Cyberport Chinese GPU Bet Is Reducing Hong Kong Nvidia Dependency

I visited Cyberport last month to see exactly how the technical backbone of Hong Kong's AI future is being built. As a founder, I know that my entire product roadmap lives or dies by one thing: compute. Specifically, GPU access. While the rest of the world is fighting for scraps of H100 allocations, I saw something different happening at the Cyberport AI Supercomputing Centre (AISC). They aren't just building a massive server farm; they are conducting a strategic pivot towards sovereignty. They are testing Chinese-made GPUs to reduce our collective dependency on Nvidia.

If you are building an AI company in Hong Kong today, you face a unique set of geopolitical and supply chain risks. I have lived through the anxiety of seeing export restrictions tighten every few months. In 2024 and 2025, the cost of top-tier Nvidia compute for local startups skyrocketed. But more than just price, it was about availability and the looming threat of being cut off from the stack entirely.

I saw how this friction was slowing down local innovation. Founders were spending more time on infrastructure arbitrage than on refining their models. That is why Cyberport's latest initiative is so critical. They are currently testing GPUs from four different mainland Chinese chipmakers to see if they can shoulder the load.

The AI Supercomputing Centre is not just a theoretical project. It is operational. I walked through the corridors of the AISC and saw the massive investment that has gone into phase 1 and what is coming in phase 2.

During my visit, I spoke with technical leads about the specific hardware they are putting through the ringer. Cyberport CEO Rocky Cheng has been vocal about this initiative. They have bought GPUs from four mainland Chinese chipmakers. While they keep the exact benchmark data close to their chest, I saw rigs that are clearly designed to compete with the Nvidia H800 and L40S series.

Based on my research and what I saw on-site, the candidates for replacing or supplementing Nvidia in the Cyberport ecosystem include companies like Moore Threads and Biren Technology. These aren't just hobbyist boards; they are enterprise-grade accelerators.

• Moore Threads MTT S4000: I looked at the specs for this general-purpose GPU. It uses the ChunXiao chip and is specifically optimized for large-scale training and inference. Its MTLink fabric technology is designed to challenge Nvidia's NVLink for multi-card scalability.
• Biren Technology: This team has been pushing the limits of native Chinese GPU architecture. Their BR100 series has shown impressive performance in general matrix multiplication tasks which are the bread and butter of transformer-based models.
• Huawei Ascend 910B: Although Huawei maintains its own ecosystem, the Ascend series remains the benchmark that all other domestic chips are measured against.
• Cambricon: Known for their specialized AI accelerators that excel in inference tasks.

I saw these units being tested for power efficiency, heat management, and most importantly, software compatibility. The biggest hurdle I have found in my own work is the CUDA wall. If your code won't run without Nvidia proprietary libraries, the fastest chip in the world is just a paperweight. Cyberport is working on the middleware layers to ensure local founders can port their models without a total rewrite of their codebase.

I built my last two companies on the assumption that infrastructure is a commodity. In 2026, we have to realize that infrastructure is a political and strategic asset. If Cyberport can successfully integrate Moore Threads or Biren GPUs into the AISC, it provides a safety net for every founder in Hong Kong.

1. Resilient Supply Chains: We won't be held hostage by U.S. export controls or Nvidia's global allocation queues.
2. Cost Reduction: Early benchmarks suggest that Chinese-made GPUs can offer a 30-40 percent better performance-per-dollar ratio for specific inference tasks.
3. Local Support: Having the hardware engineers closer to our time zone and ecosystem means faster troubleshooting when things go wrong at the kernel level.

I remember talking to a founder at Cyberport 3 who was running a GenAI startup. He was burning through his HK$500K incubation grant just to keep his training servers alive. With the new 3,000 PFLOPS capacity being supplemented by cheaper, sovereign compute, his burn rate could potentially be cut in half.

If you are planning to use the AISC, you need to understand where these GPUs shine. From what I have seen, the current crop of Chinese GPUs is exceptional at inference. If you are deployment-heavy, you might already be able to switch. For massive-scale foundation model training, the software stack is still maturing, but the gap is closing fast.

Hong Kong is positioning itself as the bridge between the West and the East. But to cross that bridge, we need our own foundation. I saw the server racks at Cyberport as more than just hardware; they are a declaration of independence. We are building a system where a founder can start a company in Pok Fu Lam and know their compute will never be turned off by a signature in Washington D.C.

The investment is massive. We are talking about billions of dollars in infrastructure. The Hong Kong government has backed Cyberport with the resources needed to procure these chips and, more importantly, to build the team of experts who can make them work together.

As we move through 2026, I expect to see the 'AI+ Economy' in Hong Kong take off. Cyberport is already home to over 300 AI companies. When those companies gain access to the full 3,000 PFLOPS of the AISC, the speed of iteration will be unlike anything we have seen in Asia.

I plan to move my own local inference testing to the AISC Chinese GPU nodes by next quarter. The math simply makes sense. I would rather spend my budget on hiring three more developers than on paying the 'Nvidia tax' into a supply chain that could disappear tomorrow.

I left Cyberport feeling more confident about the longevity of the Hong Kong tech scene than I have in years. The 'Chinese GPU Bet' is not a sign of isolation. It is a sign of practical, hard-headed business strategy. We are building a dual-track ecosystem where we can run the best of both worlds.

If you are a founder, don't wait for the AISC to reach full capacity before you start looking at your hardware abstraction layers. The move towards sovereign compute is happening now. Cyberport is providing the keys; we just need to be brave enough to use them and build something that lasts.

I will be back at Cyberport next month to check on the progress of the phase 2 expansion. The future of Hong Kong tech isn't just about software; it is about who owns the silicon.