Data Center Energy: AI's Hidden Environmental Cost

AI data centers are on track to consume 4.5% of US electricity by 2030. The industry's response? Lobby for nuclear power, buy shuttered plants, and fight local opposition.

AI's explosive growth is driving an unprecedented surge in energy demand. Data centers could consume 4.5% of US electricity by 2030, prompting tech companies to pursue nuclear deals, face local backlash, and lobby for permitting reform.

In September 2024, Microsoft announced a deal that would have seemed absurd just five years earlier: the company signed a 20-year power purchase agreement to restart a reactor at Three Mile Island — the Pennsylvania nuclear plant synonymous with America's worst nuclear accident. The reactor, Three Mile Island Unit 1 (which was not involved in the 1979 partial meltdown; that was Unit 2), had been shut down in 2019 because it couldn't compete economically with cheap natural gas. Microsoft's deal with Constellation Energy would bring it back online by 2028, with all 835 megawatts of carbon-free power dedicated to feeding Microsoft's AI data centers.

The Three Mile Island deal wasn't an outlier. It was a signal — the loudest yet — that AI's voracious appetite for electricity is fundamentally reshaping America's energy landscape. And the numbers are staggering.

The Scale of the Problem

According to the US Energy Information Administration and multiple industry analyses, AI data centers are on track to consume approximately 4.5% of total US electricity generation by 2030 — up from roughly 2.5% for all data centers (AI and non-AI) in 2023. The International Energy Agency projects that global data center electricity consumption could double between 2022 and 2026, reaching over 1,000 terawatt-hours annually — roughly equivalent to Japan's entire electricity consumption.

What's driving this explosion? AI workloads are fundamentally different from traditional data center operations:

Training runs for frontier AI models like GPT-5, Gemini Ultra, and Claude require thousands of GPUs running continuously for months. A single training run for a state-of-the-art model can consume 50-100 gigawatt-hours of electricity — enough to power 5,000-10,000 average American homes for a year.
Inference — the process of running trained models to generate responses — is less intensive per query but operates at massive scale. Every ChatGPT conversation, every Copilot suggestion, every AI-generated image draws power. OpenAI alone reportedly handles hundreds of millions of queries per day.
GPU density is rising rapidly. NVIDIA's latest AI chips (H100, B200) consume 700-1,000 watts each. A single rack of AI servers can draw 100-120 kilowatts, compared to 10-20 kilowatts for traditional server racks. This means AI data centers need 5-10x more power per square foot than conventional facilities.
Cooling demands scale with power consumption. AI chips generate enormous heat, requiring advanced cooling systems — including liquid cooling and, increasingly, water-intensive evaporative cooling — that add to both energy and water consumption.

The Stargate Project and the Infrastructure Arms Race

In January 2025, President Trump announced the Stargate project — a joint venture between OpenAI, SoftBank, and Oracle promising up to $500 billion in AI infrastructure investment over four years, with an initial commitment of $100 billion. The project aims to build a network of massive AI data centers across the United States, starting with a facility in Abilene, Texas.

The Stargate project's energy implications are enormous. Industry analysts estimate the full buildout would require 5-10 gigawatts of dedicated power capacity — equivalent to several large nuclear power plants. Securing this power in a US grid that's already facing capacity constraints in many regions is one of the project's greatest challenges.

Stargate is the largest single AI infrastructure commitment, but it's far from the only one:

Microsoft has committed to spending over $80 billion on AI data centers in fiscal year 2025 alone, with facilities planned or under construction across the US, Europe, and Asia.
Google announced $30 billion in capital expenditure for 2024, with a significant portion dedicated to AI infrastructure. The company is building data centers in multiple US states and international locations.
Amazon Web Services committed $150 billion over 15 years to data center expansion, with major facilities in Virginia, Ohio, Oregon, and Mississippi.
Meta plans to spend $60-65 billion on capital expenditure in 2025, primarily for AI infrastructure including a massive 2-gigawatt data center campus.

The Nuclear Pivot

The sheer scale of AI energy demand has driven tech companies toward an energy source that, until recently, seemed politically toxic: nuclear power. Microsoft's Three Mile Island deal was the first domino. Others followed quickly:

Amazon purchased a nuclear-powered data center campus adjacent to the Susquehanna Steam Electric Station in Pennsylvania for $650 million. The deal gives Amazon access to 960 megawatts of nuclear-generated electricity — enough to power a massive AI computing operation with carbon-free energy. Amazon also invested in small modular reactor (SMR) company X-energy.
Google signed a deal with Kairos Power in October 2024 to purchase electricity from small modular nuclear reactors, with the first reactor expected online by 2030. The company committed to purchasing 500 megawatts of SMR capacity.
Oracle chairman Larry Ellison disclosed plans to power a 1-gigawatt data center with three small modular reactors, calling nuclear the only viable option for powering AI at scale.

The nuclear pivot has created a remarkable political realignment. Tech companies that once championed solar and wind are now lobbying for nuclear permitting reform, streamlined NRC licensing, and federal support for SMR deployment. This lobbying is intensifying — OpenAI's federal disclosures reference the Data Center Energy Efficiency Act (H.R. 4425), and multiple AI companies are pushing for provisions in energy legislation that would accelerate nuclear construction timelines.

Google's Emissions Spike

In July 2024, Google released its annual environmental report with a striking admission: the company's greenhouse gas emissions had increased 48% since 2019, with the increase driven primarily by AI-related data center expansion. Total emissions reached 14.3 million metric tons of CO2 equivalent in 2023 — blowing past the company's goal of achieving net-zero emissions by 2030.

Google's report was notable for its candor. The company acknowledged that "as we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute." In other words: Google's AI ambitions and its climate commitments are on a collision course, and AI is winning.

The admission undercut years of corporate sustainability messaging. Google, along with Microsoft, Amazon, and Meta, had made aggressive net-zero and carbon-negative pledges. The AI energy surge has made those pledges increasingly difficult to honor. Microsoft's emissions rose 29% since 2020, also largely attributable to data center expansion.

State and Local Pushback

The data center building boom has triggered fierce opposition at the state and local level, particularly in communities bearing the environmental burden:

Northern Virginia — home to the world's largest concentration of data centers, hosting roughly 70% of the world's internet traffic through "Data Center Alley" in Loudoun County — has become ground zero for the backlash. Residents have organized against new facilities, citing noise pollution from backup generators, visual blight, strain on the electrical grid, and declining property values. In 2025, Loudoun County imposed new restrictions on data center zoning, and Virginia legislators introduced moratorium proposals to pause new construction pending environmental review.

Water consumption has emerged as a particularly contentious issue:

In Oregon, Google's data centers in The Dalles consumed over 1.4 billion gallons of water in 2023 for cooling — a significant draw in a state increasingly prone to drought. Local activists have pushed for mandatory water use disclosures and consumption caps.
In Iowa, Microsoft's data center cluster in West Des Moines consumed approximately 11.5 million gallons of water in a single month during peak cooling demand, according to city records obtained by the Associated Press.
In Louisiana, Meta's planned data center near New Orleans drew controversy over water rights, with environmental groups and local officials questioning whether diverting water resources for tech industry cooling was appropriate in a state facing its own water management challenges. The controversy highlighted the tension between economic development (Meta promised billions in investment and hundreds of jobs) and environmental stewardship.

The Lobbying Response

Tech companies have responded to the energy and environmental challenge with an aggressive lobbying campaign focused on several priorities:

Permitting reform: AI companies and their allies are pushing for streamlined environmental review processes for data centers and the energy infrastructure that feeds them. This puts them in alliance with traditional energy companies and in opposition to environmental groups who argue that NEPA (National Environmental Policy Act) reviews serve a critical function. The debate echoes broader permitting reform battles in Congress, but AI has added urgency and corporate dollars to the push.

Nuclear advocacy: As detailed above, the nuclear pivot has made AI companies some of the most prominent advocates for nuclear energy in Washington. This includes lobbying for the ADVANCE Act (signed into law in July 2024), which streamlined NRC licensing for advanced nuclear reactors, and for additional legislation to support SMR deployment and nuclear fuel supply chains.

Renewable energy credits: Companies are purchasing massive quantities of renewable energy credits (RECs) and signing power purchase agreements (PPAs) for wind and solar to offset their carbon footprint on paper — but critics note that these purchases don't reduce the actual grid demand created by data centers. A REC purchased in Texas doesn't power a data center in Virginia.

Federal preemption: Some industry lobbyists are pushing for federal standards that would override state and local restrictions on data center construction — arguing that a patchwork of local regulations threatens national AI competitiveness. This is fiercely opposed by local governments that argue land use decisions should remain under local control.

Environmental Groups Push Back

Environmental organizations are increasingly alarmed by AI's energy trajectory. The Sierra Club, Natural Resources Defense Council, and other groups have begun focusing on data center energy consumption as a major climate concern. Their arguments include:

AI energy demand is consuming renewable energy capacity that would otherwise displace fossil fuels on the grid — a phenomenon called "additionality displacement"
Nuclear restart deals like Three Mile Island extend the life of aging infrastructure rather than building truly new clean energy
Corporate net-zero pledges are meaningless if absolute emissions keep rising
Water consumption in drought-prone regions represents an unconscionable misallocation of a scarce public resource for private profit
The beneficiaries of AI (primarily large tech companies and their investors) are not the communities bearing the environmental costs

The tech industry counters that AI itself will help solve climate change — through more efficient energy grid management, better climate modeling, accelerated materials science research, and optimized industrial processes. This argument has some merit but also conveniently positions unlimited AI growth as both the problem and the solution.

The Policy Gap

What's striking about the data center energy debate is the near-total absence of federal policy specifically addressing it. There is no federal requirement for data centers to:

Disclose their energy consumption or carbon emissions
Meet energy efficiency standards
Report water usage
Conduct environmental impact assessments
Offset their grid impact

The Data Center Energy Efficiency Act (H.R. 4425), which OpenAI has lobbied on, would begin to address some of these gaps by establishing efficiency benchmarks and reporting requirements. But the bill has not advanced significantly, and the current political environment — with both parties eager to support AI competitiveness — makes stringent regulation unlikely in the near term.

Meanwhile, the International Energy Agency has called for urgent policy action, warning that without efficiency improvements and clean energy deployment, data center energy demand could become a significant obstacle to global climate goals. The EU is moving faster, with the Energy Efficiency Directive requiring data centers above 500 kW to report energy performance indicators starting in 2024.

Looking Ahead

The energy debate will intensify as AI capabilities — and the compute required to power them — continue scaling. Several developments to watch:

Efficiency breakthroughs: NVIDIA's next-generation chips promise better performance per watt, and algorithmic improvements (like model distillation and mixture-of-experts architectures) can reduce inference costs. But historically, efficiency gains have been overwhelmed by demand growth — a pattern known as Jevons Paradox.
SMR timelines: The small modular reactors that tech companies are betting on are still years from commercial deployment. NuScale Power, the most advanced US SMR project, canceled its first planned facility in 2023 due to cost overruns. The gap between nuclear ambitions and nuclear reality could leave AI companies dependent on fossil fuel generation for years.
Grid reliability concerns: Grid operators including PJM Interconnection (which manages the grid across 13 states including data-center-heavy Virginia) have warned that rapid data center growth threatens grid reliability. PJM has implemented a queue reform to manage the flood of interconnection requests from data center developers.
Geopolitical competition: The US, China, and EU are all racing to build AI infrastructure. Environmental constraints that slow US data center deployment could shift AI development to countries with weaker environmental standards — a "carbon leakage" argument that tech companies deploy effectively in Washington.

The inconvenient truth is that the AI revolution runs on electricity, and electricity has consequences — for the climate, for local communities, for water resources, and for the grid that powers everything else in modern life. How we manage those consequences will be one of the defining policy challenges of the next decade. Track the energy lobbying on our Follow the Money page and the relevant legislation on our Bill Tracker.