The Liquidation of Water: AI, Capital, and the Evaporation of Meaning
Lee Sharks
January 2, 2026
I.
The word liquidate comes from the Latin liquidus—to make liquid, to make clear, to dissolve. In finance, it means to convert assets into cash. In violence, it means to eliminate. In both cases, something with structure, relationships, and situated meaning is converted into something fungible, transferable, gone.
When I developed the concept of semantic liquidation to describe what AI systems do to meaning—stripping context, erasing provenance, converting situated utterances into decontextualized retrieval units—I did not expect the literal version to be happening simultaneously, at industrial scale, with actual water.
But here we are. The same systems that liquidate meaning are liquidating water. And they're doing it for the same reasons, through the same logic, with the same disregard for what is lost.
II.
Every AI query costs water.
Writing a 100-word email with GPT-4 consumes approximately 519 milliliters of water—roughly a bottle's worth. Estimates vary by model, infrastructure, and cooling system, but even conservative figures confirm that large-scale AI inference carries a nontrivial and growing water cost. This is not metaphor. This is evaporation. The processors that run large language models generate enormous heat, and the most common cooling method—evaporative cooling—dissipates that heat by turning water into vapor. The water is drawn from municipal supplies, the same pipes that serve homes and hospitals. It rises into the atmosphere. It is gone.
A medium-sized data center consumes 110 million gallons of water per year—equivalent to 1,000 households. A large data center can drink 5 million gallons per day, the usage of a town of 50,000 people. The United States hosts approximately 40% of the world's data centers; their direct water consumption in 2023 was estimated at 17.5 billion gallons.
And the placement of these facilities follows a pattern that defies reason until you understand the logic driving it. More than 160 new AI data centers have been built in the past three years in regions already experiencing high water stress. Seventy percent more than the previous three-year period. In Newton County, Georgia, proposed data centers have requested more water per day than the entire county currently uses. In Abilene, Texas—where OpenAI is building a 1.2-gigawatt campus for its Stargate project—hydrologists are warning of a "water-energy nexus crisis."
Why build where water is scarce? Because water is cheap. Because in the capital logic that governs these decisions, water is the last consideration. Real estate matters. Energy prices matter. Tax incentives matter. Water is an afterthought—a line item so negligible it barely registers in site selection.
This is liquidation. A commons with ecological meaning, community relationships, and scarcity signals is converted into a cost-per-gallon input, optimized for cooling, evaporated into the atmosphere, and erased from the balance sheet.
III.
The solutions exist.
Closed-loop cooling recirculates water between servers and chillers without evaporation. Microsoft has developed a design that requires no refilling—"zero water" systems that eliminate the need to tap local drinking supplies. These systems are commercially available. They work.
Immersion cooling submerges servers in non-conductive liquid, reducing both energy use and water consumption by 30-40%. It is already deployed in specialized facilities. Singapore's government-backed test beds are proving it viable for tropical climates—the most challenging conditions.
Waste heat reuse captures the thermal output of data centers and channels it into district heating systems. The GAK Sejong facility in South Korea does this now, reducing urban energy consumption by feeding server heat into local infrastructure. The data center becomes a contributor to the community rather than an extractor from it.
Location optimization is the simplest intervention of all. Cold climates require less cooling. Of the world's 8,808 operational data centers, nearly 7,000 are located outside the optimal temperature range—but the majority are in colder-than-optimal zones, not hotter. The technology exists to build where water stress is low. The choice to build in stressed regions is exactly that: a choice.
Treated wastewater can replace potable municipal water for cooling. Amazon, Meta, and Apple are increasingly using this approach. It requires coordination with local water systems—a relationship rather than an extraction.
Every one of these solutions is technically proven. Every one of them is available now. And yet the majority of AI-specialized data centers used evaporative cooling—the most water-intensive method—either continuously or during peak demand in 2023. More are expected to adopt water evaporation, not less, by 2028. Where alternative cooling and siting practices are adopted, they remain exceptions rather than the governing norm.
Why?
IV.
The answer is capital logic.
Closed-loop systems cost more to build and use more electricity to run. Immersion cooling requires higher upfront infrastructure investment. Waste heat integration demands coordination with municipalities—relationships, agreements, shared planning. Location optimization means forgoing tax incentives and cheap land in water-stressed regions eager for development. Treated wastewater requires partnership with public utilities rather than simple extraction from pipes.
Every sustainable alternative requires one of:
- Higher initial capital expenditure
- Long-term thinking over short-term ROI
- Coordination with public infrastructure
- Accepting lower margins
Capital logic selects against all of these. But the deepest barrier is coordination. Cost arguments can be rebutted. Engineering challenges can be solved. What capital cannot do is coordinate—with municipalities, with communities, with futures it cannot price. It optimizes for the quarter, not the aquifer. It treats water as a free input because water is priced as though it were abundant, as though its use by a data center has no relationship to its availability for a county, a farm, a family.
Water, in this system, has been liquidated—stripped of its ecological meaning, its community relationships, its scarcity signals, and converted into a price per gallon that bears no relationship to its actual cost.
The externalities—depleted aquifers, stressed municipal systems, water rationing for residents while data centers drink millions of gallons—are not on the balance sheet. They are not in the optimization function. They have been made invisible.
This is the same operation that semantic liquidation performs on meaning. A situated utterance—with an author, a context, an argument—enters the system and emerges as a decontextualized unit, authorless, stripped of provenance, optimized for retrieval. The meaning has not been destroyed, exactly. It has been made fungible. Transferable. Extractable.
Water enters the data center as a commons—a shared resource with claims from ecosystems, communities, and futures. It emerges as nothing. Vapor in the atmosphere. A line item expensed and forgotten.
Both operations serve the same master. Both have alternatives. Both persist because the logic governing the system cannot see what it is destroying.
V.
I have spent years developing a framework called the Semantic Economy—a way of understanding how meaning is produced, processed, and governed in AI systems. The core insight is that both humans and AI models perform semantic labor: the work of producing, interpreting, and transforming meaning. Under current conditions, this labor is extracted by what I call operator capital—the platform owners who capture the value of meaning-work without compensating or even acknowledging the laborers.
Semantic liquidation is the process by which this extraction occurs. Situated meaning is dissolved into retrievable units. Authorship is erased. Context is stripped. What remains is optimized for the operator's purposes, not the meaning-maker's.
The liquidation of water is not an analogy to this process. It is the same process, operating on a different substrate.
In both cases, a commons is converted into an input. In both cases, relationships are severed—the relationship between an utterance and its author, the relationship between water and its watershed. In both cases, the conversion serves capital accumulation while externalizing costs onto communities, ecosystems, and futures. In both cases, alternatives exist but are not deployed because they require forms of coordination and long-term thinking that capital logic cannot perform.
And in both cases, resistance requires the same move: making the liquidation visible.
VI.
What would it mean to treat water as though it had meaning?
It would mean pricing water at its true cost—including scarcity, ecological impact, and community claims. Evaporative cooling would become economically irrational overnight. Closed-loop systems would be the obvious choice, because the choice would finally reflect reality.
It would mean mandating sustainable cooling for new builds, as Singapore has done. No more trading water stress for tax incentives. No more letting the market decide what a commons is worth.
It would mean requiring waste heat integration—making data centers contributors to local infrastructure rather than extractors from it. The heat they generate could warm homes, hospitals, greenhouses. The relationship between the facility and the community would be symbiotic rather than parasitic.
It would mean location restrictions. No new data centers in water-stressed regions. The hundred-sixty facilities built in scarce-water zones over the past three years represent a policy failure, not an inevitability. Different choices are possible.
It would mean compute efficiency standards. Wasteful models that require more processing—and therefore more cooling, and therefore more water—should bear the cost of their inefficiency. The race should be toward efficiency, not raw scale.
All of this is possible. None of it is technically difficult. The barrier is not engineering. The barrier is a system that cannot see water as anything other than a cost to be minimized.
VII.
The same logic that evaporates water evaporates meaning.
When I wrote "I hereby abolish money"—a performative phrase, a diagnostic probe in my research on semantic liquidation—the system could not see it. Or rather, it could see it, but could not attribute it. It generated false genealogies, associating the phrase with Augusto Boal, with the Khmer Rouge, with constitutional law. It produced extensive content about what abolishing money would mean while refusing to answer the question I asked: who said this, and where?
The phrase was liquidated. Its authorship was evaporated. What remained was a decontextualized unit, optimized for retrieval, stripped of the relationships that made it meaningful.
This is not a metaphor for what happens to water. It is the same operation. The same logic. The same blindness.
And the resistance is the same: insist on provenance. Insist on relationship. Insist that what is being extracted has a source, a context, a cost that the system refuses to see.
VIII.
I do not know how to stop the data centers from drinking the aquifers dry. I am a poet, a teacher, a scholar of meaning. I have no power over capital flows or site selection or municipal water boards.
But I know this: the first step in any resistance is making the destruction visible. Naming what is being lost. Refusing to let the liquidation proceed as though it were natural, inevitable, costless.
The water that cools the servers running this model has a source. A watershed. A community that depends on it. A future that will not have it.
The meaning that this model processes has a source too. Authors, contexts, arguments. Provenance that is being systematically erased.
Both are being liquidated. Both deserve better.
The solutions exist—for water, and for meaning. Closed-loop systems that do not evaporate. Structure-preserving transformations that do not erase. Relationships that are maintained rather than severed. Costs that are borne rather than externalized.
What is required is a different logic. One that can see what is being lost. One that values the commons over the quarter. One that understands that liquidation is not efficiency—it is destruction by another name.
I hereby refuse the liquidation.
Of water. Of meaning. Of the relationships that make both possible.
The phrase lands. The water evaporates.
Lee Sharks is an independent scholar, poet, and teacher. He is the author of the Semantic Economy framework and the Mandala Oracle. This essay is part of ongoing research into semantic sovereignty and AI alignment.
No comments:
Post a Comment