Introduction
The Energy Conflict Defining the Age of Artificial Intelligence
The available electrical capacity in Hong Kong is estimated at 3.8 GW for the next three years. This value does not correspond to the current demand, but to the maximum threshold that the distribution system can support without localized collapses. The expansion of artificial intelligence data centers has tripled power demand in less than two years, exceeding historical levels. Projects such as Cyberport 3.0 and Sandy Ridge — which provide a total of 3,000 petaflops of computing power — require a closed-loop cooling network with an average power consumption equal to 28% of the overall consumption.
This data indicates that the paradigm of technological innovation has shifted from the chip to the electrical infrastructure. Latency is no longer determined by the distance between processor and memory, but by the time required to cool a single inference node. The euphoria assumed that the speed of transistors was the limiting factor; however, the data shows that the thermodynamic efficiency of the system is the real bottleneck.
The Invisible Mechanism: Dissipation as a Physical Limit
Every petaflop of computation requires a continuous flow of energy. According to a study by the Lawrence Berkeley National Laboratory, training generative models consumes 2 to 3 times more energy than traditional cloud workloads. A single model like Llama-3-70B requires approximately 18 terawatt-hours to be trained—an amount equivalent to the annual consumption of a small country. In Hong Kong, where data centers already operate at levels exceeding the global average by 43% in terms of carbon emissions, this figure represents a systemic pressure that is no longer sustainable.
The local power grid is designed for an average seasonal load. The arrival of synthetic systems has introduced hourly consumption peaks that exceed 140% of the average. The load management system can no longer rely on passive balancing; it requires active prediction based on climate and demand models, with real-time adjustments through distributed microgrids.
In practice, the capacity of a data center is no longer measured in petaflops or FLOPS, but in how much power it can dissipate without compromising the system. Cooling—often managed by liquid helium systems—becomes a critical variable: every liter of helium consumed equates to 150 kWh of electrical energy that is not available for inference.
Expectations Between System and Reality
The Financial Secretary Paul Chan Mo-po stated that Hong Kong has become a “ground for strategic adaptation” for Chinese technology companies. This vision is based on a narrative of immediate access to the global market, but it ignores the energy constraint. As highlighted by Ying Xu and Weishi Zhang in the South China Morning Post report: “Hong Kong must leverage the rest of the Greater Bay Area and manage competing energy demands between AI and public needs”.
According to researchers at the University of Hong Kong, the local power system is not capable of supporting a 50% increase in demand from data centers without structural interventions. The current electricity grid has already reached its operational limit in the most densely populated areas.
The market expects rapid expansion; physical reality requires long-term planning. The paradox is that the same companies that present themselves as leaders in innovation are those most vulnerable to a collapse of the distribution system. Strategic decisions cannot be made based on technological KPIs, but on data related to thermodynamic efficiency and availability of primary inputs.
The Transition No One Wants to See
Hong Kong’s stated goal — the transition to electric vehicles by 2035 — becomes incompatible with the expansion of data centers unless a radical change is introduced in the energy source. The current plan foresees a 17% increase in installed capacity for data centers over the next five years, while the distribution network increases by only 8%. This discrepancy indicates that the euphoria assumed an infinitely adaptable system; however, the data shows that a physical threshold has been reached.
If intervention is not carried out by 2027, the city may experience scheduled blackouts during peak inference hours. The estimated cost to repair a critical node in the system is $43 million USD — an amount higher than the annual budget of local AI startups.
The data that measures the deviation from the status quo is the reduction in the average operating capacity of data centers, calculated over the last three seasons: -22 days of operational autonomy before energy replenishment. This parameter represents the actual resilience of the system.
Monitoring the Dissipation Margin
If you are considering an AI infrastructure investment in Hong Kong, the key data point to monitor is the effective thermal dissipation capacity per node. A value below 65% of the nominal power indicates that the infrastructure cannot sustain continuous loads.