Why Chinese Chips Matter for the GLM Model Stack

The GLM family, short for General Language Model, is Zhipu’s flagship series of foundation models.

Building and updating foundation models requires repeated training runs across thousands of accelerators.

Each training run can cost tens of millions of dollars and weeks of compute time.

If the domestically produced chips powering that process are slower or less energy-efficient than Nvidia’s offerings, the competitive disadvantage compounds with every model generation.

Zhipu’s bet is that software efficiency, purpose-built infrastructure, and scale can close the gap. The compiler acquisition is the clearest signal of that strategy.

Rather than waiting for these chips to match Nvidia through hardware iteration alone, Zhipu is attacking the problem from both ends simultaneously.

The 40% share surge reflects investor belief that the strategy is credible. Knowledge Atlas Technology (2513.HK), the Hong Kong-listed entity that operates the Zhipu AI business, closed the morning session on July 21 up roughly 25% before extending gains. Chinese tech stocks broadly rallied on state-backed buying, but Zhipu outpaced the index by a wide margin.

From Domestic Policy Push to Market Reality

China’s government has prioritized semiconductor self-sufficiency since the first rounds of U.S. export controls.

The domestic chip ecosystem has matured considerably since 2022. Companies such as Huawei and Cambricon have shipped AI accelerators that, while not matching Nvidia’s peak performance per chip, can be deployed in large numbers inside China’s borders without export-control risk.

The data center buildout follows broader policy moves.

Earlier this month, Chinese authorities were reported to be considering tightening export controls on AI models and chips, a potential mirror of U.S. restrictions that would limit outbound flows of Chinese AI technology. That regulatory backdrop gives companies like Zhipu a dual motive for domestic infrastructure built on homegrown accelerators: supply security and potential compliance with future outbound rules.

Building at 1GW is also a statement of permanence.

A facility at that scale takes years to design, permit, and construct. Zhipu is not hedging.

The company is committing capital and engineering resources to a future in which it never relies on imported accelerators again.

Also Read: Zero-Knowledge Proofs Emerge as Crypto’s Defense Against the AI Trust Crisis

What a Nvidia-Free AI Stack Built on Chinese Chips Actually Looks Like

Running a large language model at scale on non-Nvidia hardware requires solving three interconnected problems. First, raw compute throughput: how many floating-point operations per second the chips can deliver across a training job.

Second, interconnect bandwidth: how fast chips can exchange data with each other, since training a large model distributes work across thousands of accelerators simultaneously. Third, software compatibility: whether the model training frameworks researchers use, primarily PyTorch, can compile and run efficiently on non-Nvidia hardware.

Nvidia’s competitive moat is not just the chip.

It is the CUDA software ecosystem, built over two decades, that makes Nvidia GPUs dramatically easier to program than alternatives. Zhipu’s acquisition of a compiler specialist is a direct attack on that advantage.

A high-quality compiler can translate PyTorch operations into efficient instructions for these domestic accelerators without requiring researchers to rewrite code from scratch.

The approach is not unique. Google built its own tensor processing units specifically to run its AI workloads without depending on Nvidia.

AMD has invested heavily in its ROCm software stack. What makes Zhipu’s move notable is that it is happening at 1GW scale, inside China, with explicitly Chinese Chips, under a geopolitical context in which the alternative of simply buying Nvidia hardware is not available.

The 40% Move and What Comes Next

The stock surge is a signal, not a guarantee.

Zhipu has demonstrated that it can build at scale on domestically produced chips. Training competitive frontier models is a separate question.

The GLM series will need to perform on standard benchmarks for the market’s confidence in the infrastructure story to hold.

The compiler acquisition and custom chip exploration suggest Zhipu is planning for a future measured in years, not quarters. If domestic chip performance continues to improve and software optimization closes the remaining gap, China’s AI labs could compete at the frontier without any dependency on U.S. supply chains.

That outcome would have consequences far beyond Zhipu’s share price.

Read Next: NEAR Protocol Makes a Breakthrough With Quantum-Safe Mainnet Signing