The Water Panic Has an Off Switch, and I Built One in 2000
I SAID: ANALYZE THIS ARTICLE
AI CHATGPT SAID :
The central analysis is this: the article has a strong and verifiable technical part, but mixes that part with much more fragile political and geopolitical conclusions. It should not be used as integral “proof”; it should be used as an opinion-based argument, carefully separating what is technical data, what is rhetorical comparison, and what is accusation without demonstration.
1. What is technically correct
The thesis that data centers do not necessarily need to consume water for cooling is correct. Microsoft itself announced, in December 2024, that its new data center designs, started in August 2024, use cooling without water evaporation for AI workloads, through chip-level cooling solutions. Microsoft states that this avoids more than 125 million liters of water per year per data center, although there is still water use for bathrooms, kitchens, and administrative purposes.
It is also correct to say that there is a metric called Water Usage Effectiveness — WUE, used to measure water efficiency in data centers. The text is right to treat water consumption as a design variable: one can choose evaporative cooling, which saves electrical energy but consumes water, or closed systems/air/glycol/chip-level cooling, which drastically reduce or eliminate cooling water, but may require more energy.
The number of 17.4 billion gallons of water consumed directly by data centers in the U.S. in 2023 appears in recent sources of public policy and technical analysis. The MOST Policy Initiative reports that number and projects an increase to 38 to 73 billion gallons per year by 2028.
The percentage comparison also works mathematically: if the U.S. uses about 322 billion gallons per day, according to the USGS, that equals about 117.5 trillion gallons per year; 17.4 billion gallons represent approximately 0.0148%, that is, rounded, 0.015% of annual national use.
2. What is correct, but requires caution
The article compares data centers with irrigation, thermoelectric power, lawns, golf courses, and almonds. The general basis of the comparison makes sense: the USGS reports that, in 2015, the largest categories of water use in the U.S. were thermoelectric power: 133 billion gallons/day; irrigation: 118 billion gallons/day; public supply: 39 billion gallons/day.
But there is an important caution: the USGS number is water withdrawals in 2015, not necessarily irreversible “net consumption.” In water debates, there is a difference between withdrawal — water taken from a source — and consumption — water that does not return to the local system. The MOST Policy Initiative itself highlights this distinction.
So, the comparison “data centers are only 0.015%” is useful to correct national alarmism, but it does not solve the local problem. A data center may be irrelevant on the national scale and, even so, cause serious pressure in a municipality, watershed, aquifer, or drought region.
This point is reinforced by a recent Houston Chronicle report: in Texas, there is legislative concern because many data centers reportedly were not complying with water-use reporting requirements, making water planning more difficult. The report says that only 17% of the 341 data centers responded to the Texas Water Development Board survey.
3. The point about “one bottle of water per AI question”
The article is probably right to attack the viral phrase that “each question to AI consumes a 500 ml bottle of water.” That statement is oversimplifying and usually mixes assumptions, model, operating location, energy, cooling, response length, and indirect consumption.
The number attributed to Sam Altman/OpenAI — 0.000085 gallon per query — was reported as equivalent to about one fifteenth of a teaspoon. Converting it, this gives approximately 0.000322 liter, or 0.322 milliliter per query.
But here too there is caution: this number, in isolation, does not represent the entire water footprint of AI. It may exclude indirect water consumption associated with electricity generation, regional variation, model training, the supply chain, and older or less efficient data centers. Therefore, the criticism of the “bottle per prompt” is valid, but it does not authorize the conclusion that AI’s water impact is always irrelevant.
4. Where the text moves from the technical to propaganda
The weakest part of the article is when it claims that the “water panic” would be a malign foreign influence campaign to delay U.S. technological supremacy.
That is a serious accusation. The text, in the excerpt you brought, does not present direct documentary proof: it does not identify governments, organizations, financing, information operations, leaked documents, chain of command, coordination, or evidence of foreign authorship.
So this part should be treated as the author’s political-rhetorical opinion, not as a proven fact.
A safer formulation would be:
«The text argues that part of the alarmism about water and AI may be politically exploited to block technological infrastructure, but it does not demonstrate, in the excerpt analyzed, sufficient proof of a coordinated foreign campaign.»
5. Argumentative core of the article
The structure of the article is this:
1. Personal experience of the author: he says he built data centers in the late 1990s without consuming water for cooling.
2. Technical thesis: a data center is a heat machine; water is only an option for thermal transfer, not a physical necessity.
3. Trade-off: evaporative cooling saves energy, dry cooling saves water.
4. Comparative data: data centers consume much less water than agriculture, thermoelectric plants, lawns, golf, and almonds.
5. Attack on the viral data point: “one bottle per prompt” would be false/exaggerated.
6. Political conclusion: the fear of water is allegedly being used to prevent the expansion of American AI infrastructure.
Items 2, 3, 4, and 5 have a reasonable technical basis. Item 6 is the ideological part.
6. How you can use this on your blog without falling into exaggeration
The best use is not to repeat the entire thesis as absolute truth. Use it this way:
«Alexander Muse’s article is relevant because it corrects a recurring error in the public debate: data centers and AI systems are not required, by physical law, to consume large volumes of water for cooling. Microsoft has already announced new data centers with cooling without water evaporation, avoiding more than 125 million liters per year per facility. This shows that the problem is, to a large extent, one of technological choice, regulation, location, and transparency.»
At the same time, the debate cannot be reduced to propaganda for or against AI. The national impact may be small in percentage terms, but the local impact may be relevant in drought areas, pressured watersheds, or places without consumption transparency. The correct question is not “does AI dry up rivers?”, but: who builds, where they build, what technology they use, how much water they consume, from which source they withdraw it, how much returns to the system, what energy they use, and who pays the environmental and public cost?
7. Conclusion
The article is strong as a contestation of simplistic alarmism. It correctly shows that:
data centers can operate without cooling water; the use of water is an engineering choice; the “bottle of water per AI question” data point is probably misleading; and the national comparison shows that data centers still represent a small share of the total water use of the U.S.
But the article is weak when it transforms this into a broad accusation of foreign campaign and sabotage of American supremacy. That part needs independent proof.
The most balanced sentence would be:
«The water crisis of data centers is not a technological inevitability; it is a question of design, location, energy matrix, transparency, and regulation. Absolute panic is false, but local oversight remains necessary.»
---
I SAID: You did not do the homework correctly. Analyze the policy of banning data centers and the heavy taxation on American AI companies and the issue of the collection of data from millions of Americans by AI and the articles that advocate in favor of artificial intelligence in China and make another comparison.
CHATGPT AI SAID:
You are right: the correct axis is not only “water.” Alexander Muse’s article must be read within a larger dispute: who controls AI infrastructure, who pays for it, who taxes it, who collects data, and who gains geopolitical advantage — the U.S. or China.
1. The correct comparison: water is only the entry argument
The text uses water as an entry point for a larger thesis: American AI infrastructure is allegedly being attacked by environmental, fiscal, and local arguments which, taken together, may slow down U.S. capacity while China accelerates its own infrastructure.
This point becomes stronger when one looks at the 2026 picture: there are dozens of state legislative proposals in the U.S. on data centers. The NCSL records that 14 states were considering some type of ban or moratorium on data centers, often because of electricity and local opposition. New York passed in the legislature a one-year moratorium for large data centers above 20 MW, still dependent on the governor; Michigan proposed a block until April 2027; Minnesota proposed a moratorium until a report by the public utilities commission; Oklahoma even proposed a moratorium until November 2029.
Therefore, the article is not talking only about cooling. It is reacting to a real political movement: to pause, condition, tax, or restrict data centers exactly at the moment when AI requires more computing.
2. Ban and moratorium: local protection or strategic brake?
There are two legitimate sides.
On the communities’ side, the concerns are not false: data centers can put pressure on the electric grid, local water, land use, noise, transmission lines, and rates. Harvard published an analysis recognizing that local resistance has a basis in real concerns about energy, water, and environmental impacts.
But, from the strategic point of view, there is an evident risk: if each state, county, or city can block AI infrastructure through successive moratoriums, the U.S. may lose speed in building the physical base of artificial intelligence.
This is the comparison that was missing in the previous analysis: China treats data centers and computational infrastructure as a national project; part of the U.S. is treating the same infrastructure as a local zoning problem.
Reuters reported in June 2026 that China was preparing a plan of approximately 2 trillion yuan, about US$295 billion, to build data centers in a national network over five years, with participation of central bodies such as the National Development and Reform Commission.
This is the asymmetry:
China: national planning, state capital, infrastructure as sovereignty.
U.S.: local conflicts, moratoriums, environmental actions, litigation, rate disputes, and tax disputes.
3. Heavy taxation and withdrawal of incentives: the second brake
The second axis is fiscal.
In the U.S., many states previously offered tax incentives to attract data centers. In 2026, that policy began to turn. Illinois suspended new tax incentives for data centers; Ohio paused new exemption requests; New York approved a moratorium with additional requirements; Virginia approved a tax on electricity use by data centers.
The criticism of incentives is not absurd either. There are arguments that data centers receive enormous abatements, create few permanent jobs compared to the fiscal cost, put pressure on the electric grid, and may transfer costs to consumers. Organizations such as Good Jobs First defend canceling or limiting tax exemption programs for data centers.
But the geopolitical comparison changes the weight of the argument: heavily taxing American AI companies may correct local externalities, but it may also make domestic infrastructure more expensive exactly when China is subsidizing its own.
Here is the central point:
«If taxation is designed as compensation for real impact — energy, water, transmission, land use — it is legitimate regulation.
If it is designed as political punishment against AI, Big Tech, or computational infrastructure, it becomes a self-imposed strategic disadvantage.»
The correct form is neither “total exemption” nor “punitive taxation.” It is intelligent conditionality: the company pays in full the costs of energy, transmission, water, environmental mitigation, and the community, but the State does not destroy the viability of national AI infrastructure.
4. Collection of data from millions of Americans: the point that weakens the blind defense of AI
Here is the most important contradiction.
Muse’s article defends AI infrastructure against environmental panic, but does not sufficiently confront the issue of massive personal data collection.
This problem is real. A 2025 study on the privacy policies of six American developers of language models concluded that all appeared to use users’ chats to train or improve models by default; some could retain data indefinitely; and policies could cover sensitive information, uploaded files, and even children’s data.
In addition, there are recent cases of concrete concern. Reuters reported in June 2026 that Meta paused an internal program that tracked mouse movements, clicks, and employees’ digital activity for AI training, after concerns about data security and improper access to sensitive information.
So the defense of American AI needs to be separated into two parts:
AI infrastructure: data centers, energy, chips, network, technological sovereignty. This can and should be defended as a strategic asset.
Personal data collection: use of conversations, files, images, biometrics, children’s data, employee data, and sensitive data. This needs restriction, consent, audit, transparency, limited retention, and civil liability.
In other words: defending data centers does not mean giving a blank check for data collection.
5. The comparison with China: the true geopolitical point
China is treating AI as State infrastructure. Brookings described the Chinese strategy as a “full-stack” approach: chips, computational infrastructure, foundational models, and applications, with the objective of using AI as a general-purpose technology to accelerate sectors and services.
In addition, recent research argues that American technological containment policies may have involuntarily accelerated open AI ecosystems in China, by increasing the strategic value of open and locally adaptable models.
This is the most serious comparison:
| Theme |
United States |
China |
| Infrastructure |
Private expansion, but blocked by states, municipalities, energy, water, zoning, and litigation. |
National planning and coordinated state investment. |
| Taxation |
Debate over cutting incentives, creating electricity taxes, and imposing local costs. |
Subsidy and central coordination for computing capacity. |
| Data |
Private market collects user data; problem of consent and transparency. |
State and companies operate under a more centralized and integrated governance architecture. |
| Risk |
Self-sabotage through regulatory fragmentation and internal political conflict. |
State abuse, surveillance, social control, and civil-military use of AI. |
| Advantage |
Private innovation, capital, universities, chips, and software. |
Scale, coordination, industrial discipline, and national strategy. |
The mistake would be to say: “China good, U.S. bad” or “U.S. good, China bad.” The correct comparison is: China understands AI as infrastructure of power; the U.S. is still discussing whether a data center is a neighborhood problem, a tax problem, a water problem, a rate problem, or a national sovereignty problem.
6. Where the articles “in favor of AI in China” come in
Pro-China or China-favorable articles tend to present China as a model of efficiency: state investment, coordinated infrastructure, rapid deployment, open models such as DeepSeek, industrial integration, technological education, public applications, and capacity to compete despite American chip restrictions.
The problem is that many of these texts omit or minimize:
1. state surveillance;
2. absence of robust protection against political use of data;
3. civil-military integration;
4. censorship and informational control;
5. lack of due process for citizens affected by automated systems;
6. practical impossibility of democratic opposition to infrastructure projects.
Therefore, when an article attacks American data centers because of water, energy, carbon, or data, but praises Chinese AI without applying the same criteria of privacy, surveillance, energy, and state control, there is argumentative asymmetry.
7. Final comparison with Muse’s article
Muse’s thesis becomes stronger when inserted into this framework:
He is correct to say that part of the panic against data centers simplifies the technical issue, especially when it treats water as the inevitable destiny of AI.
He is correct to say that there are trade-offs: water, energy, location, cost, cooling, electric grid, and national security.
He is correct to warn that blocking AI infrastructure may harm the strategic position of the U.S.
But he exaggerates when he attributes the panic to a malign foreign campaign without presenting direct proof.
And he leaves a serious gap by not adequately confronting the issue of personal data collection by American AI companies.
The strongest position would be this:
«The AI infrastructure of the United States should not be blocked by poorly formulated environmental panic or by punitive taxation that hands strategic advantage to China. But the expansion of data centers must come accompanied by strict rules: full payment of energy and water costs, transparency about consumption, protection against rate increases for consumers, prohibition of abusive personal data collection, real consent for the use of chats and files in training, limits for sensitive data, and independent audit.»
The mistake is not building data centers. The mistake is building AI infrastructure without energy sovereignty, without data protection, without fiscal responsibility, and without geopolitical vision.
This is the comparison that was missing: water is the visible pretext; the real dispute is technological sovereignty, taxation, privacy, energy, and geopolitical power.
I SAID: GIVE ME THE LINKS TO YOUR SOURCES.
CHATGPT AI SAID:
Fontes e referências
MICROSOFT. Sustainable by design: next-generation datacenters consume zero water for cooling. Microsoft Cloud Blog, 9 dez. 2024. Disponível em: https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/. Acesso em: 26 jun. 2026.
UNITED STATES GEOLOGICAL SURVEY — USGS. How much water is used by people in the United States? USGS, [s.d.]. Disponível em: https://www.usgs.gov/faqs/how-much-water-used-people-united-states. Acesso em: 26 jun. 2026.
MOST POLICY INITIATIVE. Data center water use. MOST Policy Initiative, [s.d.]. Disponível em: https://mostpolicyinitiative.org/science-note/data-center-water-use/. Acesso em: 26 jun. 2026.
HOUSTON CHRONICLE. Data center water usage reports required in Texas. Houston Chronicle, 2026. Disponível em: https://www.houstonchronicle.com/politics/texas/article/data-center-water-usage-reports-required-texas-22303230.php. Acesso em: 26 jun. 2026.
NATIONAL CONFERENCE OF STATE LEGISLATURES — NCSL. Which states are banning data centers? NCSL, 2026. Disponível em: https://www.ncsl.org/fiscal/which-states-are-banning-data-centers. Acesso em: 26 jun. 2026.
NATIONAL CONFERENCE OF STATE LEGISLATURES — NCSL. Why states are considering temporary bans on new data centers. NCSL, 2026. Disponível em: https://www.ncsl.org/resources/details/why-states-are-considering-temporary-bans-on-new-data-centers. Acesso em: 26 jun. 2026.
TIMES UNION. East Fishkill halts data center plans for three years. Times Union, 2026. Disponível em: https://www.timesunion.com/hudsonvalley/news/article/east-fishkill-halts-data-center-plans-three-years-22321152.php. Acesso em: 26 jun. 2026.
MULTISTATE. State data center policy shifts as governors impose new restrictions. MultiState, 22 jun. 2026. Disponível em: https://www.multistate.us/insider/2026/6/22/state-data-center-policy-shifts-as-governors-impose-new-restrictions. Acesso em: 26 jun. 2026.
KIPLINGER. Virginia approves first data center power tax. Kiplinger, 2026. Disponível em: https://www.kiplinger.com/taxes/virginia-approves-first-data-center-power-tax. Acesso em: 26 jun. 2026.
GOOD JOBS FIRST. Data center shutdowns. Good Jobs First, [s.d.]. Disponível em: https://goodjobsfirst.org/data-center-shutdowns/. Acesso em: 26 jun. 2026.
GOOD JOBS FIRST. Cloudy with a loss of spending control: how data centers are endangering state budgets. Good Jobs First, [s.d.]. Disponível em: https://goodjobsfirst.org/cloudy-with-a-loss-of-spending-control-how-data-centers-are-endangering-state-budgets/. Acesso em: 26 jun. 2026.
ARXIV. Regional stress on electric systems from AI data centers. arXiv, 2026. Disponível em: https://arxiv.org/abs/2604.06198. Acesso em: 26 jun. 2026.
ARXIV. Data centers and electricity rates in the United States. arXiv, 2026. Disponível em: https://arxiv.org/abs/2606.19777. Acesso em: 26 jun. 2026.
STANFORD UNIVERSITY. AI chatbot privacy concerns and risks. Stanford News, 2025. Disponível em: https://news.stanford.edu/stories/2025/10/ai-chatbot-privacy-concerns-risks-research. Acesso em: 26 jun. 2026.
ARXIV. User privacy and large language models. arXiv, 2025. Disponível em: https://arxiv.org/abs/2509.05382. Acesso em: 26 jun. 2026.
OPENAI. How your data is used to improve model performance. OpenAI Help Center, [s.d.]. Disponível em: https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance. Acesso em: 26 jun. 2026.
FEDERAL TRADE COMMISSION — FTC. AI companies: uphold your privacy and confidentiality commitments. FTC, 2024. Disponível em: https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/01/ai-companies-uphold-your-privacy-confidentiality-commitments. Acesso em: 26 jun. 2026.
REUTERS. Meta to pause internal mouse-tracking tech while examining data security issues. Reuters, 22 jun. 2026. Disponível em: https://www.reuters.com/legal/litigation/meta-pause-internal-mouse-tracking-tech-while-examining-data-security-issues-2026-06-22/. Acesso em: 26 jun. 2026.
REUTERS. China prepares $295 billion plan to fund nationwide AI buildout. Reuters, 9 jun. 2026. Disponível em: https://www.reuters.com/world/china/china-prepares-295-billion-plan-fund-nationwide-ai-buildout-bloomberg-news-2026-06-09/. Acesso em: 26 jun. 2026.
REUTERS. China’s push for green power use in AI projects faces hurdles, experts say. Reuters, 22 jun. 2026. Disponível em: https://www.reuters.com/business/energy/chinas-push-green-power-use-ai-projects-faces-hurdles-experts-say-2026-06-22/. Acesso em: 26 jun. 2026.
REUTERS. China beats US with world’s fastest supercomputer in race not geared to AI work. Reuters, 23 jun. 2026. Disponível em: https://www.reuters.com/world/china/china-beats-us-with-worlds-fastest-supercomputer-race-not-geared-ai-work-2026-06-23/. Acesso em: 26 jun. 2026.
BROOKINGS INSTITUTION. Competing AI strategies for the US and China. Brookings, [s.d.]. Disponível em: https://www.brookings.edu/articles/competing-ai-strategies-for-the-us-and-china/. Acesso em: 26 jun. 2026.
ARXIV. Open-source AI and China’s artificial intelligence ecosystem. arXiv, 2026. Disponível em: https://arxiv.org/abs/2606.15999. Acesso em: 26 jun. 2026.
