本报告以问答形式系统梳理人工智能建设的经济学。核心结论：AI资本开支仍有上行空间；变现能否覆盖资本成本是唯一关键问题；当前最佳机会在科技以外，即AI采用者、大宗商品与实体基建。超大规模云厂商WACC已从2023年的4%升至10%，自由现金流预计到2027年下半年才由负转正；AI资本开支增速50%远超收入增速20%，而到2030年的算力需求对应约10万亿美元支出，高出市场预测2万亿。企业端若仅10%美国工人被替代，对应最高每年3万亿美元AI潜在市场。基准情形下2030年AI超额回报约每年2000亿美元，情景区间从每年亏损5000亿到盈利近1万亿。建设1GW数据中心成本250亿至450亿美元，约60%投向AI服务器与GPU，电力、冷却、网络及土建环节受益。对标普500未来12个月风险回报持均衡看法，信息技术与通信服务均给予标配。
Using a Q&A format, the report examines the economics of the AI buildout. Key conclusions: AI capex likely has further to run; monetization versus cost of capital is the only question that matters; and the best opportunities now may lie outside Tech—in AI adopters and commodities/physical buildout plays. The WACC for hyperscalers has risen to 10% from 4% in 2023, and their free cash flow is expected to turn positive only by 2H27. AI capex has surprised by a much wider and accelerating margin than revenue (50% vs. 20% this year), while estimated compute demand through 2030 implies ~$10tn of spend, $2tn above forecasts. If just 10% of US workers are rationalized, enterprise AI TAM could reach as much as $3tn annually. Base case excess returns are ~$200bn p.a. by 2030, ranging from -$500bn to nearly +$1tn p.a. Building 1GW of data center capacity costs $25-45bn, with ~60% for AI servers/GPUs, benefiting power, cooling, networking and construction. The team sees balanced risk/reward for th