A campaign in Quilicura, Chile, asked residents to be the AI for a day. According to the campaign's own site, on January 31, from 8 AM to 8 PM, more than 40 people from the town answered more than 25,000 prompts by hand. The stunt, an initiative of Fundación Avina and Corporación Ngen, was meant to make visible the water cost of the AI data centers clustered in Quilicura. This post separates what the campaign claims from what permits and press reporting document. It is a story about AI infrastructure, a topic our readers follow closely after the Ohio data center ballot fights and similar siting disputes.
TL;DR: what is claimed and what is documented
| Claim | Source | Status |
|---|---|---|
| 40+ residents answered 25,000+ prompts on January 31, 8 AM to 8 PM | Campaign site | Campaign's own account |
| Chats lasted an average of 11 minutes 4 seconds, with 68 countries taking part | Campaign site | Campaign's own account |
| Quilicura sits in the Maipo basin and hosts many AI data centers | Campaign site; Rest of World | Documented |
| Chile is in a drought expected to last until 2040 | Campaign site; Rest of World | Reported by both |
| Google's Quilicura center was authorized to extract 50 L/s | Rest of World, from official documents | Documented authorization, not measured use |
| Quilicura will run out of water in 10 years | Campaign site, citing "projections" | Not independently verified; no source given on the page |
| Data centers caused the wetland to dry | Activists | Contested; no study isolating the cause found |
What did the campaign actually do?
The campaign site describes a simple idea: people in Quilicura took the role of a chatbot. Visitors from around the world sent ordinary prompts, such as a request for an image of a sloth playing in the snow, a cake recipe, or a question about why hot dog buns and sausages come in different pack sizes. Residents and local artists replied by hand. The point, in the campaign's words, is that "someone casually using A.I. to create funny images in NYC or Berlin" can affect water supply in Quilicura.
The numbers on the page are small and readable: more than 40 people, more than 25,000 prompts, 12 hours, 68 countries. The page does not give a year for the January 31 session, does not say how the prompts were routed, and does not give a water-per-prompt figure. It is an awareness campaign, not a measurement, and we treat it that way.
What do permits and reporting show?
For the facts underneath, the best public source we found is Rest of World's 2024 investigation of data centers in Chile. Key points, as that report states them:
- Over 12 years, 16 data centers were approved in Santiago's metropolitan area, and Chile has had a recorded drought since 2010 expected to last until 2040.
- Google opened its Quilicura data center in 2015. Official documents seen by Rest of World show that by 2018 it was authorized to extract 50 liters of water per second from underground wells, more than 1 billion liters a year.
- Google's second Chilean data center was authorized to extract 228 liters per second, more than 7 billion liters a year. After protests by the Mosacat group and a review by Santiago's environmental tribunal, the project was suspended until Google reassessed its environmental impact, as of Rest of World's May 2024 report.
- A small 1 MW data center with water-based cooling needs about 25 million liters a year, Rest of World says, citing the general figure for such systems.
- Microsoft's Santiago-area data center, approved in 2023, was said by activists to rely partly on water despite earlier public claims about its cooling. Microsoft and Google did not respond to Rest of World's request for comment. An Amazon Web Services spokesperson said the company complied with its environmental declarations and aims to return more water than it uses by 2030 across its operations.
Please note the age: those reports are from 2024, and the status of each project may have changed. Check current records before you rely on them.
Timeline: how Quilicura became a data center hub
The sequence below uses only dates in the sources above.
| Year | Event | Source |
|---|---|---|
| 2010 | Chile's recorded drought begins; expected to last until 2040 | Rest of World |
| 2012 onward | 16 data centers approved in Santiago's metropolitan area over 12 years | Rest of World |
| 2015 | Google opens its Quilicura data center, its first in Latin America | Rest of World |
| 2018 | Quilicura site authorized to extract 50 L/s from wells | Rest of World, official documents |
| 2019 | Google announces a second Chilean data center; Mosacat protests begin | Rest of World |
| 2020 | Second Google site authorized at 228 L/s; Microsoft announces a $317 million site | Rest of World |
| 2023 | Microsoft's environmental approval granted; nearly 50 citizen observations filed on Amazon's Huechuraba site | Rest of World |
| 2024 | Environmental tribunal suspends Google's second project pending reassessment; Chile's government says it will launch a national data center plan | Rest of World |
The pattern is a country that used trade agreements to attract investment, and a community that found out about water terms by reading permit files. Activists told Rest of World that the companies "use water, dry up the territory, but don't give anything back to the population." That is a quote from an activist, not a finding, but it explains the anger the campaign taps.
A droplet above a cooling basin, showing the data center water use that the Quilicura human AI campaign tries to make visible
What have the companies said?
Statements are thin. Rest of World reported that Microsoft and Google did not respond to its request for comment. An Amazon Web Services spokesperson said AWS complied with the requirements of its environmental declarations for its Padre Hurtado and Huechuraba data centers, and that the company is committed to returning more water than it uses in its direct operations globally by 2030. That is a global pledge, not a figure for Quilicura. Chile's president, announcing 28 new data centers in 2024, assured the public that the investment would be "environmentally responsible," according to the same report.
Nothing in the sources we could read gives measured water use for any of these sites, and the campaign page does not name a company. If you work at an operator, the useful response is not a slogan but a site-level number, published on a schedule, with the cooling method stated. If you are a customer, ask for it.
What is still unmeasured?
Authorizations are ceilings, not meters. A permit to extract 50 liters per second does not tell you how much a data center actually pulled. Activists quoted by Rest of World have pushed for transparency on the water and energy use of each site, which implies those numbers are not easy to find. Without them, nobody outside the companies can say how much of Quilicura's groundwater decline comes from data centers versus farming, households and drought.
The campaign's "run out of water in 10 years" projection is the weakest link. The page cites "projections" without naming one. We could not trace it to a hydrological study, so we do not repeat it as fact. The drought itself is well supported by the documents above.
Why this matters beyond Chile
Water is the quiet constraint behind AI capacity. We have covered power constraints and siting fights across the U.S., from Senate permitting talk on who pays for transmission to OpenAI's grid protection pledge in Georgia, and abroad in Oman and the UK's Selby site. Water adds a second axis: where there is power there may not be enough water for the cooling method chosen.
Operators have options. Closed-loop and air-based cooling cut water use at a cost in electricity. For a different extreme in the siting debate, see our post on Google's energy chief saying space data centers could cost up to 100x more than Earth sites. The Chilean case shows that a cooling choice is also a local political choice.
What can people who use AI do?
The campaign's message is guilt-adjacent, so it is worth being precise. No source here gives a per-prompt water number, and the per-prompt effect of one person's chat is tiny next to industrial cooling choices. The useful levers are about demand patterns and disclosure:
- Prefer providers that publish site-level water and cooling data. Ask for it in procurement.
- Route simple tasks to smaller models, which cost less compute and so less cooling.
- Cache repeated requests and avoid regenerating the same output.
- Ask where inference runs. A region with abundant water and cool air differs from a drought basin.
- Follow local permit processes if a data center is proposed near you. Chile's system lets citizens file observations on environmental declarations, and activists used it heavily.
Bottom line
The Quilicura campaign is a clever way to make a hidden cost visible, and its headline numbers describe the stunt, not the water. The documented story is permits, drought and a contested wetland, with Google, Microsoft and Amazon at the center and many numbers undisclosed. If you want to cite water figures for AI, cite measured disclosures when they exist and label permit ceilings as ceilings.
Details are accurate as of November 3, 2026, based on the campaign site and Rest of World's 2024 reporting. Project statuses may have changed.