作者回忆自己在1990年代末于东京采访寿险公司倒闭：日本资产泡沫破裂后，不到4年就有7家寿险公司破产，波及数以百万计保户。这段经验让他对「以退休金与寿险资金投入AI资料中心」格外警惕，尤其当SoftBank提到其在美国规模高达5,000亿美元的Stargate资料中心计划可能引入寿险与退休基金作为长期金主。

他指出，市场也出现跨国大型配置：加拿大退休金投资委员会与澳洲Goodman Group拟在欧洲资料中心投资「数十亿美元」。支持者主张资料中心属长期基础设施，规模越大、边际算力成本越低，但作者认为目前仍缺乏从投资热潮走向「可验证的高额利润」的清晰路线图。

他特别质疑愿景式算术：孙正义声称，若「实体AI」在10年后可赚到全球GDP的10%，而他估计当时GDP约200兆美元，则未来10年即使投入10兆美元，也能在6个月回本。作者提醒，保守资金应避免被「数兆美元」级支出叙事与「拥抱AI或灭亡」的话术推著走；即便替代投资只占资产很小一部分，寿险私募化案例（如有人缴了17年保费后因资本缺口而家属仍在追讨）也显示风险不容低估，且氛围像网路泡沫前的基础设施过度投资。

The author recalls covering life-insurer failures in late-1990s Tokyo: after Japan’s asset bubble burst, 7 insurers collapsed in under 4 years, hurting millions of policyholders. That history shapes his unease as AI data-center megaprojects court long-term capital. SoftBank has floated a US “Stargate” build costing $500 billion, with potential backers including life insurers and pension funds—money meant to be exceptionally conservatively managed.

He notes similar scale elsewhere: Canada’s pension investment board and Australia’s Goodman Group plan to invest billions in European data-center projects. Proponents argue long-horizon funds match long-horizon AI infrastructure, and that bigger builds drive down the marginal cost of compute. But he argues the industry lacks a fully accounted roadmap that connects today’s capital surge to “lavish” profits sufficient to justify it.

He highlights how optimistic math can outrun reality: Masayoshi Son claimed that if “physical AI” earns 10% of global GDP in 10 years, and GDP reaches about $200 trillion, then spending $10 trillion over the next decade could be recouped in six months. The author worries that “trillions” talk from figures like Sam Altman, plus exhortations to “embrace AI or perish,” could pull retirement and insurance savings into a dot-com-like infrastructure boom—despite past cautionary tales such as a policyholder who paid premiums for 17 years before his insurer’s private-equity ownership left a capital shortfall.