You might not realize that every time you ask ChatGPT to write a 100-word weekly report or have Claude tweak a few lines of code, about 500 milliliters (equivalent to a bottle of Nongfu Spring) of pure freshwater somewhere on Earth evaporates into white steam.
Over the past two years, the grand narrative of the AI arms race has been firmly locked within a logical cycle of chips, computing power, and nuclear energy.
Huang Renxun loudly proclaimed the staggering computational power of ten-thousand- and hundred-thousand-GPU clusters at Computex Taipei; Musk tirelessly acquired land and built structures in Silicon Valley, stacking the largest supercomputer in human history—Colossus—in just 122 days, capable of handling up to 230,000 NVIDIA GPUs.
Capital markets are betting wildly on these "silicon-based myths." Yet everyone seems to have selectively forgotten the most fundamental, cold physical constraint—these sizzling silicon brains need water, and they drink freshwater, the very resource humans depend on for survival.
The United Nations University (UNU) has released a new global study on the environmental costs of AI, using a set of cold, hard figures to strip away the warm, virtual, low-carbon facade of AI: global daily AI prompt processing has surged to 2.5 billion.
By 2030, the global AI infrastructure is projected to consume a staggering 9.3 trillion liters (9.3 gigaliters) of water annually.
This amount is just enough to meet the basic drinking water needs of 1.3 billion people on Earth for an entire year.
From a super data center by the Mississippi River in Memphis, USA, to severely drought-stricken regions in Europe, a "freshwater arms race" driven by Physical AI and the massive computational demands of large models has officially begun in the summer of 2026.
I. Silicon Devours Carbon: The Insatiable Appetite of the Supercomputing Factory
Why do AI large models become an insatiable "water-sucking behemoth"? The answer lies hidden in the cooling architecture of data centers.
Currently, high-end GPUs such as NVIDIA’s latest Blackwell and the next-generation Vera Rubin architectures consume between 700 and 1200 watts per chip when operating at full load.
When thousands of these chips are densely packed into a server room, the entire data center essentially becomes a massive "high-heat boiler." If the heat is not removed within milliseconds, multi-hundred-million-dollar chips can instantly be destroyed by overheating.
To achieve optimal cost efficiency, over 70% of data centers worldwide use evaporative cooling systems.
The principle of this system is extremely primitive and wasteful: pumping large volumes of cold freshwater into the data center to absorb heat generated by the chips, then releasing approximately 80% of the water as vapor directly into the atmosphere.
This means that the vast majority of this consumed water cannot be recycled locally but instead vanishes directly from local groundwater and public water supply systems.
Let’s take a look at the undeniable real bills hidden in the sustainability reports of major corporations:
OpenAI (GPT series): According to independent researchers and investment banks, training GPT-4 once in a virtual environment consumed approximately 600 million liters of purified water—enough to fill 237 Olympic-sized swimming pools. The next-generation flagship model, currently under closed-door training, is expected to exceed one billion liters of water usage per training run due to an exponential surge in computational scale.
Google and Microsoft: In its latest environmental data disclosure, Google’s annual water consumption has surpassed 8.1 billion gallons (approximately 30 billion liters), marking a significant year-over-year increase; meanwhile, Microsoft’s water usage near its major model training hubs in Quincy and Iowa has nearly doubled over the past three years. Local residents in Iowa have begun protesting, as Microsoft’s five data center campuses are consuming millions of gallons of groundwater per day, directly competing with local farmland for this vital resource.
The bottomless pit of large models is evolving into a physical strain on Earth's real-world resources.
II. Memphis' "Watergate": Musk, Huang, and Outraged Residents
The most intense direct confrontation of this "freshwater takeover" occurred this year in Memphis, Tennessee, USA.
In 2024, Musk’s xAI team built the supercomputer cluster Colossus in Memphis in just 122 days—with Silicon Valley speed—to train the Grok large model. To keep this massive system running, which contains 230,000 chips, Colossus draws up to 1 million gallons (approximately 3.8 million liters) of potable water daily from Memphis’s local underground aquifer.
Because Musk adopted an aggressive “build first, ask questions later” strategy, bypassing environmental hearings, public outrage erupted in late 2025 when Memphis residents suddenly discovered their water bills had skyrocketed and groundwater levels had abnormally dropped during summer. Environmental groups and local communities sued xAI and local government, accusing the tech giant of “stealing the next sip of clean water from children’s mouths.”
Facing a massive legal and public relations crisis, Musk and Huang were forced in spring 2026 to make an exceptionally rare compromise: xAI urgently announced a $80 million investment to rapidly construct a "Greywater Recycling Plant" next to the data center.
Elon Musk’s solution: Since potable water isn’t available for residents, my AI will have to “drink wastewater.” The facility plans to use secondary-filtered industrial and municipal wastewater discharged from Memphis’s wastewater treatment plant to cool Colossus’s cooling towers, replacing purified freshwater.
Memphis's "Watergate" marks a pivotal turning point in the global history of Physical AI. It has proven to all enthusiastic tech investors that, beginning in 2026, the ultimate bottleneck limiting AI's expansion will no longer be TSMC's production capacity or Altman's dollars—but rather the local government's approval of water access rights.
Three: Wall Street's "New Anxiety" and the Tech Giants' "Zero Water Lie"
Facing growing public protests and a severe drought that affected nearly 63% of North America in 2026, the CEOs of tech giants are desperately trying to "tell new stories" on earnings calls and tech summits to reassure Wall Street.
At the recently concluded Microsoft Build 2026 conference at the end of May, CEO Satya Nadella dedicated a ten-minute segment to explaining Microsoft’s “Zero Water Revolution.”
Nadella declared in his speech: “Microsoft’s latest hyperscale data centers have completely eliminated evaporative cooling and adopted a brand-new ‘closed-loop, waterless chilled water system.’ We fill the cooling pipes with water once during construction, after which the system circulates endlessly between the servers and condensers, much like a household refrigerator—resulting in an annual water consumption equivalent to that of an average restaurant.”
But is this really the solution? In the eyes of self-media and academia, it resembles more of a "power consumption sleight of hand"—a game of pass-the-parcel.
The cost of a closed-loop cooling system: While a closed-loop system eliminates water evaporation, its heat dissipation efficiency is far lower than that of open evaporative cooling. To achieve the same cooling effect, the data center must rely on massive external fans and chillers with enormous power demands, causing electricity consumption to surge by 20% to 30%.
Transfer of indirect water footprint: The surge in electricity demand means power plants must operate at full capacity. Whether coal, gas, or nuclear power plants, their turbines require astronomical amounts of cooling water to generate electricity. According to calculations by the Lawrence Berkeley National Laboratory, while direct water consumption from evaporation in data centers amounts to 17.4 billion gallons, the indirect water footprint resulting from their electricity use reaches a staggering 211 billion gallons!
Microsoft saved water in its data centers, but caused power plants in another state to evaporate even more water. This kind of "treating the symptom, not the cause" greenwashing cannot hide the fact that AI is becoming an ecological disaster.
IV. Conclusion: The Key to Breaking Through in China’s AI, Smart Home, and Embodied Intelligence Sectors
While Western hyperscale data centers are being halted by environmental groups and regulations due to water resource and carbon emission constraints, this harsh physical reality about "freshwater" serves as a wake-up call for China’s rapidly expanding AI industry—but also provides a clear roadmap for its breakthrough.
China's AI industry must not blindly replicate Silicon Valley’s “industrial behemoth” model of recklessly stacking hundreds of thousands of chips in the cloud and consuming millions of gallons of freshwater daily. By 2026, the solutions we see should be more pragmatic and sophisticated:
First, there is the natural geographic hedge in computing power deployment. China inherently possesses the forward-looking national strategy of "East Data, West Computing." By firmly locating large training facilities—which require massive amounts of cooling—in regions like Guizhou and Inner Mongolia, which naturally feature karst underground rivers or extremely low average annual temperatures enabling passive air cooling, geographic advantages are leveraged to mitigate water resource concerns.
Second, and most importantly, the core technological breakthrough lies in the hybrid computing rearchitecture we have repeatedly mentioned earlier: “cerebellum at the edge, cortex in the cloud.”
Smart home brands represented by Haier Casarte, along with Chinese embodied AI robotics companies such as Agibot and Unitree, are fully advancing the development of lightweight edge-side chips.
If our vacuum robots, in-car intelligent cockpits, and industrial screwdrivers can resolve 90% of physical interaction tasks locally using just a十几-watt edge chip paired with a lightweight "spatial world model," without needing to send high-energy, multimodal prompts to a cloud server thousands of kilometers away for every minor action, then we would have cut AI’s water and power consumption at the foundational level by ninety percent.
Leave AI’s soul to the algorithms, and leave AI’s burden at the edge.
The battle over freshwater resources, already escalated between Musk and Huang, is forcing global AI to shed its superficial facade.
Is AI truly the ladder that propels human civilization to a higher dimension, or is it a silicon-based monster that will ultimately compete with humans for the planet’s last pure water? This summer of 2026, the answer is becoming increasingly clear as the water evaporates.
This article is from the WeChat public account "Xin Mang xAI," authored by Green and Dong Yizhen.