The explosive growth of artificial intelligence is redrawing the map of global data centers. When OpenAI used more than ten thousand AI agents to crack a 90-year-old unsolved math problem in 88 hours, the compute it consumed was enough to make a mid-sized city's power grid tremble. This reveals an increasingly clear core contradiction: AI's capability ceiling is increasingly determined by whether data centers can secure enough electricity, water, and—more importantly—smarter ways of operating.
The Physical Cost of Runaway Compute
Stanford University's 2026 AI Index Report reveals a set of sobering numbers: the United States has 5,427 AI data centers, more than ten times that of any other country; AI data center power capacity has climbed to 29.6 gigawatts, roughly equivalent to New York State's entire peak demand.
Water consumption is equally striking. GPT-4o's annual inference water use is estimated to exceed the drinking water needs of 12 million people. In a white paper jointly released by China's CAICT and Grundfos, it is projected that by 2030, U.S. data centers alone will add between 697 and 1,451 million gallons of daily water supply demand just to meet cooling needs.
"The mismatch between water use efficiency and the pace of compute expansion further amplifies local water supply-demand conflicts," the report notes. In the AI era, the core competitive edge has shifted from simply comparing compute scale to a comprehensive contest of compute, energy, and water resource coordination.
When AI Starts Managing Itself
Facing this predicament, the industry's answer is surprisingly elegant: use AI to optimize AI's own infrastructure.
At NTT DATA's data center in Bonn, Germany, a hybrid AI-based cooling optimization system is already running. The system combines physical models with data-driven AI algorithms to create a "digital twin" of the cooling system, simulating and optimizing chiller operations in real time. Initial results show a 19.1% reduction in cooling system energy consumption, with projected annual savings of up to 25%.
Digital Realty deployed Phaidra's AI platform at its Northern Virginia data center. The company's 2025 impact report shows that while its data center portfolio grew 34%, water consumption increased by only 3%. This "huge divergence" is attributed to AI-driven optimization systems that monitor pumps, filters, and the full infrastructure stack in real time.
AWS has gone further. Its network operations now incorporate AI agents that automatically correlate telemetry from multiple monitoring systems, completing root-cause identification in seconds rather than the traditional minutes. For routine operational issues, the AI reviews tickets, compares historical patterns, and in many cases resolves problems without human intervention.
A New Paradigm for Grid Flexibility
The real breakthrough may come from the concept of "flexibility." Emerald AI's Conductor software represents the most radical approach: letting data centers actively sense grid needs and modestly reduce compute load during power stress—without affecting critical task execution.
In one test, Conductor took control of 256 Nvidia A100 GPUs—hardware consuming roughly as much power as about 170 U.S. homes. When the grid came under pressure, the system reduced chip power by 25% for three hours while maintaining acceptable compute performance.
This "speed-to-power" model is drawing attention from utilities. In Hillsboro, Oregon, Aligned Data Centers agreed to install a 31-megawatt battery system to reduce draw during grid congestion. This flexibility measure, combined with others, allowed local utility Portland General Electric to add 80 megawatts of supply capacity for surrounding data centers without building new power plants.
Google has been shifting processing loads from high-demand regions to less stressed facilities since 2023, and has signed agreements with five utilities adding up to 1 gigawatt of flexibility.
The Battle Over Cooling Technologies
The choice of cooling method directly determines a data center's resource footprint. According to the CAICT report, open-loop cooling systems rely on evaporation to remove heat, consuming large amounts of water; closed-loop dry cooling systems circulate coolant through pipes, reducing evaporative water consumption to extremely low levels.
The reason northern China has become a compute hub—Ulanqab in Inner Mongolia, Zhongwei in Ningxia, Qingyang in Gansu—is not only abundant green power but also climatic conditions that allow low-water cooling. But an easily overlooked contradiction is that these regions are precisely the most water-scarce in the country. A data center's "absolute water consumption" may not be large, but its "relative pressure on water resource carrying capacity" cannot be ignored.
The value of liquid cooling extends beyond mere cooling efficiency. Grundfos technical experts point out that liquid cooling systems can return water at temperatures of 40-50°C, making it possible to heat surrounding buildings. "The significance of liquid cooling isn't just high cooling efficiency—it's that it allows data centers to transform from pure energy consumers into resource nodes that can export thermal energy. "
Investment Frenzy and Rational Concerns
Capital continues to pour in. Bell Canada announced it would expand its AI data center in Saskatchewan from a planned 300 megawatts to 1.2 gigawatts, increasing investment to 52 billion Canadian dollars—the largest private capital investment in the province's history. Yet Wall Street has begun weighing the possibility of a slowdown in AI development. Anthropic CEO Dario Amodei's proposal to slow the pace of frontier model development directly triggered a selloff in AI infrastructure stocks—GE Vernova fell nearly 9%, Vertiv nearly 8%.
This tension reveals a deeper problem: the current data center construction boom is, to a large extent, a bet on sustained compute demand from a handful of frontier AI companies. Any change in the pace of model development transmits rapidly through the supply chain.
Meanwhile, a new safety concern is emerging. The evolution of AI model architectures—such as "recurrent depth" techniques—is migrating more reasoning processes into a "latent space" unobservable to humans. This means the "chain-of-thought" window developers use to monitor whether models are engaging in deception or reward hacking may gradually narrow.
The relationship between AI and data centers is evolving from a simple "demand and supply" dynamic into a complex symbiosis. Data centers use AI to manage their own energy and water consumption, while AI's progress in turn defines the shape of data centers. The ultimate test of this transformation may be this: between compute hunger and resource constraints, can the industry find a path that sustains capability growth without overdrafting environmental and social trust?
