What training and running AI actually costs the planet — and why the numbers are contested
The researcher's paper landed like a grenade. Training GPT-3, she calculated, produced roughly 552 tonnes of CO2 equivalent — about the same as five average American cars over their entire lifetimes, or 125 round-trip flights from New York to San Francisco.
The AI company disputed the methodology. A different researcher said the number was too low. A third said it was too high. A fourth noted that the electricity grid mix mattered enormously — the same training run on a coal-heavy grid versus a renewables-heavy grid could produce dramatically different emissions. Everyone agreed the number was significant. Nobody agreed on what it was.
AI's energy footprint is real, large, and poorly measured. The three components: Training — the one-time (or periodic) process of building a model from data, extremely energy-intensive for large frontier models. Inference — running the trained model to generate responses, less intensive per query but happening at enormous scale continuously. Infrastructure — the data centers, cooling systems, networking, and hardware manufacturing that support both.
Training a large frontier model consumes between hundreds of thousands and millions of kilowatt-hours of electricity. The exact figures are rarely disclosed. Inference at the scale of major AI services (hundreds of millions of queries per day) accumulates significant energy consumption even at modest per-query cost. Infrastructure — particularly data center cooling and hardware manufacturing — adds further environmental cost that is rarely included in reported figures.
The Transparency Problem
Major AI companies do not publicly disclose the energy consumption of specific training runs or inference operations. The estimates that circulate in research papers use indirect methods — comparing to known hardware power consumption, extrapolating from data center electricity use — with significant uncertainty. The difficulty of measuring AI's environmental impact is itself an accountability problem: you cannot govern what you cannot measure.
Energy is not the only environmental cost. Water: data center cooling consumes enormous quantities — estimates suggest training a large model uses hundreds of thousands of liters of water. In water-stressed regions, data center water consumption competes with agricultural and domestic use. Hardware: manufacturing AI chips (GPUs, TPUs) requires rare earth metals, energy-intensive fabrication, and generates electronic waste when hardware is retired. The full environmental footprint of AI includes supply chains that extend into mining and manufacturing in addition to the electricity used in operation.
Choose a major AI company or service — OpenAI, Google (Gemini), Microsoft (Copilot), Meta (Llama), or Anthropic. Research what environmental disclosures they have made: energy consumption figures, water use, renewable energy commitments, carbon claims. Assess: How transparent are they? What can and can't be verified? What do they not disclose?
Start with: "I want to analyze the environmental disclosures of [company] — here's what I found: [your research]"
Where AI genuinely helps with climate — and where the claims outrun the evidence
The energy company's press release announced that its new AI grid management system had reduced energy waste by 15% — the equivalent of taking 200,000 cars off the road. The claim was in every major newspaper.
A researcher read the technical appendix. The 15% figure was measured against a theoretical baseline of what the grid would have wasted without any optimization. The company had been optimizing its grid with non-AI software for years. Against the actual pre-AI baseline, the improvement was closer to 3%. Both numbers were real. Only one was in the headline.
AI has demonstrated real utility in specific climate-relevant applications. Grid optimization: AI can improve the efficiency of electricity grid management — better matching supply to demand, integrating variable renewable sources, reducing transmission losses. Google's DeepMind application to data center cooling achieved documented efficiency improvements. Climate modeling: AI is accelerating climate science — enabling higher-resolution models, faster simulation of scenarios, and better downscaling of global models to regional predictions. Emissions monitoring: satellite-based AI can detect methane leaks and deforestation at scale and speed that human monitoring cannot match. Materials discovery: AI is accelerating the search for better battery materials, catalysts for carbon capture, and solar cell improvements.
The Rebound Effect
Efficiency improvements enabled by AI may increase total consumption rather than reduce it — a phenomenon called the rebound effect. If AI makes energy use more efficient, it can make energy-intensive activities cheaper and more accessible, leading to more of them. Historical evidence from other efficiency technologies (more fuel-efficient cars leading to more driving, more efficient lighting leading to more lighting) suggests rebound effects are real and can partially or fully offset efficiency gains.
The gap between AI climate potential and AI climate reality is large. Most AI climate applications are in early or demonstration stages — the path from "AI demonstrated X in a pilot" to "AI is reducing global emissions by X" involves deployment challenges, scalability questions, and behavioral responses that pilot studies don't capture. Several specific claims warrant skepticism:
AI will solve climate change: AI is a tool with specific applications, not a general-purpose solution to a multi-decadal political, economic, and technical problem. AI's climate benefits will outweigh its costs: this is possible but not demonstrated — the costs are measurable now while benefits are projected and uncertain. AI-enabled efficiency reduces overall emissions: without accounting for rebound effects and the energy cost of AI deployment itself, efficiency claims may overstate net benefits.
Find a specific AI climate benefit claim — from a company press release, research paper, or news article. Evaluate it: What is the claimed benefit? What baseline is it measured against? Is the benefit demonstrated or projected? What rebound effects might apply? What would need to be true for the claim to hold at scale?
Start with: "I want to evaluate this AI climate claim: [describe the claim and its source] — here's my initial assessment: [your analysis]"
Where AI infrastructure lands — and who bears the environmental costs
The town had welcomed the data center. It had brought construction jobs, a few permanent positions, and tax revenue. What it had also brought — which nobody had mentioned clearly in the approval process — was a water withdrawal permit for 2.5 million gallons per day from a river system already stressed by drought.
The farmers downstream had noticed the river level dropping. The town's water utility had noticed pressure changes. The data center's website said the company was committed to water neutrality by 2030. The commitment was for 2030. The water was flowing now.
Data center location decisions are driven by: cheap land, reliable electricity, available water for cooling, fiber connectivity, and favorable tax and regulatory environments. These factors pull data centers toward specific geographies — rural areas with cheap land, areas near cheap electricity sources (hydroelectric in the Pacific Northwest, natural gas in Texas), and jurisdictions offering tax incentives.
The environmental impacts of data centers are localized. Electricity consumption shows up in grid demand and emissions from the source generation. Water consumption affects local watersheds. Land use affects local ecosystems. Heat rejection affects local microclimates. These local impacts often fall on communities that had little voice in the siting decision and receive few of the economic benefits.
The Carbon Grid Lottery
The carbon intensity of AI computation depends heavily on where it runs and when. The same AI inference run on a data center powered primarily by coal produces dramatically more emissions than one powered primarily by renewables. Data centers that can schedule flexible workloads to run during periods of high renewable generation — and avoid running during peak fossil fuel periods — can significantly reduce their carbon footprint. Whether AI companies optimize for this is a governance and incentive question as much as a technical one.
Environmental justice concerns arise when the burdens of environmental harm fall disproportionately on communities with less political and economic power. Data center siting follows this pattern in several ways: rural communities with fewer resources to negotiate siting terms; communities near cheap (often fossil fuel) electricity that already bear pollution burdens; water-stressed communities in arid regions where data center cooling competes with agricultural and domestic use.
The communities hosting data center infrastructure rarely use the AI services those centers power. A rural Virginia county hosting a major data center cluster bears the land use, water, and electricity infrastructure impacts — while the AI services run there are consumed primarily in major urban markets. The extraction of environmental value from a community for consumption elsewhere is a classic environmental justice pattern, newly applied to digital infrastructure.
Choose a real data center cluster or major AI company's infrastructure footprint — Northern Virginia's data center corridor, Google's data centers in the Dalles (Oregon), Microsoft's data centers in Iowa, or Meta's infrastructure in the Southwest. Analyze: What are the local environmental impacts? What community impacts have been documented? What governance mechanisms exist, and are they adequate?
Start with: "I want to analyze the environmental impacts of [data center location/company] — here's what I know: [your description]"
What disclosure, accountability, and environmental governance of AI could look like
The EU's AI Act required documentation of many things. Energy consumption for high-impact AI systems was not among them — at least not with any meaningful specificity. The GPAI systemic risk provisions required reporting energy consumption for models above a certain compute threshold, but the methodology for measurement was not specified.
The AI company filed its energy report. It listed a number. The number was not independently verifiable. There was no standard for what to include or how to measure it. The report satisfied the legal requirement. The legal requirement had not been designed to produce accountability. It had been designed to produce a report.
Current environmental governance of AI is thin. The EU AI Act includes energy reporting requirements for high-impact AI systems, but without standardized methodology. The SEC's climate disclosure rules (contested and evolving) would require disclosure of climate-related risks from AI companies. Some jurisdictions have data center-specific water and energy requirements for permitting. Corporate sustainability commitments — renewable energy pledges, net zero targets — are voluntary and vary in rigor and verifiability.
What doesn't exist: standardized, methodology-specified energy reporting requirements for AI training and inference. Mandatory lifecycle assessment of AI hardware. Binding community impact requirements for data center siting. Cooling water consumption reporting tied to watershed impact assessment. The governance gap is large relative to the scale of environmental impact.
The Greenwashing Problem
AI companies' environmental commitments are often expressed in terms that sound ambitious but are difficult to verify: "carbon neutral by 2030," "water positive by 2030," "100% renewable energy." These commitments frequently rely on purchasing renewable energy certificates rather than actually running on renewable electricity, and on carbon offsets whose quality is variable. Without standardized reporting and independent verification, these commitments function as marketing rather than accountability.
Effective environmental governance of AI requires several components currently absent. Standardized measurement: methodology-specified requirements for measuring and reporting training energy, inference energy, water consumption, and hardware lifecycle impact — comparable to how financial reporting requires standard accounting methods. Independent verification: third-party verification of reported figures, similar to financial audit requirements. Scope completeness: reporting that includes full scope — not just direct electricity consumption but grid emissions factors, water consumption, and hardware supply chain impacts. Community rights: meaningful participation requirements for communities hosting data center infrastructure, with legally enforceable commitments on water and land impacts.
The Dual Challenge
AI's relationship to the environment presents a genuine dual challenge: AI is simultaneously contributing to environmental harm through its resource consumption and potentially contributing to environmental solutions through climate applications. Effective governance needs to address both — reducing AI's environmental footprint while enabling AI climate applications. These are not contradictory goals, but achieving both requires governance that currently doesn't exist.
Design a specific environmental governance requirement for AI — an energy reporting standard, a water disclosure rule, a data center siting requirement, or a hardware lifecycle assessment requirement. Specify: What must be reported or done? How is it measured? Who verifies it? What are the consequences for non-compliance? Who enforces it?
Start with: "I want to design a [type of requirement] for AI environmental governance — here's my proposed standard: [your framework]"
15 questions. Complete all to finish the module.