Cyberport, Hong Kong’s government-backed tech incubator, is considering adding Chinese-made graphics processing units (GPUs) to its AI Supercomputing Centre to reduce reliance on Nvidia amid ongoing China-US tensions.

The centre currently uses Nvidia H800 chips to supply 1,300 petaflops of computing power and plans to expand by another 1,700 petaflops by year-end, reaching a total of 3,000 petaflops.

Cyberport has tested GPUs from four mainland Chinese chipmakers at its AI lab and will consider performance and cost before making a decision.

Cyberport is building new facilities with a 60,000 sq ft data center, dedicating two-thirds to AI computing.

.source-ref{font-size:0.85em;color:#666;display:block;margin-top:1em;}

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Geopolitical tensions force strategic diversification in critical computing infrastructure

Cyberport’s move reflects a broader trend where tech hubs must diversify suppliers due to US-China tensions, even when existing solutions work well. The center currently achieves 98% capacity utilization with Nvidia’s H800 chips across 1,300 petaflops of computing power, yet still seeks alternatives due to supply chain risks rather than performance issues1. This parallels Nvidia’s own strategy of creating region-specific chips—the company recently proposed a B30 chip delivering 80% of standard Blackwell performance specifically for the Chinese market2. Hong Kong’s position as a bridge between Chinese and international markets makes this diversification particularly critical, as the city aims to expand computing capacity to 15,000 petaflops by 20303. The strategic shift demonstrates how geopolitical considerations now override pure technical merit in infrastructure decisions, forcing even successful facilities to hedge their technology dependencies.

Chinese GPU alternatives demonstrate competitive performance in specialized applications

Recent breakthroughs show Chinese GPUs can outperform Nvidia chips in specific use cases, with researchers achieving notable performance improvements in supercomputer simulations4. The performance gains were achieved using Hygon processors and domestic GPUs for flood defense planning simulations, demonstrating real-world applications rather than theoretical benchmarks4. Multiple Chinese GPU companies, including Biren Technology and Moore Threads, showcased competitive products at Shanghai’s World Artificial Intelligence Conference, indicating a maturing domestic ecosystem5. Cyberport’s testing of four different Chinese GPU suppliers suggests these alternatives now offer “similar performance” to established options, with cost becoming the primary differentiator rather than capability1. ……