In June 2026, SpaceX revised its prospectus before going public.
The revisions weren't about rocket technology, satellite internet, or Mars colonization plans. The newly added risk warnings pointed to something more fundamental: water. The document stated that water scarcity, drought conditions, competition for local water resources, or regulatory restrictions on water use could limit the company's ability to obtain sufficient cooling water, thereby slowing data center expansion, or even force the adoption of more expensive alternative cooling solutions. Electricity, processors, and water were listed side-by-side in the filing, constituting the core resource constraints for AI computing power expansion.
This was the first time SpaceX systematically highlighted water risk in a public document. A company known for Starship and Falcon rockets was reminding its potential shareholders to pay attention to the stability of the water tap.
The original sentence of that risk warning read: "water scarcity, drought conditions, competition for local water resources, or regulatory restrictions on water use could limit our ability to obtain sufficient water for cooling... delay or limit expansion... or require us to implement alternative cooling techniques that may be more costly." The wording was flat, the restrained tone of standard legal documents. But its very appearance in an IPO filing was the signal.
SpaceX's AI business entity is xAI. According to a TechCrunch report in May, xAI posted an operating loss of $64 billion in 2025, with annual revenue of $32 billion and capital expenditures continuing to soar. This rate of cash burn corresponds to the frantic rollout of server rooms, servers, and computing clusters. When hardware and infrastructure investments of tens of billions of dollars per quarter become the norm, supply fluctuations of any physical resource are no longer just a cost item for operations departments to optimize. It becomes a risk that must be disclosed to investors.
The shift of water from an operational cost into a risk disclosure framework is worth marking.
Operational costs are things a company can manage itself, by saving money or changing technical solutions. But risk is different. Risk is an external variable, something a company may not be able to control. Drought is a weather problem; local government tightening of water permits is a policy problem; community opposition is a political problem. These are issues a company cannot necessarily solve by throwing more money at them.
TechCrunch noted in its report that this revision reflects how the AI industry's reliance on natural resources is triggering new attention at the regulatory and investment levels. Analyzing this judgment requires first answering a more basic question: How much water do AI data centers actually use?
17 Billion Gallons, Just the Direct Cooling Part
Lawrence Berkeley National Laboratory provides a set of estimates. In 2023, water consumption in U.S. data centers for direct cooling was approximately 17 billion gallons, about 64 billion liters. And that's just direct cooling. Data center operations rely on electricity, and the power generation process itself consumes vast amounts of water. Cooling for thermal and nuclear power plants, evaporation for hydropower — adding these indirect water uses together, the same estimates show a staggering figure of 211 billion gallons.
17 billion gallons of direct water use, 211 billion gallons of indirect water use. The latter is over 12 times the former. When discussing the water footprint of AI, the direct cooling data you see above the surface is just the tip of the iceberg.
This estimation also provides a trend projection: by 2028, direct cooling water use for U.S. data centers could increase by 2 to 4 times. The wide range exists because it depends on the pace of AI computing expansion, choices in cooling technology, and the geographic distribution of new data center construction. A doubling is the most conservative scenario; a quadrupling is the aggressive expansion scenario. Either way, the direction is sharply upward.
These numbers themselves are abstract. When applied to specific companies, the scale becomes clearer.
Google disclosed in its sustainability report that it consumed 6.4 billion gallons of water in 2023, with 95% used for its data centers. This means Google's data centers alone drank about 6 billion gallons of water that year. One site, the data center in Council Bluffs, Iowa, consumed 1 billion gallons of potable water in 2024 alone.
Meta's figures are slightly smaller but still substantial. In 2023, Meta consumed 813 million gallons of water globally, with 95% also for data centers.
Putting these numbers together, the water usage of Google's data centers alone roughly amounts to over one-third of the Lawrence Berkeley Lab's estimated total direct cooling water for all U.S. data centers. The single site in Council Bluffs, Iowa, uses enough water annually to support a medium-sized city.
Where does this water go?
Most large data centers use evaporative cooling technology. The principle isn't complicated: water contacts hot air in cooling towers, evaporation carries away heat, turning into water vapor released into the atmosphere. This process is called "consumptive water use." The water is used up; it does not return to rivers, lakes, or underground aquifers. This differs from residential water use where shower and dishwater, after treatment, can be returned. What comes out of a data center cooling tower is steam. It's consumed, in the literal sense.
A study published in 2021 in npj Clean Water, a Nature journal, gave a technical scale: a typical 1-megawatt IT load data center, using traditional evaporative cooling, consumes about 25.5 million liters of water annually. A 1-megawatt IT load roughly corresponds to the computing power of several hundred servers. Large data centers often have tens or even hundreds of megawatts. Multiplying at this scale, a 50-megawatt data center cluster can easily consume over 1 billion liters of water annually for cooling.
What this consumption means in arid regions needs no further explanation.
Building Water Tigers Next to Deserts
In April 2025, a Guardian investigation reported that Amazon, Microsoft, and Google were operating and expanding data centers in some of the world's driest regions, with the collective scale of the three companies' data centers projected to expand by 78%. Behind these numbers lies a series of ongoing conflicts.
In Mexico's central Querétaro state, 17 of 18 municipalities suffer severe drought. At the same time, the state has become a data center cluster for international tech giants. Local residents held signs outside data centers: "No queremos centros de datos, queremos agua" — We don't want data centers, we want water. BBC provided detailed coverage of this conflict.
In Mesa, Arizona, USA, according to a Business Insider report in June 2025, Meta reached a water agreement allowing its facility to use up to 4 million gallons of water per day. What does 4 million gallons represent? Based on an average U.S. resident's daily water use of about 82 gallons, this is equivalent to the daily water use of nearly 49,000 people. Arizona itself is one of the most water-stressed regions in the U.S., with Colorado River levels declining year after year and states bickering over water allocations. A single data center drawing 4 million gallons daily is legal, compliant, but not without controversy.
Similar voices have emerged in Australia. A December 2025 report by The Guardian Australia noted that as large-scale data center construction accelerates, drinking water supplies in some regions are facing direct competition. The acute impact of sudden jumps in data center water usage even in developed countries with relatively mature water planning systems indicates this is not an isolated case of poor governance in a particular region, but a universal contradiction between scale expansion and resource scarcity.
The commonality in these disputes isn't that tech companies are "violating water use rules." They are not violating any rules. Every water use agreement has undergone legal approval, every water resource fee is paid according to regulations. The root of the problem lies in the fact that the existing water allocation frameworks were established in an era when data centers were not major water users. When a single data center's daily water consumption equals that of a town, compliance itself becomes the problem. The system hasn't kept pace with the growth of the water tigers.
A Guardian investigation in October 2025 also revealed another dimension. Amazon has long refused to disclose detailed water usage for its data centers, accused of strategically concealing its full water footprint. Google discloses per-site data, Meta releases global aggregated data, while Amazon provides the least information. This divergence in disclosure is starting to be seen by analysts as a variable in risk assessment. The less a company is willing to tell you about its water usage, perhaps precisely the more likely its water usage is to cause controversy.
Projects Halted, Water is the Reason
Water resource disputes have moved beyond the realm of public opinion. They are now materially blocking project implementation.
An industry report from Data Center Watch shows that over the two years since mid-2024, approximately $64 billion worth of data center projects in the U.S. have been blocked or delayed due to local community opposition. Water consumption is one of the core protest reasons, alongside power grid strain and noise pollution. The report documents 142 grassroots opposition groups across the U.S., spanning different states and political spectra, achieving rare consensus in opposing mega data centers.
Water is becoming a new weapon of NIMBYism. Past NIMBY movements focused on substations, waste treatment plants, highways. Now data centers have joined this list. The reasons have changed, but the logic remains. Residents' logic is simple: You say your data center contributes to the economy, but if the cost is lower water pressure at my home, rising water bills, falling well water levels, I won't accept that cost.
Once such opposition forms, it cannot be resolved by a company holding a few community presentations or promising a handful of jobs. Electricity can come from new power plants, fiber can be newly laid, land can be bought at a premium. But in the eyes of residents, water has no substitute. Things with no substitute have very little room for negotiation.
Throughout 2025, according to industry statistics, about half of the data center projects originally planned to go online in 2026 were canceled or delayed. This proportion is enough to make any company planning AI infrastructure expansion reconsider its site selection logic. In the past, data center site selection prioritized: electricity, fiber, land price, climate. Now the weight of water is catching up.
In February 2026, the University of California, Berkeley's Center for Law, Energy & the Environment (CLEE) released a dedicated report exploring how to regulate data center water use in California. This is the first time academia has directly addressed this issue in a thematic report. The release of the report itself is an indicator: when top law schools and energy policy think tanks begin systematically studying regulatory frameworks for data center water use, it shows the issue has crossed the boundary of internal industry discussion and entered the public policy agenda.
Investors Start Calculating the Water Bill
The capital markets are also following up.
In April 2026, according to the Journal Record, investors formally urged Amazon, Microsoft, and Google to disclose more data on data center water consumption. The report also cited a macro figure: North American data centers already used nearly 1 trillion liters of water in 2025.
1 trillion liters is a number difficult to intuitively grasp. To put it another way: it's roughly equivalent to the storage capacity of a large freshwater lake. The Lawrence Berkeley Lab's 2023 estimates were large enough, but compared to the actual consumption in 2025, they might be conservative.
The investment community's changing attitude has a clear trajectory. Previously, water resources appeared in ESG reports, listed alongside other environmental indicators, more as a form for corporate social responsibility departments to fill out. Now it's different. Water has jumped from the "corporate image" section to the "operational risk" section. Shareholders care less about environmental protection than about whether there is enough water to keep the servers running. When the stability of water supply begins to affect revenue projections, it's no longer an ESG topic, but a financial one.
Clear differentiation has emerged in the response strategies of different companies. Google continues to publish per-site water usage data and claimed in 2024 to have replenished 4.5 billion gallons of water through water replenishment projects. Meta publishes aggregated data. Amazon, following The Guardian's investigation, still hasn't publicly disclosed detailed water usage per site. This divergence further reinforces a viewpoint: the transparency of water usage data itself is becoming a variable for analysts to assess the risk exposure of AI infrastructure companies.
Companies are also attempting technical responses. Switching to air cooling can reduce direct water use but often increases power consumption. Liquid cooling technologies can use water at higher temperatures (NVIDIA's Vera Rubin platform supports cooling with 45°C water), but system deployment costs are higher. Each technical pathway makes a trade-off between water consumption and electricity consumption; there is no perfect universal solution. Ultimately, what determines a data center's cooling solution may not be the technically optimal one, but local water prices, electricity prices, and policy tolerance. Technological choices become compromises under resource constraints.
An Ironic Contrast
In March 2026, OpenAI CEO Sam Altman said a widely circulated line in a public speech. According to Business Insider, he expressed it this way: "We see a future where intelligence becomes a utility like electricity or water, where people buy it from us by the meter."
This statement sparked much discussion at the level of copyright and business models, but its other layer of meaning is more tangible. When Altman compares AI to water and electricity, the actual physical operation of AI is consuming real water in the real world. The industry's imagination for the business model is to package AI as an inexhaustible infrastructure, metered and paid for like turning on a water tap. At the same time, SpaceX's prospectus is frankly admitting: without enough water, AI might not run at all.
Before a service is compared to water and electricity, its infrastructure has already incurred massive bills for water and electricity. This contrast itself is the most accurate description of the cognitive state of the AI industry in 2026.
Looking back chronologically, the narrative path is clear.
From 2023 to 2024, annual water usage data of global leading cloud providers was passively disclosed through sustainability reports. Lawrence Berkeley National Laboratory released estimates, providing the first macro picture of U.S. data center water usage. Community conflicts in places like Querétaro, Mexico, and Mesa, Arizona began entering mainstream media view.
In 2025, systematic investigative reports by The Guardian, BBC, and others pushed the connection between data center expansion in arid regions and local water stress into public discourse. Data Center Watch released quantitative statistics on $64 billion in blocked projects. Investors began formally requesting greater transparency on water footprints.
In 2026, SpaceX extracted this chain from public discussion and industry reports and placed it in the "Risk Factors" section of its IPO prospectus. This is the formal transition of the water issue from a public opinion topic to an investment pricing factor. Before an investor subscribes to SpaceX stock, they need to sign off acknowledging they are aware: this company's AI business could face problems due to water scarcity.
This is how capital markets price resource constraints. They don't care about sentiment, corporate promises, or sustainability visions in PR releases. They care about one thing: what, under what circumstances, could cause expected returns to fall short. Water supply is affected by weather, water prices by policy, water access by community opposition — these three things a company cannot control. Things that cannot be controlled are risks. Risks need to be written into the document to remind investors.
This mechanism itself is reshaping the logic of AI infrastructure expansion.
In recent years, the main narrative of the AI race has been the computing arms race. Chips, electricity, talent were the three elements. Water was a hidden condition, assumed to be available locally. Now that taken-for-granted premise is shaken. In arid regions, in cities where local water is already tight, in places where regulation is beginning to tighten water quotas, "local water availability" is no longer an automatically valid assumption.
The expansion of AI infrastructure is no longer just a game of technology and capital. It is entering a stage requiring simultaneous negotiation over resource allocation with four groups: local residents, local governments, regulators, and investors. The speed of the computing race may not be determined by the fastest company, but by the slowest water meter.
