I am strongly pro-artificial intelligence. I believe AI will become increasingly important to education, scientific research, medicine, business, engineering, and nearly every other knowledge-intensive profession. I also believe efforts to broadly slow or prohibit AI development are unlikely to be either practical or beneficial. However, supporting AI advancement does not mean that we should ignore the physical infrastructure required to make that advancement possible.
A recent Politico Magazine article, The New Luddites, examines some of the growing resistance surrounding artificial intelligence and the infrastructure emerging around it. I am hesitant to categorize legitimate questions about energy, water, and infrastructure as simply anti-technology sentiment. There is an important distinction between opposing technological progress and asking whether we are developing that technology intelligently.
AI may appear almost entirely digital when we interact with ChatGPT, Claude, Gemini, or another model on a laptop or smartphone, but the computational infrastructure behind these systems is decidedly physical. AI requires processors, servers, cooling systems, electrical generation, transmission infrastructure, land, and increasingly large data centers. As our appetite for computational power grows, so does the physical footprint required to supply it.
The scale deserves attention. Lawrence Berkeley National Laboratory reported that U.S. data centers consumed approximately 176 terawatt-hours of electricity in 2023, equal to about 4.4 percent of total U.S. electricity consumption. Its more recent 2025 update estimates that data centers could represent approximately 11.8 percent of U.S. electricity consumption by 2030, with modeled scenarios ranging from 9.5 to 15.3 percent. Those numbers do not provide an argument against AI. They provide an argument for making energy efficiency and computing infrastructure major components of the AI research agenda.
Water creates another challenge, particularly for those of us living in the American West. Arizona, Nevada, California, and other Western states already face difficult decisions surrounding groundwater, the Colorado River, agriculture, municipal development, population growth, and long-term water security. In that environment, adding another water-intensive industry deserves careful analysis. This is particularly important because some conventional data-center cooling systems depend upon evaporation, meaning water consumption can become a continuous component of heat removal. The Department of Energy notes that cooling-tower water consumption is directly connected to heat generated by IT equipment and describes several techniques that can reduce both water and energy demands.
At the same time, the issue should not be oversimplified into the claim that every AI query consumes some fixed quantity of water or that every data center operates the same way. They do not. Location, climate, cooling technology, electricity sources, hardware efficiency, server utilization, and facility design all influence environmental impact. Some modern facilities are moving toward closed-loop cooling systems that continually recirculate coolant rather than depending as heavily upon evaporative water consumption. The Department of Energy has specifically pointed to these systems as a means of substantially reducing water requirements.
This distinction is important because it changes the nature of the problem. If all AI development necessarily required enormous and permanently increasing quantities of fresh water, we might eventually face an unavoidable conflict between computational growth and environmental resources. Instead, much of the evidence suggests that engineering choices can dramatically influence the environmental footprint of computation. That means sustainability is not merely a constraint on AI development. It is another technological problem waiting to be solved.
That is where I believe much more innovation is needed.
The next generation of AI infrastructure should advance closed-loop cooling, direct-to-chip liquid cooling, dry cooling where appropriate, reclaimed-water systems, waste-heat recovery, renewable and low-water electricity generation, more efficient processors, improved workload scheduling, and intelligent systems capable of optimizing their own energy and cooling requirements. There will inevitably be tradeoffs. A system that dramatically reduces water use may require additional electricity, while another design may reduce electrical consumption while relying more heavily on evaporative cooling. The goal should therefore be optimization across the entire system rather than simply minimizing one environmental metric in isolation.
There is an interesting irony here. Artificial intelligence itself may ultimately help solve many of these infrastructure challenges. Digital twins, predictive analytics, reinforcement learning, sensor networks, and AI-driven control systems can potentially optimize cooling, predict equipment loads, coordinate workloads with available electrical capacity, identify inefficiencies, and reduce unnecessary resource consumption. Recent research is already exploring data-center cooling optimization using digital twins and sophisticated control strategies, demonstrating that infrastructure efficiency is itself becoming an active area of computational research.
For Western states, however, technology alone should not eliminate the need for thoughtful planning. A data center constructed in a region with abundant renewable electricity and water resources presents a fundamentally different resource equation than an identical facility constructed in a desert community dependent upon stressed groundwater or Colorado River allocations. Infrastructure decisions should reflect those regional realities.
I would like to see the conversation evolve from asking simply, “How many data centers can we build?” toward asking, “How much computing capacity can we create per unit of energy and water consumed?” Those are very different questions. The second treats efficiency as a competitive advantage and creates incentives for researchers, technology companies, utilities, engineers, and policymakers to improve the underlying architecture of AI.
We should also become more sophisticated about how we measure environmental performance. Power Usage Effectiveness has long been used to evaluate data-center energy efficiency, while Water Usage Effectiveness provides a similar mechanism for examining water consumption relative to IT energy use. As AI infrastructure expands, these measures—and perhaps new measures that combine energy, water, carbon, computational output, and regional resource availability—could become increasingly important.
Ultimately, I do not believe the choice should be between artificial intelligence and environmental sustainability.
If AI is truly the transformational technology that many of us believe it to be, we should expect more from the infrastructure supporting it. We should expect engineers to develop better cooling systems. We should expect chip manufacturers to produce greater computational output per watt. We should expect facilities to reuse water when possible, recover waste heat when practical, integrate increasingly efficient energy systems, and locate computational resources where they can be sustained.
Most importantly, we should see these challenges as opportunities for innovation rather than reasons to abandon technological progress.
The United States should continue pursuing leadership in artificial intelligence. But leadership should mean more than building the largest models or accumulating the greatest number of GPUs. It should also mean developing the world’s most efficient, resilient, and environmentally sustainable computing infrastructure.
For those of us in Arizona and throughout the American West, that distinction is especially important. Water is simply too valuable a resource to treat as an unlimited input into technological expansion.
We do not need less AI.
We need better AI infrastructure.
And I suspect that some of the most consequential innovations of the next decade will not occur only inside the models themselves, but inside the data centers that make those models possible.
You can read the Politico article below:
https://www.politico.com/news/magazine/2026/08/02/the-new-luddites-01017824