Cyberport AI Supercomputing Centre Is 3000 PFLOPS of Sovereign Compute Power

The industrial hum of three thousand petaflops echoing across the Cyberport waterfront is the most visceral evidence yet that Hong Kong has finally stopped talking about digital sovereignty and started building it. I remember standing in a coworking space in Cyberport 3 back in 2018, looking out at the South China Sea and thinking how the entire startup ecosystem here was effectively a tenant on someone else's land - not just in terms of real estate, but in terms of the intellectual and physical infrastructure we used to build our code. We were tenants of AWS in Oregon, tenants of Azure in Singapore, and tenants of Google in Tokyo. In 2026, that period of structural dependency is officially drawing to a close.

The Artificial Intelligence Supercomputing Centre (AISC) is more than just a massive server room filled with liquid-cooled racks. It is a 3,000 PFLOPS (Petaflops) statement of intent. For my fellow non-technical founders, a petaflop represents one quadrillion (10^15) floating-point operations per second. When you scale that to 3,000, you are looking at a computational engine capable of training the next generation of frontier LLMs entirely within our own city limits. This is the "Sovereign Stack" in physical form, and it is the single most important development for the Hong Kong tech ecosystem since the founding of the Innovation and Technology Bureau. It is about reclaiming the means of production in the digital age.

Let's look at the numbers. The first phase of the AISC, which went live in late 2024, provided around 1,300 petaflops. By the middle of 2025, that capacity has scaled to the target of 3,000 petaflops. This isn't just about raw power; it's about the security and continuity that power provides. We live in an era where the export of H100s and other high-end GPUs is governed by the shifting winds of Washington policy. As a Hong Kong founder, relying on a US-controlled cloud provider is no longer just a technical choice - it is a geopolitical risk. Every developer I speak with in the Cyberport ecosystem has felt the sting of uncertainty. One day a model is available, the next it is geofenced. One week a chip is shipable, the next it is on a restricted list. This volatility is the enemy of enterprise-grade software.

When you use the AISC, your data never leaves the 852. Your training weights are stored on localized storage, and your inference happens on silicon that is physically located in Pok Fu Lam. This is why the "Sovereign Stack" matters. If a policy change tomorrow freezes global API access to a particular American model, the firms running local instances at Cyberport will be the only ones left standing in the Hong Kong market. This localized resilience is our competitive edge in a fragmented world. We are not just building for today; we are building for a future where digital borders are as real as physical ones.

One of the most strategic decisions made by Cyberport CEO Rocky Cheng and his team was to avoid a mono-culture of silicon. For years, the industry was locked into a "Nvidia or nothing" mindset. But the AISC was designed to be hardware-agnostic. During the early ramp-up phases, the facility began testing accelerators from four different mainland Chinese chipmakers alongside traditional international options. This diversity is crucial. In the current global climate, being tied to a single vendor is a single point of failure.

This is a brilliant hedge. By building a software layer that can distribute workloads across different hardware architectures, Cyberport is ensuring that Hong Kong's AI capability is not tethered to the survival of a single supply chain. This is the kind of engineering foresight that defines a "Sovereign Stack." We aren't just buying chips; we are building a system that can consume any chip that meets our performance standards. Whether it is Blackwell, Ascend, or an emerging local accelerator, the AISC is the bucket that can hold them all. This flexibility allows us to optimize for cost, performance, and legal compliance simultaneously. It is the architectural equivalent of a multi-currency fund, protecting us against the devaluation of any single technological "currency."

In the early days of any AI startup, the "GPU Tax" is often what kills the business. I have seen countless promising Hong Kong teams run out of runway because their cloud bills for R&D were hitting six figures USD before they even had a Beta product. The Hong Kong government recognized this as a systemic bottleneck and responded with the Artificial Intelligence Subsidy Scheme (AISS). This wasn't just a political gesture; it was a targeted strike at the highest barrier to entry in the modern tech landscape.

The AISS is backed by a HK$3 billion allocation over three years. The mechanics are simple and devastatingly effective for local founders - - Standard Subsidy: Eligible startups and research bodies can have up to 70% of their compute costs covered by the government. This is a massive reduction in burn rate. - Exceptional Cases: For projects deemed of high strategic value to the SAR - such as those in healthcare, finance, or smart city logistics - that subsidy can climb as high as 90%. - Direct Credit: The funds are often applied directly to the compute usage, meaning you don't have to wait months for a reimbursement check that may never come. You get the benefit in real-time.

Imagine the competitive advantage this gives a Hong Kong-based AI agent firm. If your competitor in Singapore or London is paying 100% of their AWS bill, and you are paying 10%–30% of your AISC bill for equivalent (or higher) compute power, your runway is effectively 3x to 5x longer. In the world of tech, time is the only resource that truly matters. That extra year of runway is the difference between finding product-market fit and joining the startup graveyard. It allows us to iterate faster, fail cheaper, and eventually dominate the regional market.

For developers, the move to AISC involves a mental shift from "serverless" convenience to "bare metal" optimization. You aren't just hitting an endpoint; you are managing a node in a high-performance cluster. This requires a deeper understanding of orchestration. You have to think about data throughput, GPU memory management, and interconnect speeds. It's more demanding than calling a GPT-4 API, but it's also far more rewarding. You have absolute control over the environment.

Here is a conceptual look at how we architect an agentic failover system that prioritizes local AISC compute while maintaining a backup on global cloud providers - what I call the "Hybrid Sovereign Architecture." This ensures 100% uptime even in a worst-case geopolitical scenario.

import os
import requests
import json
import logging

# Set up logging for our sovereign operations
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("SovereignNode")

class SovereignAgent:
    def __init__(self):
        # Local Cyberport AISC Endpoint - The heart of the stack
        self.primary_endpoint = "https://aisc.cyberport.hk/v1/inference"
        # Global Backup (e.g., AWS/OpenAI) - Use only when necessary
        self.backup_endpoint = "https://api.openai.com/v1/chat/completions"
        self.auth_key = os.getenv("AISC_SECRET_KEY")
        self.backup_key = os.getenv("OPENAI_API_KEY")

    def run_inference(self, prompt, context=None):
        logger.info(f"Initiating request on AISC Sovereign Stack...")
        try:
            # Try the Sovereign Stack first to ensure data residency and low latency
            response = requests.post(
                self.primary_endpoint,
                headers={
                    "Authorization": f"Bearer {self.auth_key}",
                    "X-Sovereignty-Strict": "True" # Force HK-local execution
                },
                json={
                    "model": "deepseek-v3-local", 
                    "prompt": prompt,
                    "max_tokens": 1024,
                    "temperature": 0.7
                },
                timeout=8 # Strict timeout for local low-latency requirements
            )
            response.raise_for_status()
            logger.info("AISC Local Compute Success.")
            return response.json()["text"]
            
        except requests.exceptions.Timeout:
            logger.warning("AISC local node timeout. Checking network and load...")
            return self.fallback_to_global(prompt)
        except Exception as e:
            logger.error(f"Sovereign Node Error: {e}")
            # Fallback to global infrastructure if local metal is undergoing maintenance or restricted
            return self.fallback_to_global(prompt)

    def fallback_to_global(self, prompt):
        logger.info("Switching to Global Backup Infrastructure...")
        try:
            # Fallback logic for global provider - ensures continuity
            response = requests.post(
                self.backup_endpoint,
                headers={"Authorization": f"Bearer {self.backup_key}"},
                json={"model": "gpt-4o", "messages": [{"role": "user", "content": prompt}]},
                timeout=20
            )
            response.raise_for_status()
            return response.json()["choices"][0]["message"]["content"]
        except Exception as e:
            logger.critical(f"Total System Failure: Both local and global endpoints unreachable. Error: {e}")
            return "ERROR: Service Interruption"

We cannot talk about Cyberport without talking about the Greater Bay Area (GBA). The AISC is not just a Hong Kong asset; it is the data lighthouse for the entire GBA. With 15,000 petaflops as the eventual long-term target, Hong Kong is positioning itself to be the AI training ground for the Pearl River Delta. This region, encompassing 11 cities including Shenzhen and Guangzhou, is the manufacturing heart of the world. By housing the brain (the AISC) of these operations in Hong Kong, we are creating an unbreakable bond between manufacturing scale and computational intelligence.

The integration of data flows between Hong Kong and the mainland (under the GBA cross-boundary data flow agreement) means that a company in Shenzhen can use the AISC to process data while staying within a high-standard regulatory framework that is recognized globally. This is the "arbitrage" that makes Hong Kong unique. We have the common law legal system, we have the global connectivity, and now we have the physical compute to back it up. We are the trusted bridge where Western data standards meet Eastern industrial power.

1. Finance (the Central Segment): Banks are traditionally the most risk-averse institutions. The ability to run LLMs on-premise at a government-backed facility like Cyberport is the only way many of these firms will ever move beyond "experimental" AI into full production. For a bank in Central, data security isn't a feature; it is the product. The AISC provides a "Private Cloud" experience with "Public Cloud" scale. This is the holy grail for fintech founders.
2. Logistics: Hong Kong's terminal operators are using AI to optimize container flows. For these critical operations, a 10ms latency difference between a local AISC node and a Singapore-based AWS region is the difference between a smooth operation and a pile-up. When you are managing millions of TEUs (Twenty-foot Equivalent Units), speed is safety.
3. Medical Research: With our world-class medical schools at HKU and CUHK, the AISC provides the heavy-duty compute needed for genomic analysis and drug discovery - data that is too sensitive to ever leave the territory. We are seeing breakthroughs in personalized medicine that would have been impossible without the local compute capacity to process massive bio-data sets in real-time.
4. Education and EdTech: Local universities are now integrating AISC access into their computer science curricula. We are training a generation of engineers who aren't just "prompt engineers" but "infrastructure engineers" who understand how to milk every ounce of performance out of a cluster.

There is a common misconception that local "sovereign" compute is somehow slower or less advanced than the "magical" hyperscalers. The reality is that an H100 in Cyberport performs exactly the same as an H100 in a Google data center in Iowa. The laws of physics do not respect national borders. The difference is the speed of the fiber and the proximity of the user.

For a local enterprise, the network round-trip time to Cyberport via the HKIX (Hong Kong Internet eXchange) is measured in single-digit milliseconds. When you are running a complex multi-agent system where a single user prompt triggers 20 different AI-to-AI calls, that latency adds up exponentially. 10ms vs 150ms over 20 calls is the difference between a 2-second response and a 5-second response. In user experience terms, that is the difference between a tool that feels like a natural extension of your mind and a tool that feels like a chore. Furthermore, by optimizing our models for the specific interconnect architecture of the AISC (using techniques like RoCE v2 or InfiniBand), we can often achieve higher training efficiency than on generic public cloud instances.

I've followed the recent remarks from Cyberport’s leadership closely. The term they use, "New Quality Productive Forces," is more than just a buzzword. It refers to a shift away from traditional, labor-heavy growth toward growth driven by high-tech, high-efficiency, and high-quality innovation. It's about moving up the value chain. Hong Kong used to be a port for physical goods; now we are a port for digital intelligence.

The AISC is the engine of these new forces. By providing "3,000 PFLOPS of sovereign compute power," Cyberport is effectively handing a high-powered drill to a workforce that has been digging with spoons. The question for us as founders is no longer "Where is the infrastructure?" but "How fast can we build?" We are seeing a new class of "AISC-native" startups emerging. These are companies that didn't start on AWS; they started on the sovereign stack. They are building architectures that are natively optimized for the local ecosystem, giving them a structural advantage over "global" firms trying to port their solutions into the city.

If you are a founder and you are still paying full price for US-based GPU instances, you are committing a double sin - you are wasting capital and you are building on a fault line. You are exposing your business to risks that your competitors are already hedging against. Here is my roadmap for the "Sovereign Shift" -

1. Get Certified: If you aren't a Cyberport incubatee or tenant, look at the "AISS Eligible User" criteria immediately. You don't necessarily have to move your office to Pok Fu Lam to use the power, but you do need to be part of the official ecosystem.
2. Audit Your Stack: Identify every "kill switch" in your architecture. Does your app break if a US API goes dark? Does it break if a Singapore data center has a routing error? If yes, start building the failover logic to a local instance at AISC.
3. Test Local Models: Stop assuming that closed-source US models are the only option. Models like DeepSeek-V3 or Qwen-2.5, when fine-tuned on local compute at AISC, are often superior for Hong Kong-specific tasks - ranging from Cantonese instruction following to GBA-compliant financial reporting.
4. use the 70%: Use the subsidy to run experiments you previously thought were too expensive. Want to train a specialized model on your industry's niche data? Want to run a massive simulation of the city's traffic flow? Now you can do it for 30 cents on the dollar. This is the time to be bold.

The industrial hum of the AISC is the sound of a city maturing. We are moving from being a city of "traders" of technology to being a city of "makers" of it. When 3,000 PFLOPS are buzzing in the background, the excuse that "Hong Kong is too small" or "we don't have the resources" evaporates. We have more compute power per square kilometer than almost anywhere else on earth.

The Sovereign Stack is here. The power is physical. The money is on the table. The only missing ingredient is your code. We are entering an era where intelligence is a manufactured commodity, and the AISC is our primary factory. As a founder, your job is to design the products that this factory will produce. Whether it's agentic workflows for global trade or real-time translation for GBA tourism, the tools are ready.

Secretary Sun Dong has already hinted at the next phase. 3,000 PFLOPS is just the base camp. The climb to 15,000 petaflops will put Hong Kong on par with the most advanced specialized AI regions in the world. As we look toward the 2026-2027 fiscal year, the goal is to make compute so abundant in Hong Kong that it becomes a utility, like water or electricity. This "compute abundance" will trigger a Cambrian explosion of innovation. When the cost of intelligence drops toward zero, we can afford to apply AI to everything - from optimizing the garbage collection in Sham Shui Po to predicting the next major financial trend in Central.

When compute is a utility, the competitive advantage of a city is determined by its talent and its regulatory environment. On both fronts, Hong Kong is winning. We have the best developers in the region and a government that is actually putting its weight behind the infrastructure of the future. We are building a "Digital Hong Kong" that is resilient, sovereign, and incredibly powerful.

We are no longer tenants. With the AISC reaching 3,000 PFLOPS, we are finally the masters of our own digital destiny. The industrial-scale AI supercomputing center at Cyberport is the foundation upon which we will build the next decade of Hong Kong tech. It is our fortress in a world of digital storms.

The period of structural dependency on foreign infrastructure was a necessary phase - it allowed us to grow quickly when we didn't have our own resources. But it was just that - a phase. As we look at the waterfront of Cyberport today, we aren't just seeing a beautiful sunset over the South China Sea. We are seeing the birthplace of the Sovereign Stack, where the hum of 3,000 petaflops is the heartbeat of a new, independent, and powerful Hong Kong tech ecosystem.

It’s time to stop renting and start building. The metal is waiting for your models. The chips are cooled, the power is on, and the subsidy is ready. The only question left is: What will you build with the sovereign power of 3,000 petaflops?

The term "Sovereign" often gets misinterpreted as "isolationist." Nothing could be further from the truth. A Sovereign Stack is about resilience. It's about having the option to operate independently if the global systems we normally rely on fail us. In the volatile world of 2026, resilience is the most valuable asset a founder can have. It is the insurance policy for your business.

By using the AISC, you aren't just saving money; you are securing your company’s future. You are ensuring that your intellectual property, your customer data, and your operational capability are all protected within the jurisdiction of Hong Kong. That is not just good business - it’s a strategic necessity. It is how we ensure that the next "Unicorn" from this city is built on a foundation that no foreign government can take away.

The hum of those 3,000 petaflops is the sound of opportunity. I’ll see you at the waterfront, ready to code the future of the 852. We are just getting started. This is not just a building; it is a movement. A movement toward a self-sufficient, high-tech Hong Kong that serves as the literal brain of the Greater Bay Area. We are the architects of this new reality, and the foundation is solid metal and liquid cooling. Let's get to work.

To truly appreciate the scale of 3,000 petaflops, one must understand the evolution of supercomputing in the region. For decades, Hong Kong's involvement in high-performance computing was limited to theoretical academic research. We had small clusters at HKUST and HKU, but they were never intended for the industrial-scale demands of a trillion-parameter LLM. The AISC changes that narrative permanently. We have moved from the "playground" phase to the "industrial" phase of AI.

This transition is mirrored in the way local startups are hiring. We are no longer just looking for "full stack developers" who can glue together third-party APIs. We are looking for "AI infrastructure engineers" who can optimize CUDA kernels, manage Distributed Training sessions, and understand the nuances of NVLink vs. PCIe Gen5 bandwidth. The AISC is creating a high-skills job market that didn't exist in Hong Kong three years ago. This talent density is what will ultimately keep the 852 competitive on a global scale.

Furthermore, the environmental sustainability of the AISC is a point of pride. Located in the Pok Fu Lam coastal area, the facility uses advanced cooling technologies to maintain an incredibly low PUE (Power Usage Effectiveness). In an industry often criticized for its carbon footprint, the AISC is proving that sovereign compute can also be green compute. This alignment with global ESG standards makes it even more attractive for institutional investors who want to back the next generation of "Ethical AI" from Hong Kong.