The REAL Cost of Data Centers: Infrastructure, Power, and Geopolitics | Part 1

My notes

Summary

Data centers are hitting a hard physical ceiling: the electricity grid cannot scale fast enough to meet AI demand, with global data center power consumption expected to rival India’s by 2030. The US, China, and Gulf states are locked in a geopolitical race where energy access, chip control, and geographic positioning determine who builds AI infrastructure - and China’s early misallocation toward training (not inference) data centers has created a structural disadvantage despite having superior energy capacity.

Key Insight

• Electricity is the binding constraint, not hardware. Every watt entering a data center converts entirely to heat. Cooling alone takes 40% of total power. Current global data center consumption equals Germany or France; by 2030 it tracks toward India-level. Ireland already hit 20% of national grid from data centers and imposed a moratorium.
• Grid expansion is decade-scale. New power plants take ~10 years, transmission lines ~8 years, generator grid connections ~5 years. US electricity demand was flat for decades so the grid was never built for this. AI data centers run at near 100% utilization (unlike variable dotcom-era loads), making the problem worse.
• IMF modeled two scenarios: if energy supply responds fast, small price increases. If not (current trajectory), price surges steep enough to slow AI industry growth. There is a hard ceiling on AI scaling without radical energy breakthroughs.
• China built wrong. In 2022, China placed 500+ data centers in the west for cheap power, optimized for training workloads. But by 2024, 80% of new capacity sits unused because demand shifted to inference (latency-sensitive). Western China facilities cannot serve Shanghai users with acceptable latency. However, China added more electricity generation in 2024 alone than the US has in a decade.
• Gulf states hold the wildcard. Saudi Arabia and UAE have cheap energy (4x cheaper than US), abundant land, government-as-landowner efficiency, and geographic positioning to serve Europe, South Asia, and East Africa simultaneously. Abu Dhabi’s G42 was forced to choose between US and Chinese tech partnerships (chose US after Biden administration ultimatum).
• The US forced a geopolitical alignment. G42 divested Chinese partnerships, removed Huawei equipment, migrated to Microsoft Azure. Access to US chips is a “death sentence” without - giving the US leverage over Gulf AI infrastructure decisions.