艾睿铂：2027年数据中心市场展望报告：需求超供给与监管趋严下的行业关键期（英文版）

AI training demands robust computational power, the kind GPUs were built to deliver. Inference has different criteria: It is memory-constrained, time-sensitive, and lower-precision. Winning in inference will require purpose-built silicon, not repurposed training hardware. The mismatch necessitates physical plant changes Facilities built between 2022 and 2024 were engineered for 20–40kW GPU training racks. Inference-native silicon increasingly demands 80–150kW+ of power, direct liquid cooling,