✦   ✦   ✦

breaking news

News, read through The Primal Race
← 全部评论 · all commentary

AI 喝掉的水,是谁的饮用水AI Drinks the Water, the Community Pays

科技 结构层 · 文化层 Rest of World ↗ 2026-10-31 § 链接
新闻摘要2022年ChatGPT发布后,大型科技公司数据中心的水资源消耗引发关注。加州大学研究估计,训练初版ChatGPT的大语言模型耗水18.5万加仑。在智利、乌拉圭等地,谷歌等公司的数据中心项目因当地干旱和水资源紧张而遭到居民反对。
The newsAfter ChatGPT's release in 2022, water use by Big Tech data centers drew attention. A University of California study estimated training the model behind the first ChatGPT required 185,000 gallons of water. In Chile and Uruguay, Google data center projects faced local opposition amid drought and water scarcity.
数据中心用社区的水冷却,把账单留给不喝这口水的人。
Data centers cool themselves with community water and leave the bill to those who never drink it.

数据中心的水不是技术问题,是分配问题。它要的是干净的、能直接喝的饮用水,因为处理自己的水太贵,而公共供水系统既便宜又现成。于是它和社区抢同一口水,抢赢了,代价由不喝这口水的人付。

智利 Cerrillos 的数字最清楚:谷歌自己预测的数据中心每秒用 169 升水,一天用掉的水等于当地 8 万居民一整年的饮水量(来源:Rest of World)。它要从 2005 年起就受限、审批前几周刚升级为全面禁止新取水的地下含水层抽水。这不是疏忽,是算过账的:水从哪里来最便宜,就从哪里来。

结构层的暴力在这里不需要一个坏人。环境评估照程序走,项目照计划批,水照抽。居民要反对,得先读懂用技术语言写的报告——用一位研究者的话说,那是一个没几个律师、没几个生态专家的工人阶级社区。信息不对称本身就是结构暴力的一部分:不是不让你反对,是让你看不懂你在反对什么。

MOSACAT 做的事,是把解释权从技术报告里抢回来。他们翻材料、开会、传播,把项目对水的需求翻译成居民能判断的数字,然后推动 2019 年 12 月的公投:49% 反对,38% 支持。公投不具强制力,但它制造了一个公司无法忽视的事实。2022 年 2 月谷歌带着新方案回来,说改用风冷、不用水。这是居民赢下的一局,赢的方式是把“技术细节”变成“公共事实”。

但别急着庆祝。谷歌在智利遇到阻力后,转头去了乌拉圭——当地研究者要上法庭才拿到环境部手里的用水数据,发现 72 英亩的选址计划每天用 200 万加仑水。乌拉圭人二十年前刚把饮用水写进宪法当基本权利,现在发现这项权利在一点点滑走。资本会移动,社区不会。一个地方守住了,水就流向另一个没守住的地方。

所以真正的账本不是“AI 用了多少水”,而是“谁有权决定这些水先给谁”。数据中心承诺带来就业和繁荣,实际带来的是短期施工岗位和长期的水缺口。这套“繁荣换资源”的话术,和第三章里那些武器化的叙事是同一个结构:先定义什么是进步,再让被定义的人为进步买单。

巴西 Fortaleza 郊外那座相当于 12 个足球场、造价约 100 亿美元、实际服务于 TikTok 的数据中心,选址在一个 2003 到 2024 年间有 16 年处于紧急干旱状态的地区。这不是巧合,是选址逻辑:哪里便宜、哪里监管松、哪里的人最没有能力说“不”,就去哪里。

水是原初种族被掠夺的第一样东西——生育、身体、劳动,然后是水。当社区把水权重新攥回手里,他们缩小的不是某个公司的利润,而是 Violence = Potential − Actual 里那个被当成“现状”的差额。这个差额还没缩完:公投赢了,公司换了个国家;风冷方案来了,下一个社区还在等法庭排期。

The water a data center uses is not a technical problem. It is a distribution problem. It wants clean water fit for drinking, because treating its own is expensive and the public supply is already there. So it competes with a community for the same glass of water, wins, and the cost lands on people who never drink it.

The numbers from Cerrillos, Chile, are the clearest: by Google's own projection, the data center would use 169 liters of water every second — in one day, as much as all 80,000 residents use for drinking in a year (source: Rest of World). It wanted to draw that water from an aquifer restricted since 2005, upgraded to a full ban on new extraction just weeks before approval. That is not an oversight. It is an accounting: take water from wherever it is cheapest.

Structural violence here needs no villain. The environmental review follows procedure, the project is approved on schedule, the water is drawn. To object, residents first have to read reports written in technical language — in what one researcher calls a working-class community where few are lawyers or ecology experts. The information gap is itself part of the structure: it is not that you are forbidden to object, it is that you cannot understand what you are objecting to.

What MOSACAT did was take the power of interpretation back from the technical report. They read, met, organized, translated the project's water demand into numbers residents could judge, and forced a December 2019 referendum: 49% against, 38% for. The vote was not binding, but it manufactured a fact the company could not ignore. In February 2022 Google returned with a new proposal: air-cooling, no water. That is a round the residents won, and they won it by turning a technical detail into a public fact.

Do not celebrate too early. After hitting resistance in Chile, Google moved to Uruguay, where a researcher had to go to court to get the water-use figures from the Environment Ministry and found a 72-acre site planning 2 million gallons a day. Uruguayans had written drinking water into their constitution as a basic right two decades earlier; now they watched that right slip. Capital moves. Communities do not. One place holds the line, and the water flows to the next place that did not.

So the real ledger is not how much water AI uses. It is who gets to decide who drinks first. Data centers promise jobs and prosperity; they deliver short-term construction work and a long-term water gap. The prosperity-for-resources pitch is the same structure as the weaponized narratives in Chapter 3: first define what progress is, then make the people it is defined against pay for it.

Outside Fortaleza, Brazil, a data center the size of 12 football fields, costing an estimated $10 billion and built to serve TikTok, is sited in a region under emergency drought conditions for 16 of the 21 years between 2003 and 2024. That is not coincidence. It is siting logic: go where it is cheap, where regulation is loose, where people are least able to say no.

Water is the first thing taken from the Primal Race — reproduction, body, labor, then water. When a community takes its water rights back into its own hands, it shrinks not some company's profit but the gap in Violence = Potential − Actual that was being passed off as the status quo. The gap is not closed: the referendum was won, and the company changed countries; the air-cooling plan arrived, and the next community is still waiting for a court date.