文章将AI资料中心扩张比作「新罗马帝国」：从早期主机、1990年代网际网路机房、2000年代云端虚拟化一路演进，如今转为专为生成式AI打造的「超大型」资料中心。叙事强调规模与版图意象（跨三大洲、约占地球圆周的九分之一），并点出更快、更高效晶片与更大算力需求正把资本与基础设施推向新高。
投资与交易以惊人的金额与能耗指标衡量：Stargate 计划先承诺 1,000 亿美元、未来拟上看 5,000 亿美元；后续合作以 4.5 吉瓦容量与约 10 万个工作为指标。微软称 2025 年初预计投入约 800 亿美元扩建AI资料中心。9月辉达称最多投 1,000 亿美元入股OpenAI，条件是OpenAI采购最多 10 吉瓦的辉达系统；隔月AMD称若OpenAI于 2030 年前部署最多 6 吉瓦AMD GPU，AMD可给到最多 10% 股权，呈现「你付我、我再投你」的循环。
下游成本被量化为能源、水与交通外部性：估计显示全球AI用电需求将在今年底超过比特币挖矿；资料中心散热需大量用水且未必充分揭露，导致地方井水干涸或疑虑。路网压力也被具体化：路易斯安那州里奇兰郡一处（Meta 的 270 亿美元 Hyperion 资料中心所在地）今年车祸暴增 600%。同时需求端被描绘为庞大（每周约 8 亿人使用ChatGPT），但作者质疑成长、资源与回报的「数学」是否成立，并以泡沫风险收束。
The piece frames AI data-center expansion as a new “Roman Empire”: from mainframes to 1990s internet server farms to 2000s cloud virtualization, the industry is now building huge, AI-wired facilities. It stresses scale and footprint, invoking an empire spanning three continents and about one-ninth of Earth’s circumference, and argues generative AI pushes demand for faster chips and far more compute.
Capital spending is described through striking dollar figures and power targets. The Stargate effort is pledged at $100 billion to start with plans up to $500 billion; a later partnership is measured at 4.5 gigawatts and around 100,000 jobs. Microsoft said in early 2025 it aimed to invest about $80 billion in AI-enabled data centers. Nvidia then said it would invest up to $100 billion in OpenAI if OpenAI used up to 10 gigawatts of Nvidia systems; AMD said it could give OpenAI up to 10% of AMD if OpenAI deployed up to 6 gigawatts of AMD GPUs by 2030.
Downstream impacts are quantified as energy, water, and local disruption. Estimates suggest global AI electricity demand will surpass bitcoin mining by year’s end; cooling draws on municipal water with incomplete disclosure, and wells can run dry or seem unsafe. Near construction, traffic and crashes rise; one area of Richland Parish, Louisiana—home to Meta’s $27 billion Hyperion facility—saw a 600% spike in vehicle crashes this year. While demand is portrayed as massive (about 800 million weekly ChatGPT users), the author questions whether the economic and ecological math—and bubble risk—adds up.