[Eyes on China] The AI boom is a building physics problem

What struck me while examining China’s AI infrastructure is that this is no longer simply a computing story. It is a building physics and territorial planning story. By 2025, AI accounted for approximately 79% of China’s computing capacity. General CPU-based computing has nearly plateaued, while GPU-based AI capacity is expanding rapidly. Before 2022, China added roughly 150 EFLOPS per year. It is now adding closer to 600 EFLOPS annually. In one year, the system is adding what previously took about four years. This changes the building.

GPUs concentrate far more electricity and heat inside each rack. Conventional air-cooled racks typically operate at up to approximately 30 kW. AI racks can exceed 100 kW, making direct-to-chip liquid cooling increasingly necessary.

China’s data centers could consume 774 TWh by 2030, around 6% of national electricity demand. At a PUE of 1.25, approximately 155 TWh would be consumed by non-IT systems such as cooling, pumps, fans, and electrical infrastructure. But nearly every kilowatt-hour used by the IT equipment also ultimately becomes heat. We are therefore building enormous, continuously operating heat-producing buildings, then spending more energy to reject that heat where nobody needs it.

China’s east-west computing strategy tries to address the energy geography. The east contains the users, technology companies, and digital demand. The west offers land, renewable electricity, and, in some regions, cooler conditions. Energy-intensive AI training and batch workloads can move west because they tolerate delay. Real-time inference must remain near eastern users. Yet by June 2026, 55.9% of intelligent computing capacity was still located in the east, compared with 32.6% in the west. Computing does not simply follow cheap electricity. It follows users, data, fiber networks, technical expertise, and latency.

And moving data centers west creates another contradiction: better access to renewable electricity, but fewer opportunities to reuse heat in cities, industries, greenhouses, or district heating networks. China may solve part of AI’s electricity geography while worsening its thermal geography. The real question is not simply where to place the servers. It is which workloads should run where and when, using which electricity, cooling technology, water source, and heat-recovery system.

AI may look like software. At a national scale, it is buildings, energy infrastructure, heat, electricity, and water.

#EyesOnChina #ArtificialIntelligence #DataCenters #BuildingPhysics #Architecture #GPU #LiquidCooling #WasteHeat #EnergyTransition