Pennsylvania Voters Unite Across Party Lines Against Data Centers: The Human Cost of AI Compute Expansion

Community Backlash Against the AI Infrastructure Boom

As artificial intelligence enters a period of explosive growth, hyperscale data centers are expanding across the United States at an unprecedented pace. However, this infrastructure frenzy driven by tech giants is encountering increasingly strong resistance from local communities.

Recently, voters across multiple Pennsylvania localities have united in a rare cross-partisan alliance to oppose data center construction in their communities. The sentiment "People are going to get screwed" captures ordinary residents' deep anxiety about this AI boom.

This phenomenon deserves reflection from the tech industry: as AI's compute demands climb exponentially, who ultimately bears the physical costs behind it—land, electricity, water resources?

Why Data Centers Have Become a Target

Rising Electricity Prices: Direct Pressure on Residential Bills

One of Pennsylvania residents' core concerns about data centers is their impact on the local power grid and electricity rates. Modern AI data centers are massive power consumers—a single large facility can use as much electricity as a mid-sized city. Take a typical AI training cluster: a single NVIDIA H100 GPU has a thermal design power (TDP) of 700 watts, and a large-scale training cluster often deploys tens of thousands or even hundreds of thousands of GPUs. GPU power consumption alone can reach tens to hundreds of megawatts. Add storage systems, networking equipment, cooling facilities, and power distribution losses, and a hyperscale AI data center's total power requirement can easily exceed hundreds of megawatts or even reach gigawatt scale. The industry typically uses PUE (Power Usage Effectiveness) to measure data center energy efficiency—the ratio of total facility power consumption to IT equipment power consumption. Current industry average PUE is about 1.5 to 1.6, meaning for every unit of electricity consumed for computing, an additional 0.5 to 0.6 units are needed for cooling and auxiliary systems like power distribution. According to U.S. Energy Information Administration (EIA) estimates, data centers already account for approximately 4% of total U.S. electricity consumption, and this proportion is rapidly climbing—potentially doubling by 2030.

When these high-energy facilities connect to regional grids, they often drive up local wholesale electricity prices, and this cost ultimately transfers to ordinary residents' and small businesses' utility bills. In the PJM power market that covers Pennsylvania (the largest regional transmission organization in the U.S., spanning 13 states and Washington D.C.), surging data center demand has already caused capacity auction prices to rise significantly over the past two years, with power capacity bids in some areas increasing several-fold compared to previous levels. This means all users connected to that grid—whether or not they use AI services—are indirectly paying for compute infrastructure expansion.

For many working families, this means passively increased monthly living expenses without receiving any direct benefits. Those "getting screwed" are precisely ordinary residents who cannot obtain employment or tax benefits from data centers.

Water Resource Competition and Environmental Concerns

Beyond electricity consumption, data center cooling systems also require massive water resources. Understanding this requires knowledge of mainstream data center cooling technologies. Currently the most common approach is evaporative cooling, which operates on the same principle as natural water evaporation removing heat: water is introduced into cooling towers, and through evaporation absorbs heat generated by servers, thereby lowering the temperature of circulating air or coolant. Water evaporated in this process cannot be recovered, constituting data centers' "consumptive water use." A 100-megawatt data center is estimated to consume hundreds of millions of gallons of water annually, equivalent to the annual water consumption of thousands of households. Google and Microsoft environmental reports show that in 2022, the two companies' global data centers used approximately 5.6 billion and 6.4 billion gallons of water respectively, and water consumption continues to climb annually as AI operations expand. In recent years, some tech companies have begun exploring liquid cooling technologies (including direct contact and immersion cooling) to replace traditional air cooling plus evaporative cooling approaches, but large-scale deployment of these technologies remains in early stages.

In some regions, data centers' massive water demands directly compete with local agricultural irrigation and residential daily water use. Particularly in water-stressed areas of the U.S. Midwest and Southwest, this competition has already triggered intense community conflicts. Additionally, sprawling data center campuses alter community land use patterns, bringing sustained noise (mainly from large cooling fans and backup diesel generators), light pollution, and long-term impacts on local ecosystems.

These environmental externalities are difficult to simply quantify, yet they tangibly erode community residents' quality of life.

Cross-Partisan United Resistance: An Intriguing Political Signal

The most striking aspect of Pennsylvania's protests is how they break through deeply entrenched partisan divisions in American politics. Both conservative and liberal voters have found common ground on the data center issue.

This reflects a more fundamental structural contradiction: a severe mismatch exists between the national or even global benefits of tech capital and the localized burden of infrastructure costs.

Profits created by data centers flow to tech giants and their shareholders, while ordinary communities hosting the facilities bear the costs. Regarding employment, while data centers involve massive investment—construction of a large data center campus can cost billions or even tens of billions of dollars—long-term operational positions after completion are actually quite limited, typically requiring only dozens to hundreds of full-time operations staff. Most processes rely on automated management, including intelligent temperature control systems, remote monitoring platforms, and robotic inspection technologies, dramatically reducing reliance on human labor. The truly high-paying technical positions—such as AI researchers, algorithm engineers, cloud architects—do not land in the remote communities hosting data centers, but concentrate in tech hub cities like Silicon Valley, Seattle, and New York.

This characteristic of "high investment, low employment, localized external costs" makes community resistance naturally universal across the political spectrum. From a conservative perspective, this involves private property rights protection and local autonomy; from a liberal perspective, it concerns environmental justice and community equity. Two vastly different ideological frameworks achieve rare convergence on the demand of "don't build a data center in my backyard."

The Hidden Bill Behind AI Compute Expansion

Who Bears the Physical Cost of Compute

The resistance movement in Pennsylvania actually reveals a commonly overlooked proposition in the entire AI industry boom: compute power doesn't come from nowhere. Every large model training run, every AI application inference call, corresponds to real-world electricity consumption, water resource occupation, and land requisition.

To intuitively understand the scale of this consumption, we can review the leap in AI large model training compute in recent years. GPT-3, released in 2020, had 175 billion parameters and consumed approximately 3,640 PetaFLOP/s-day of compute during training (computing at quadrillions of floating-point operations per second continuously for 3,640 days). While OpenAI hasn't disclosed detailed parameters for GPT-4 in 2023, industry estimates suggest its training compute increased by at least 50 to 100 times compared to GPT-3. This growth follows so-called "Scaling Laws"—the power-law relationship between model performance improvements and training data volume, model parameter count, and compute amount—meaning each generation of significantly improved models requires multiples or even orders of magnitude more compute. According to International Energy Agency (IEA) forecasts, by 2026, global data center electricity consumption could exceed 1,000 terawatt-hours, approaching Japan's total national electricity consumption. This doesn't yet account for potential additional compute demand explosions from emerging application scenarios like AI agents, embodied intelligence, and video generation.

While the industry buzzes about parameter scales, model capabilities, and business valuations, few seriously discuss: who is paying the cost for this massive physical infrastructure? The answer is often local communities that lack voice and cannot participate in revenue sharing.

Regulatory Gaps and Imbalanced Benefit Distribution

Residents' dissatisfaction lies not just in costs themselves, but in the lack of fair benefit distribution mechanisms and effective community participation channels. Many data center projects receive government tax breaks and preferential electricity rates while providing no corresponding compensation or long-term benefit guarantees to host communities.

In fact, an intense "data center recruitment competition" exists among U.S. states. To attract tech giant investments, Virginia, Texas, Ohio, Indiana, and other states have rolled out highly attractive tax reduction packages, including 10 to 25-year property tax exemptions, sales tax waivers, and even direct cash subsidies. Take Loudoun County, Virginia—home to the world's largest data center cluster—while data centers contribute about one-third of the county's property tax revenue, local residents also endure increasingly severe traffic congestion, grid strain, and falling groundwater levels. More commonly, some economically underdeveloped regions introduce data center projects with extremely low barriers and without adequate assessment or community consultation. The "race to the bottom" effect of tax competition causes local governments' actual fiscal revenue to fall far short of expectations, while community external costs remain undiminished. This inequality created by institutional design itself is the deeper cause of sustained protest sentiment.

Three Directions the Tech Industry Must Address

Pennsylvania voters' united resistance is likely just one microcosm of similar conflicts nationwide and globally. From Ireland imposing temporary bans on new data center projects, to Amsterdam suspending data center land approvals, to Singapore once freezing new data center construction—communities and governments worldwide are reassessing the true cost of AI infrastructure expansion. As AI infrastructure continues expanding, tech companies will inevitably face increasing resistance from local communities.

This poses urgent requirements for the entire industry:

• Cost Transparency: Data center operators need to more candidly disclose their actual impact on local grids, water resources, and environment, allowing communities to participate in decisions on an informed basis. This means not only providing environmental impact assessment reports, but regularly publishing actual operational data including electricity consumption, water usage, carbon emissions, and other key metrics, while accepting independent third-party audits.

• Benefit-Sharing Mechanisms: Through reasonable tax contributions, community development funds, or electricity price protection measures, allow local residents to genuinely share in the dividends brought by AI prosperity. Some promising models have already emerged: for example, some European countries require data center operators to connect waste heat to district heating networks, providing low-cost heating to nearby residents; some tech companies have established dedicated education funds supporting STEM education and vocational training programs in data center host communities, attempting to create local value over the long term.

• Sustainable Technology Investment: Increase R&D investment in green energy supply, efficient cooling solutions, and water recycling technologies to fundamentally reduce data centers' external costs. In this direction, industry frontiers have already seen multiple promising technology paths. Microsoft is testing underwater data centers to utilize seawater natural cooling; Google's Finland data center uses seawater cooling, completely avoiding freshwater consumption; Meta and multiple startups are massively advancing immersion cooling technology (fully submerging servers in non-conductive special coolant), which can reduce cooling energy consumption by over 40%. On the energy supply side, Microsoft has signed the largest corporate clean energy procurement agreement in history (over 10 gigawatts) and is collaborating with nuclear energy startups to explore small modular reactors (SMR) powering data centers; Amazon directly acquired a data center campus adjacent to a nuclear power plant to obtain stable zero-carbon baseload power. While these efforts are moving in the right direction, they remain considerably distant from fundamentally solving the environmental costs of AI compute expansion.

Conclusion: AI Development Should Not Come at Community Expense

The plain protest "People are going to get screwed" is a wake-up call for the AI era's breakneck pace. Beneath the halo of technological progress lie unavoidable issues of energy consumption, water resource scarcity, and community equity.

Truly sustainable AI development should not be built on sacrificing ordinary community interests. How to find balance between compute expansion and community welfare will become a long-term issue that tech enterprises and policymakers must seriously confront. This concerns not only tech ethics, but will directly impact the pace of AI infrastructure construction—when more and more communities say "no" to data centers, physical bottlenecks to compute expansion may prove more difficult to overcome than chip shortages.