The Hidden Cost of Your Chatbot: What's Really Powering the AI Boom

Aug 1
7 min read

The existing Crusoe campus in Abilene – OpenAI/Stargate
An aerial view of the Stargate construction site in Abilene, Texas. (OpenAI)

Every time you ask a chatbot a question, that request travels to a building most of us will never see, a data center packed with tens of thousands of processors working around the clock. I'd always known that AI "costs" energy in some abstract sense. What I didn't fully appreciate until I started digging into recent investigative reporting is just how much, how dirty, and how close to people's homes that cost is actually landing.

This isn't an anti-AI piece. It's a research-led look at a genuine sustainability blind spot: an industry expanding faster than the regulations meant to keep it in check, and what a more circular, accountable version of this infrastructure could actually look like.

In short: AI's computing power comes from physical infrastructure with real environmental costs, energy, water, and emissions, that are often built faster than they're regulated or disclosed.

The Scale Nobody Quite Grasps

Combined Cycle Power Plant
Gas Turbine Power Plant _ Aeroderivative and Heavy-Duty Gas Turbines – DYNF

Take Stargate, one of the largest AI data centers in the world, currently under construction in Abilene, Texas. It will eventually span roughly 4 million square feet across 1,100 acres, larger than Central Park, and its on-site power plant is expected to deliver more than 1.7 gigawatts, enough to power over a million homes for a year, while emitting more than 7.8 million tons of greenhouse gases annually, comparable to about 2 million cars.

Zoom out, and the pattern is global. According to the International Energy Agency, data centers consumed around 415 terawatt-hours of electricity in 2024, and that figure is projected to roughly double by 2030 as AI workloads become the dominant driver of growth.¹ In 2025 alone, electricity demand from AI-focused data centers reportedly surged by 50%.² For scale, the IEA has previously compared this level of demand to the entire electricity consumption of Japan.

In short: Individual AI data centers now rival small power plants in scale, and globally, this sector's electricity use is on track to double within the decade.

The Regulatory Blind Spot

Data Center Rack Cooling
Data Center Rack Cooling
Data Center Fundamentals

Part of why this is possible comes down to how permitting works. Investigative reporting on Stargate found that some of its initial approvals came through a lower-tier permitting category typically used for facilities like auto body shops or dry cleaners, one that requires no public notice or input before construction begins. By the time more appropriate permits are sought, the facility is often already operational, making it far harder to challenge or reverse.

Regulators, for their part, are candid about being outpaced. Advocacy groups working directly with state environmental agencies have described enforcement backlogs stretching a decade, even as new facilities are approved and built in a fraction of that time. In several of the Memphis-area cases, EPA guidance issued in early 2026 explicitly clarified that gas turbines, including ones companies described as "temporary", require permits regardless. That clarification came only after turbines had already been running, sometimes for months, near homes, schools, and churches.

In short: Data centers are frequently being built and powered faster than environmental review, public input, or enforcement can keep pace with.

The People Living Next to It

Behind the permitting debates are real communities. Residents near Stargate in Abilene have described replacing HVAC filters every two weeks due to construction dust, and learning about the facility's fossil-fuel power plans only after the fact, with no public meeting, no advance notice, and in some cases, neighbors asked to sign non-disclosure agreements about what was even being built. Near Memphis, residents living across the street from unpermitted turbines, some managing conditions like severe asthma, have described emptying their homes in anticipation of leaving, unsure of what the long-term air quality impact will be.

Fine particulate matter from gas turbines is well understood to be linked to respiratory and cardiovascular illness. One estimate tied to a comparable facility in Virginia put potential healthcare costs for nearby residents as high as $100 million a year. Data centers also generate localized "heat island" effects, in some cases raising nearby temperatures by as much as 2°C.

In short: The environmental cost of AI infrastructure isn't abstract, it's landing directly on specific communities, often without their knowledge or consent, and often in neighborhoods already facing health disparities.

Why Efficiency Gains Alone Won't Fix This

Here's a genuinely important, often-overlooked wrinkle: even as AI chips and cooling systems get more efficient, total energy demand keeps climbing. This is a real phenomenon with a name, the Jevons Paradox, first described by economist William Stanley Jevons in the 1860s in relation to coal. Jevons observed that making coal-burning more efficient didn't reduce coal consumption; it made coal power cheaper and more attractive, which increased overall demand.

The same logic applies to AI hardware today. Manufacturers are making real strides, for instance, newer UPS (uninterruptible power supply) systems using silicon carbide components can now hit efficiency levels above 98%, meaningfully cutting the energy wasted as heat during power conversion.⁶ But if usage keeps growing faster than efficiency improves, and by most current measures, it is, those gains get absorbed rather than translated into lower overall consumption.

In short: Efficiency improvements are real, but on their own they won't offset AI's rapidly growing energy footprint, a pattern with a 160-year-old precedent in the Jevons Paradox.

What Real Solutions Look Like

It's not all bad news, there's a genuine, research-backed playbook for making data centers more sustainable, and parts of it are already being implemented:

Measure before you change anything. Tracking electrical usage, temperature, and server load is step one, it identifies which workloads can be consolidated onto fewer, more efficient machines.
Physical layout matters. Simple hot-and-cold aisle configurations, which group equipment by temperature and airflow needs, can meaningfully cut HVAC energy use without new hardware.
Replace legacy hardware. Older servers and switches generate more heat and waste more energy per unit of computing than modern, high-efficiency equipment, and virtualization can reduce the number of physical machines needed altogether.
Higher-efficiency power conversion. Newer UPS systems using silicon carbide semiconductors can maintain efficiency above 97% across a wide range of load levels, directly cutting the energy lost, and the cooling load created, during power conversion.
Source renewable power where it's genuinely available, and track vendor progress using independent frameworks like CDP or Sustainalytics rather than taking self-reported claims at face value.
Offset what can't yet be eliminated — while treating offsets as a last resort, not a substitute for genuine reduction.

In short: Concrete, already-available engineering and operational changes — better monitoring, smarter layouts, modern hardware, and independently verified renewable sourcing — can meaningfully cut a data center's footprint without waiting for a breakthrough technology.

Where Circular Economy Thinking Actually Fits

A few areas stand out as genuine opportunities to close loops rather than just reduce waste at the margins:

Waste heat reuse. Data centers reject enormous amounts of low-grade heat. Several European facilities already feed this heat into district heating networks, a model far more developed in the EU than the US, and one that fits neatly within existing German and Nordic district heating infrastructure.
Closed-loop water cooling. Systems like Stargate's already recirculate water rather than consuming it continuously, but broader adoption of closed-loop and even air-cooled designs could meaningfully reduce the sector's water footprint, especially in drought-prone regions, a concern directly comparable to the water-efficiency lessons from the Almería greenhouse case study.
Component-level material recovery. Semiconductor materials like silicon carbide require significantly more energy to manufacture than standard silicon, meaning end-of-life recovery and reuse of these components matters more, not less, as adoption grows.
Transparency as infrastructure. Arguably the most "circular" fix available is procedural: public permitting, disclosed emissions data, and community notice before construction, closing the information loop between operators, regulators, and the people living nearest to these facilities.

In short: Beyond hardware efficiency, genuine circular economy gains sit in reusing waste heat, closing water loops, recovering high-footprint materials, and building transparency into the permitting process itself.

How Could This Be Implemented in Practice?

For a European, research-led audience specifically, a few paths look genuinely actionable:

EU-style efficiency mandates, similar to the Energy Efficiency Directive's data center reporting requirements, could be a template for other regions currently operating with far less disclosure.
Mandatory waste heat recovery agreements for any new large-scale data center, tying construction approval to a district heating integration plan wherever feasible.
Independent, pre-construction environmental review as a non-negotiable step, not a lower-tier permit category designed for auto shops, for any facility above a defined power threshold.
Public disclosure requirements for water and energy usage, modeled on existing corporate sustainability reporting frameworks, rather than voluntary hyperscaler pledges.

In short: Several of the EU's existing efficiency and disclosure frameworks are already close to a workable template, the main gap elsewhere is political will to apply them before construction, not after.

Key Takeaways

The scale is genuinely enormous. Individual AI data centers now rival power plants, and the sector's total electricity demand is on track to double globally by 2030.
"Renewable" promises and current construction don't match. Most near-term AI infrastructure expansion is being powered by on-site gas and diesel generation, not clean energy.
Regulation is structurally behind the industry's pace. Permitting loopholes and enforcement backlogs mean facilities are often operational long before proper environmental review catches up.
Real people bear the immediate cost. Communities near these sites face air quality, noise, and water impacts, frequently without advance notice or a say in the process.
Efficiency alone won't solve this the Jevons Paradox is real. Meaningful progress requires efficiency gains combined with transparency, better siting, and genuine circular design, not efficiency improvements alone.
Concrete circular solutions already exist. Waste heat reuse, closed-loop water cooling, and material recovery are proven approaches that could be mandated rather than left voluntary.

References

International Energy Agency (2025). Energy and AI / IEA Electricity 2026 reporting on global data center electricity consumption trends, 2024–2030.
International Energy Agency (2026). Key Questions on Energy and AI — Executive Summary.
Southern Environmental Law Center (2026). xAI Built an Illegal Power Plant to Power Its Data Center; Earthjustice (2026), Illegal Pollution from Data Center Power Plants.
U.S. Senate Committee on Environment and Public Works (2026). Whitehouse Calls for Answers About Musk-Backed xAI's Pattern of Operating Illegal Data Center Gas Plants.
TechTimes (2026). xAI Ran 59 Unpermitted Gas Turbines in Black Communities, DOJ Now Shields Them.
Riello UPS (2026). Technical presentation, Making Data Centres More Sustainable — on silicon carbide UPS efficiency data.
Floodlight / KVUE investigative reporting (2026). We Saw What AI Data Centers Don't Want You to See.
Data Center Frontier (2024). IEA Study Sees AI, Cryptocurrency Doubling Data Center Energy Consumption by 2026.