When people debate the environmental impact of AI, the conversation almost always lands on energy and carbon emissions. AI water consumption rarely gets the same attention — and that’s a significant oversight.

AI’s water footprint is large, growing fast, and largely invisible to the people building and deploying the systems that drive it. Understanding where that water goes, why it’s needed, and what can actually be done about it is increasingly urgent — not just for environmentalists, but for infrastructure architects, data center operators, and anyone making long-term bets on AI at scale.

The numbers are striking. Researchers estimate — based on available data center performance metrics rather than AI-specific disclosures — that the water footprint of AI systems could reach 312.5 to 764.6 billion liters, figures projected from 2025 data and likely higher today, roughly equivalent to the global annual consumption of bottled water.

At the individual model level, the scale is equally sobering. Training the GPT-3 language model in Microsoft’s U.S. data centers is estimated to have directly evaporated 700,000 liters of clean freshwater. At the prompt level, researchers at the University of California, Riverside estimate that each 100-word AI prompt uses roughly one bottle of water — 519 milliliters. Multiply that across billions of daily interactions and the cumulative demand becomes significant.

Regionally, the picture is just as stark. A study by the Houston Advanced Research Center found that data centers in Texas alone will use 49 billion gallons of water in 2025, potentially rising to 399 billion gallons by 2030 — equivalent to drawing down Lake Mead, the largest reservoir in the US, by more than 16 feet in a single year.

The answer comes down to heat — and how the industry has chosen to manage it.
A data center’s water footprint is calculated across three categories: on-site water usage for cooling, water used by power plants supplying electricity to the data centers, and water consumed during the manufacturing of processor chips.

The dominant factor is cooling. GPU clusters — the backbone of modern AI training and inference — generate enormous amounts of heat. Most data centers use a combination of chillers and cooling towers to prevent chip overheating, with water evaporation being the most common and cost-effective method. The challenge is that much of this water doesn’t return to the local supply — it evaporates in the process.

Even a mid-sized data center can consume as much water as a small town, while larger facilities require up to 5 million gallons per day — equivalent to a city of 50,000 people. And unlike energy, which can be sourced and transmitted across distances, water availability is fundamentally local. A data center drawing from a regional aquifer in a drought-stressed area creates concentrated, irreversible pressure on that community’s water supply.

The major cloud and AI companies have not ignored this issue. Most have made public commitments to water replenishment — pledging to return more water to watersheds than they consume by a target date.

The problem is that replenishment pledges don’t reduce consumption. They offset it — and often in different watersheds than where the water was drawn. Google reported that its data centers consumed approximately 5.6 billion gallons of water in 2023, a 24 percent increase compared to the previous year, even as its sustainability commitments remained in place. Voluntary efforts and efficiency improvements help at the margins, but the trajectory of AI’s water footprint indicates a rapidly escalating challenge — one that efficiency trends alone are unlikely to reverse if current architectural practices continue.

Newer cooling technologies — direct-to-chip liquid cooling, immersion cooling — can reduce water intensity per kilowatt-hour. But they address the symptom, not the cause. As long as AI workloads continue to be driven by power-hungry GPU clusters running backpropagation-based training, the underlying demand for cooling will keep growing.

Here is the part of the conversation that rarely gets said plainly: AI’s water problem is not primarily an infrastructure problem. It is an architectural one.

The processors inside modern data centers — GPUs and TPUs running deep learning workloads — are doing an enormous amount of high-precision floating-point arithmetic. But the arithmetic itself isn’t even the main energy drain. Research from Stanford University finds that up to 90% of the energy consumed during AI training is spent simply moving data in and out of memory, not performing calculations. Every time a processor fetches parameters from memory, shuffles activations between compute clusters, or synchronizes gradients across devices, it expends far more energy than the underlying math. That data movement generates heat, heat requires cooling, and cooling requires water.

This is a consequence of the Von Neumann architecture that underpins virtually all modern AI hardware: compute and memory are separate. Every training operation requires data to be shuttled back and forth across that divide, billions of times per training step. The waste is structural, not incidental.

It’s worth putting this in perspective: the biological brain — the original intelligence system — operates with roughly six orders of magnitude less energy than modern AI systems performing equivalent tasks. That gap isn’t a hardware problem waiting to be solved with better cooling. It’s a signal that the architectural approach itself may be fundamentally misaligned with efficiency.

Replenishment pledges and more efficient cooling towers are reasonable steps. But they are downstream fixes to an upstream problem.

Brain-CA’s approach targets the architectural root cause rather than the cooling infrastructure built around it.

The problem with conventional AI isn’t just that it uses a lot of power — it’s why it uses a lot of power. Backpropagation requires computing partial derivatives across billions of parameters using high-precision floating-point arithmetic. Every training step demands millions of multiply-accumulate operations. And as established above, it’s not just the arithmetic — it’s the massive data movement that arithmetic requires. That combination generates heat, heat requires cooling, and cooling requires water. The waste is structural, not incidental.

Brain-CA’s architecture eliminates both problems at once.

At the core of Brain-CA’s approach is the Estimator — a patented hardware primitive that converges on probability estimates through a process more like a binary search than a calculation.

Think of it this way: if you wanted to guess an unknown probability, a classical approach would count observations, divide, and apply Bayes’ rule. That requires division, growing counters, and significant memory. The Estimator takes a different path. It maintains a single integer — a Markov state index — and updates it using only bit comparisons and ±1 nudges. No multiplication. No division. No floating-point arithmetic. No explicit application of Bayes’ rule.

Yet the results are Bayesian-quality. The Estimator’s expected value at every observation count has been proven to fall between two classical Bayesian posterior means — converging on the correct probability the way a binary search converges on a target value, zooming in logarithmically fast. The equilibrium that emerges is what a Bayesian statistician would compute, at a fraction of the hardware cost.

An 8-bit Estimator requires fewer transistors than an 8×8 multiplier — the kind of multiplier that conventional AI relies on billions of times per training step. At the scale of thousands or millions of Estimators running in parallel on a single chip, this difference is decisive.

Conventional AI’s energy crisis is as much a data movement problem as a computation problem. The Estimator solves both simultaneously.

Because the Estimator’s update operations — bit comparisons and a conditional ±1 nudge — are simple enough to execute directly where the data is stored, each Estimator updates its own state locally, in parallel, without competing for bus bandwidth. There is no need to shuttle data to a central processor and back. No data movement. No Von Neumann bottleneck.

Less arithmetic. Less data movement. Less heat. Less cooling. Less water.

The Estimator is the foundational unit — but Brain-CA’s Learning Engine is designed to scale. Multiple Estimators can work in concert, tracking relationships across data streams and building progressively more sophisticated models of the environment they’re observing. The system is designed to discover structure rather than have it engineered in — moving from single-stream probability estimation toward the kind of relational learning that complex real-world applications require.

This scalability is what makes the architecture relevant not just for simple sensors, but for the data-intensive workloads that currently drive the heaviest demand on data center cooling infrastructure.

This is not vaporware. The Estimator is protected by three issued U.S. patents. The architecture was presented at the International Symposium on Computer Architecture (ISCA) 2025, where it won Most Innovative Paper at the Energy Efficiency workshop. A proof-of-concept chip is currently being fabricated at TSMC.

The long-term vision is a class of AI hardware that achieves sub-linear scaling — performing real-time learning at power budgets measured in milliwatts, not because of more efficient cooling, but because the underlying computation is genuinely, structurally simpler than what it replaces.

AI’s water consumption is becoming a business and infrastructure risk, not just an environmental one.

Data centers are increasingly competing with agriculture, municipalities, and ecosystems for access to freshwater in drought-prone regions. Regulatory pressure is growing — several countries and US states are beginning to require environmental disclosures from data center operators, and mandatory reporting requirements are likely to tighten. Communities are pushing back on new data center permits when local water resources are already stressed.

For organizations planning long-term AI infrastructure, water availability is becoming a site selection criterion in the same way that energy costs and connectivity already are. Architectures that reduce compute intensity — and therefore cooling demand — will carry a material advantage as these constraints tighten.

AI’s thirst is not an inevitable consequence of intelligence. It is a consequence of a specific architectural approach to intelligence — one that relies on brute-force arithmetic and massive data movement running on hardware that generates significant heat as a byproduct.

Replenishment pledges and better cooling towers are reasonable short-term responses. But they treat the symptom. Brain-CA is working on the cause — building toward a class of AI hardware where the underlying computation is structurally simpler, the data movement is structurally reduced, the heat generated is structurally lower, and the water consumed is structurally less.

The question isn’t whether AI can be more efficient. It’s whether the industry is willing to look far enough upstream to find out why it isn’t.

Sources: de Vries-Gao, Patterns/Cell Press (Dec 2025); Lincoln Institute of Land Policy (Oct 2025); Environmental and Energy Study Institute (Jun 2025); TRENDS Research & Advisory (Feb 2026); Nature — “Fixing AI’s Energy Crisis” (Nov 2024)