SenseTime and Hong Kong Plan 40,000-Petaflop AI Computing Hub by 2030

China’s most ambitious domestic AI infrastructure push yet is being anchored not in Beijing or Shenzhen, but in Hong Kong.

AI giant SenseTime and the Hong Kong Science and Technology Park (HKSTP) have announced a partnership to build a home-grown AI data centre in three phases, targeting 40,000 petaflops of computing power by 2030—with phase one expected to be operational by the end of 2026.

For global technology investors and AI infrastructure professionals, this project is a clear indicator of how China is engineering its way around one of the most consequential geopolitical chokepoints of the decade: access to advanced AI compute.

A Critical Partnership for China’s AI Independence

SenseTime, best known for its computer vision and large-model platforms, brings deep AI engineering expertise to the partnership, while HKSTP provides a uniquely advantageous location: Hong Kong’s special administrative status affords it regulatory and financial frameworks distinct from the mainland, making it a viable hub for attracting international capital and talent even as geopolitical pressures mount elsewhere.

The partnership is emblematic of a broader Chinese industrial strategy to vertically integrate the AI supply chain—from chip design and data centre infrastructure all the way to model training and deployment. Chinese tech majors including Baidu Cloud and Alibaba, which has outlined its own proprietary chip roadmap through the Yitian 710 series, are pursuing parallel tracks, underscoring that this is a sector-wide imperative, not a single company’s bet.

Three-Phase Build-Out Targets 40,000 Petaflops by 2030

The project is structured in three phases, with the first tranche of capacity expected online by the close of 2026. While exact per-phase capacity splits have not been publicly disclosed, the 40,000-petaflop end-state puts the facility in the same conversation as hyperscaler-grade deployments operated by global cloud providers. That figure represents an enormous volume of floating-point operations per second—sufficient to train and fine-tune frontier large language models and multimodal AI systems at scale.

The phased approach is strategically deliberate: it allows SenseTime and HKSTP to validate hardware configurations, manage capital expenditure, and adapt to the rapidly shifting landscape of available domestic and allied-nation accelerator chips—a landscape that changes with every new round of US export control adjustments.

What This Means for China’s AI Race

Compute availability is now arguably the single most critical bottleneck in frontier AI development. Training a state-of-the-art large language model requires clusters of thousands of high-end GPUs running for weeks or months. Any nation or company that cannot independently access sufficient compute is, by definition, constrained in its ability to develop leading AI systems.

Hong Kong’s role in this equation is more than symbolic. Its position as an international financial centre means that infrastructure projects here can attract foreign investment and operate under common-law legal frameworks, providing a degree of commercial confidence that pure mainland projects may not offer to international partners. This project signals that Hong Kong intends to leverage that status as a node in China’s domestic AI build-out—not merely as a financial hub, but as a genuine technology infrastructure anchor.

Broader Context: Compute as Geopolitical Leverage

The United States has progressively tightened export controls on advanced AI accelerators—most notably Nvidia’s H100 and successor chips—specifically targeting China. These restrictions have accelerated Chinese investment in domestic semiconductor alternatives, including Huawei’s Ascend 910B and 910C series, which are now being deployed at scale in mainland data centres despite performance gaps relative to leading US hardware.

China’s national “East Data, West Computing” initiative—a large-scale programme that routes data processing tasks from China’s economically active eastern cities to lower-cost computing clusters built across western provinces such as Guizhou, Inner Mongolia, and Ningxia—demonstrates the scale at which Beijing is approaching compute infrastructure as a strategic resource. The programme spans eight national computing hub nodes and ten national data centre clusters, effectively creating a distributed national AI compute backbone. The HKSTP–SenseTime project complements this national grid by adding a high-specification, internationally connected node at China’s southern gateway. Should access to foreign chips remain constrained—a likely scenario given current US policy trajectories—projects like this one, designed from inception to run on domestically sourced or allied-nation hardware, will prove increasingly vital to China’s ability to remain competitive in frontier AI development.

Key Takeaways

• 40,000-petaflop target: The HKSTP–SenseTime data centre will rival global hyperscaler deployments when complete in 2030.
• Phase one by 2026: Initial capacity will prioritise large-scale AI model training, giving Chinese AI developers a near-term domestic compute option.
• Strategic geography: Hong Kong’s unique legal and financial status makes it an internationally credible node within China’s broader compute infrastructure strategy.
• Geopolitical driver: US export controls on advanced AI chips are directly accelerating Chinese investment in independent compute infrastructure at every level of the supply chain.