Where the Cloud Touches Ground 云落在哪里

"The cloud" is the most successful piece of evasive language of our era. Every time someone says send it to the cloud, a specific concrete building somewhere in a specific American county receives the bytes, runs them across silicon, and sends them back. The cloud is geography — it just isn't a geography most of us are ever asked to look at.

That geography is concentrated, and concentrated by design. Counties get chosen for their tax breaks, their cheap electricity, their proximity to the fiber backbones laid down in the 1990s. Loudoun County, Virginia, is denser than the rest by an order of magnitude: its data centers carry roughly seventy percent of global internet traffic. Around it, a handful of secondary clusters: Phoenix, Dallas, Hillsboro, Quincy, Des Moines, New Albany. Together they consume an amount of electricity that would, on its own, rank around the twenty-fifth-largest country in the world.

The bill comes due locally. Phoenix data centers compete with cities for the dwindling Colorado River. Loudoun residents have been in transmission-line battles with their utility for two years. Ohio farmers near New Albany lost row crops to Intel's semiconductor expansion and the data centers it pulled in. The job a map can do here is small but specific: take the per-prompt water number that floats around in articles and put it back where it actually lands — on a county, in a basin, beside a transformer.

AI demand is going to make this harder. Generative models have already pushed the annual power-demand growth curve away from a flat plateau and onto a near-vertical line. The map you're looking at is a 2026 snapshot; in three years it will need to grow about three times as many dots, and many of those will land on the same handful of counties — because once the substation is built and the fiber is in, the next tenant comes for the same reasons.

A small disclosure. Capacity figures are illustrative. Operators rarely publish exact numbers; the values plotted here are reconstructed from public filings, news coverage, and industry reports. Locations are accurate to within the metropolitan area. The water-stress wash across the western states is a sketched climatological footprint, not a hydrological dataset.

「云」是这一代最成功的回避词。每次有人说「放到云里就行」，地球上某个具体县的 某栋具体建筑就接收了那串字节，让它们在硅上跑一遍，再发回来。云是地理—— 只是大多数人从来没被邀请看一眼那张图。

这片地理是高度集中的，而且是有意为之。一个县被选中，是因为它有税收减免、 电力便宜、紧贴 1990 年代铺下的光纤主干。其中弗吉尼亚州的 Loudoun 县密度 比其他所有地方高一个数量级：这里的数据中心承载了大约百分之七十的全球 互联网流量。围绕 Loudoun，还有几个二级集群：凤凰城、达拉斯、希尔斯伯勒、 昆西、得梅因、新奥尔巴尼。这些集群一起消耗的电量，单算下来能在世界国家 排名里挤进前二十五。

代价是地方承担的。凤凰城的数据中心在和市民争夺日益干涸的科罗拉多河水。 Loudoun 县居民已经和电力公司打了两年的输电线官司。俄亥俄新奥尔巴尼的 农民为了让 Intel 半导体扩建工厂、以及随之而来的数据中心，失去了大片 庄稼地。一张地图能做的事很小但很具体：把那个偶尔被引用的「ChatGPT 每条提问耗水多少毫升」的数字，重新放回它真实的落地处——某个县、 某条流域、某座变电站旁边。

AI 需求会让这件事更难。生成式模型已经把年电力需求增长曲线从一条平的 高原推上了几乎直立的斜线。你现在看的这张图是 2026 年的快照；三年之后 它需要长出大约三倍的点，而其中很多还会落在这几十个相同的县—— 因为一旦变电站和光纤已经铺好了，下一家进驻者也是看重了同样的理由。

附注：容量数据为示意性数字。运营方很少公开精确容量；图中数值 由公开备案文件、新闻报道与行业报告反推。地理位置精确到大都市区。 西部各州那层水资源紧张的暖色铺底，是大致的气候带勾勒，不是水文数据。