> "We do not inherit the earth from our ancestors; we borrow it from our children."
Learning Objectives
- Quantify AI's environmental footprint, including energy consumption, water use, and hardware lifecycle impacts
- Identify applications of AI for environmental benefit, including climate modeling and energy optimization
- Evaluate the trade-off between AI capability and environmental cost
- Analyze 'green AI' initiatives and assess their effectiveness
- Assess the environmental implications of scaling AI systems and the rebound effect
In This Chapter
- Chapter Overview
- 18.1 The Carbon Cost of Intelligence: Training, Inference, and Data Centers
- 18.2 Water, Minerals, and E-Waste: The Hidden Environmental Costs
- 18.3 AI for Climate: Modeling, Monitoring, and Mitigation
- 18.4 AI for Energy: Smart Grids and Optimization
- 18.5 Green AI: Making Models More Efficient
- 18.6 The Rebound Effect: When Efficiency Doesn't Save
- 18.7 Chapter Summary
- Optional Python Exercise: Simple Carbon Footprint Estimator
- Spaced Review
- 🎯 Project Checkpoint: AI Audit Report — Step 18
- What's Next
"We do not inherit the earth from our ancestors; we borrow it from our children." — Attributed to various Indigenous leaders
Chapter Overview
Here is a number that should stop you cold: training a single large language model can emit as much carbon dioxide as five cars produce over their entire lifetimes — manufacturing included.
And here is another number that should make you pause: AI-powered climate models have helped scientists predict extreme weather events with accuracy that was impossible a decade ago, potentially saving thousands of lives and billions of dollars in damage.
Both of these statements are true. And the tension between them is the central challenge of this chapter.
Artificial intelligence has a relationship with the environment that is genuinely paradoxical. On one hand, AI systems consume enormous amounts of energy, water, and raw materials. The data centers that power them are among the fastest-growing sources of electricity demand on the planet. The hardware they run on requires mining rare minerals, and when it is discarded, it becomes toxic electronic waste. On the other hand, AI is one of the most powerful tools available for understanding and fighting climate change — optimizing energy grids, monitoring deforestation, predicting natural disasters, and designing more efficient systems of every kind.
This chapter asks you to hold both realities in your head at the same time. We are not here to declare AI "good" or "bad" for the environment. We are here to give you the knowledge you need to evaluate the trade-offs for yourself — and, through your AI Audit Report, to apply that analysis to a real system.
If the recurring theme of this book is "who benefits and who is harmed," this chapter extends that question to the planet itself.
In this chapter you will learn to:
- Quantify AI's environmental footprint — energy, water, hardware, and carbon
- Identify applications of AI for environmental benefit
- Evaluate the trade-off between AI capability and environmental cost
- Analyze "green AI" initiatives and their effectiveness
- Assess the environmental implications of scaling AI systems
Learning Paths
Fast Track (50 minutes): Read sections 18.1, 18.3, 18.6, and 18.7. Complete the Check Your Understanding prompts and the Project Checkpoint.
Deep Dive (2.5–3.5 hours): Read all sections, work through the carbon footprint estimation exercise, complete the Check Your Understanding prompts, read both case studies, and add the environmental analysis to your AI Audit Report.
18.1 The Carbon Cost of Intelligence: Training, Inference, and Data Centers
In Chapter 3, you learned that machine learning models "learn" by processing vast quantities of data, adjusting their internal parameters through millions or billions of iterative calculations. What we did not dwell on in that chapter is what all of those calculations cost in physical terms.
Every computation requires electricity. Every electron flowing through a processor generates heat. Every data center housing those processors needs cooling. And most of that electricity, in most of the world, still comes from burning fossil fuels.
Training: The Headline Number
The environmental cost of AI first entered public consciousness through a 2019 paper by researchers at the University of Massachusetts Amherst. They estimated that training a single large natural language processing model (a Transformer with neural architecture search) produced approximately 284 metric tons of CO2 — roughly five times the lifetime emissions of an average American car, including manufacturing.
That number was shocking, and it was widely reported. It is also important to understand what it does and does not mean.
What it does mean: training large AI models is computationally expensive, and that expense has real environmental costs. As models have grown — from millions of parameters (2018) to hundreds of billions (2023) to over a trillion (2024–2025) — the energy requirements have grown accordingly. Training a model like GPT-4 is estimated to have consumed thousands of megawatt-hours of electricity — enough to power hundreds of homes for a year.
What it does not mean: not every AI model is this expensive. The headline-grabbing numbers come from the very largest models — the ones trained by a handful of well-funded companies pushing the frontier of capability. A small machine learning model trained on a laptop for a class project has a negligible carbon footprint. The environmental cost of AI varies enormously depending on the model's size, architecture, training duration, and the energy source powering the hardware.
💡 Intuition: Think of AI training like driving. A quick trip to the grocery store uses a trivial amount of fuel. But driving a semi-truck across the country, fully loaded, burns an enormous amount. Most AI models are the grocery trip. The models that make headlines are the transcontinental haul. Both are "driving," but the environmental impact differs by orders of magnitude.
Inference: The Quiet Majority
Here is something that surprises most people when they first learn it: inference — the process of using a trained model to make predictions — often consumes more total energy than training.
Training happens once (or a few times). Inference happens every time someone asks ChatGPT a question, every time a recommendation algorithm serves a video, every time a voice assistant processes a command. A single inference query uses very little energy. But multiplied across billions of daily queries, the aggregate energy consumption is staggering.
One widely cited estimate suggests that a single ChatGPT query consumes roughly 10 times the energy of a standard Google search. If generative AI tools continue to grow in popularity — and every indication suggests they will — inference energy will become the dominant component of AI's environmental footprint.
📊 Evidence Evaluation: Interpreting Energy Claims
When you encounter claims about AI energy consumption, ask:
- Is this about training or inference? They have very different profiles.
- Is this a one-time cost or an ongoing cost? Training is one-time; inference is continuous.
- What is the comparison? "10x a Google search" sounds alarming, but a Google search uses very little energy. What matters is the aggregate — billions of queries multiplied by the per-query cost.
- What energy source is assumed? The same computation produces very different emissions depending on whether the data center runs on coal, natural gas, or renewable energy.
- Is the number specific or extrapolated? Some widely shared estimates are rough extrapolations, not direct measurements. Precision matters.
Data Centers: The Physical Infrastructure
All of this computation happens in data centers — large facilities filled with thousands of servers, networking equipment, and cooling systems. The world's major technology companies operate hundreds of data centers globally, and AI workloads are a rapidly growing share of their energy consumption.
A few numbers to put this in perspective:
- Data centers globally consumed approximately 1–1.5% of the world's electricity in 2023, according to the International Energy Agency (IEA). AI-specific workloads are a growing fraction of that total.
- The IEA projected that data center electricity consumption could double by 2026, driven in significant part by AI demand.
- Individual data centers operated by major cloud providers can consume as much electricity as a small city — 100+ megawatts in some cases.
The metric used to evaluate data center energy efficiency is power usage effectiveness (PUE) — the ratio of total facility energy to the energy used for actual computing. A PUE of 1.0 would mean all energy goes to computation; a PUE of 2.0 means half the energy is used for cooling, lighting, and other overhead. State-of-the-art data centers achieve PUE values around 1.1–1.2. Older facilities may have PUE values of 1.5 or higher.
🔄 Check Your Understanding: Explain the difference between the energy cost of training an AI model and the energy cost of inference. Which one is likely to grow faster as AI usage increases, and why?
18.2 Water, Minerals, and E-Waste: The Hidden Environmental Costs
Carbon emissions get the most attention, but AI's environmental footprint extends well beyond electricity and CO2.
Water
Data centers generate enormous amounts of heat, and cooling that heat requires water — lots of it. Cooling towers, which are the most common method, evaporate water to dissipate heat. A single large data center can consume millions of gallons of water per day.
A 2023 study by researchers at the University of California, Riverside estimated that training GPT-3 directly consumed approximately 700,000 liters of fresh water for cooling — enough to produce roughly 370 BMW sedans. When indirect water consumption (the water used to generate the electricity) is included, the total is significantly higher.
This raises acute concerns in regions facing water scarcity. Data centers in the western United States, the Middle East, and other water-stressed areas are competing for water resources with agriculture, residential use, and ecosystems. As AI workloads grow, this competition will intensify.
Minerals and Mining
AI runs on specialized hardware — GPUs (graphics processing units), TPUs (tensor processing units), and other accelerators designed for parallel computation. Manufacturing this hardware requires rare earth elements and other minerals: lithium, cobalt, tantalum, neodymium, and others.
The extraction of these minerals carries significant environmental and human costs:
- Environmental damage: Mining disrupts ecosystems, contaminates water supplies, and produces toxic waste. Rare earth mining in China's Inner Mongolia region, which produces a significant share of global supply, has created large toxic waste ponds and contaminated surrounding farmland.
- Human costs: Cobalt mining in the Democratic Republic of Congo — which supplies roughly 70% of the world's cobalt — has been linked to child labor, unsafe working conditions, and community displacement.
- Geopolitical concentration: The supply chains for critical minerals are concentrated in a small number of countries, creating economic dependencies and geopolitical tensions.
E-Waste
AI hardware has a relatively short operational life. GPUs used for cutting-edge AI training are often replaced every two to four years as more powerful chips become available. The discarded hardware becomes electronic waste (e-waste) — and e-waste is one of the fastest-growing waste streams in the world.
The United Nations estimated that the world generated approximately 62 million metric tons of e-waste in 2022, with less than 25% formally recycled. E-waste contains toxic materials — lead, mercury, cadmium, brominated flame retardants — that can leach into soil and water when improperly disposed of. Much of the world's e-waste is shipped to developing countries, where it is processed under unsafe conditions by informal workers, including children.
🌍 Global Perspective: Who Bears the Cost?
The environmental costs of AI are not distributed evenly. The companies and countries that benefit most from AI — primarily in North America, Europe, and East Asia — are not the same as the communities that bear the heaviest environmental burden.
- Data centers consume water in regions that may already face scarcity
- Minerals are mined in countries where environmental and labor protections are weakest
- E-waste disproportionately ends up in the Global South
- Carbon emissions affect everyone, but climate impacts fall hardest on the most vulnerable nations and communities
This is the "who benefits, who is harmed" question applied at a planetary scale.
Life Cycle Assessment
To properly account for AI's environmental impact, researchers use a framework called life cycle assessment (LCA) — a method for evaluating the environmental effects of a product or system across its entire lifespan, from raw material extraction through manufacturing, use, and disposal.
For AI systems, a comprehensive life cycle assessment would include:
- Raw materials: Mining and processing of minerals for hardware
- Manufacturing: Energy and emissions from chip fabrication (semiconductor fabs are among the most energy- and water-intensive factories in the world)
- Transportation: Shipping hardware to data centers worldwide
- Operation: Electricity for computation and cooling; water for cooling
- End of life: Recycling, disposal, or (often) informal processing of e-waste
The carbon emissions from hardware manufacturing — called embodied carbon — can be substantial. For some AI workloads, embodied carbon accounts for a significant fraction of total lifecycle emissions, challenging the assumption that the "use phase" is always the dominant source.
🔄 Check Your Understanding: List three environmental costs of AI besides electricity consumption and carbon emissions. For each, explain who is most affected.
18.3 AI for Climate: Modeling, Monitoring, and Mitigation
Now for the other side of the ledger. AI is not only an environmental burden — it is also one of the most powerful tools available for understanding and addressing climate change.
Climate Modeling
Climate models are computer simulations of the Earth's climate system — atmosphere, oceans, land surface, ice, and their interactions. These models are essential for predicting future climate conditions, evaluating policy scenarios, and understanding how human activities affect the global climate.
Traditional climate models are extraordinarily complex but computationally limited. They divide the Earth's surface into grid cells and simulate physical processes within each cell. The finer the grid (and thus the more detailed the simulation), the more computation is required. State-of-the-art climate models can take weeks to run on the world's most powerful supercomputers.
AI is being used to accelerate and improve climate modeling in several ways:
- Emulators: Machine learning models trained to approximate the outputs of traditional climate simulations — producing similar results in a fraction of the time. This allows scientists to run many more scenarios and explore a wider range of possibilities.
- Downscaling: AI can take coarse-resolution global predictions and produce fine-grained local forecasts — predicting how climate change will affect a specific city or watershed, not just a broad region.
- Extreme event prediction: Neural networks trained on historical weather data have shown promising accuracy in predicting extreme events — hurricanes, heat waves, flooding — days or weeks in advance, potentially enabling earlier evacuation warnings and better emergency response.
In 2023, Google DeepMind's GenCast model demonstrated the ability to produce medium-range weather forecasts that outperformed the leading traditional model (the European Centre for Medium-Range Weather Forecasts' operational system) on a majority of metrics. This was a significant milestone, though researchers caution that AI weather models and traditional physics-based models have complementary strengths.
Environmental Monitoring
AI is also being used to monitor environmental conditions at scales and speeds that would be impossible for humans alone:
- Satellite imagery analysis: Computer vision systems can detect deforestation, track glacier retreat, monitor urban sprawl, and identify illegal mining from satellite images — processing thousands of images daily that would take human analysts months to review.
- Biodiversity monitoring: AI models analyze audio recordings from forests and oceans to identify species by their sounds — tracking bird populations, whale migration patterns, and the health of coral reefs.
- Pollution detection: Machine learning systems can identify sources of methane emissions from satellite data, helping regulators and companies locate and fix leaks. Methane is a potent greenhouse gas, and leak detection is one of the most cost-effective ways to reduce emissions in the near term.
- Agricultural optimization: AI tools help farmers reduce water use, optimize fertilizer application, and predict crop yields — potentially reducing the environmental footprint of agriculture, which accounts for roughly 10% of U.S. greenhouse gas emissions.
📊 Research Spotlight: AI and Deforestation
Global Forest Watch, a monitoring platform developed by the World Resources Institute with support from Google, uses machine learning to analyze satellite imagery and detect forest loss in near real-time. The system processes images from the Landsat satellite constellation and can identify deforestation events within days of their occurrence — dramatically faster than traditional manual analysis.
In Brazil, the DETER system (developed by Brazil's National Institute for Space Research) uses AI to monitor the Amazon rainforest, providing weekly deforestation alerts that inform law enforcement operations against illegal logging. During periods of active enforcement, DETER-informed operations have been credited with significant reductions in deforestation rates.
These systems demonstrate a powerful use case for AI: automating the analysis of data that exists but is too vast for humans to process manually. The environmental benefit is real and measurable. But it depends on political will — monitoring tools only reduce deforestation if the information they produce leads to action.
Carbon Capture and Materials Science
AI is accelerating research into technologies that could help mitigate climate change:
- Carbon capture: Machine learning is being used to identify new materials for capturing CO2 from the atmosphere and from industrial emissions, screening millions of chemical compounds far faster than traditional laboratory methods.
- Battery design: AI tools are helping researchers discover new battery materials that could improve energy storage for renewable power — a critical bottleneck in the transition away from fossil fuels.
- Nuclear fusion: AI is being used to control plasma in experimental fusion reactors, which could eventually provide virtually unlimited clean energy.
These applications are at varying stages of maturity. Some (satellite monitoring, weather forecasting) are operational and producing real results. Others (fusion, novel battery materials) are still in the research phase and may take years or decades to reach deployment.
🔄 Check Your Understanding: Name two ways AI is being used to fight climate change. For each, explain what specific capability of AI (pattern recognition, optimization, processing speed) makes it well-suited to the task.
18.4 AI for Energy: Smart Grids and Optimization
Energy is both AI's biggest environmental cost and one of the areas where AI offers the most environmental benefit. This creates a particularly interesting feedback loop — one that could be virtuous or vicious depending on how it plays out.
Smart Grid Optimization
Electrical grids are astonishingly complex systems. They must continuously balance electricity supply (from power plants, wind farms, solar installations, and other sources) with demand (from homes, businesses, factories, and — increasingly — data centers). Supply and demand must match at every moment; even small imbalances can cause outages.
AI is being used to optimize grid operations in several ways:
- Demand forecasting: Machine learning models predict electricity demand hours or days in advance, allowing grid operators to plan generation more efficiently and reduce the need for polluting "peaker" plants that are fired up to meet unexpected demand spikes.
- Renewable integration: Wind and solar power are intermittent — they produce electricity only when the wind blows or the sun shines. AI helps predict renewable generation and manage the variability, enabling grids to absorb more renewable energy without sacrificing reliability.
- Load balancing: AI systems can shift flexible electricity demand (charging electric vehicles, running industrial processes, heating water) to times when renewable energy is abundant and cheap, reducing waste and emissions.
Google has reported that its DeepMind AI system reduced the energy used for cooling its data centers by approximately 40% — a significant efficiency gain that directly reduces both costs and environmental impact. The company has since made this technology available through its cloud services.
Building Energy Efficiency
Buildings account for roughly 40% of energy consumption in many developed countries. AI is being used to optimize building energy use:
- HVAC optimization: Machine learning systems learn the thermal characteristics of a building and adjust heating, ventilation, and air conditioning to minimize energy use while maintaining comfort. These systems can reduce building energy consumption by 10–30%.
- Lighting and occupancy: AI-controlled lighting systems that adjust based on occupancy and natural light can significantly reduce electricity use.
- Predictive maintenance: AI systems that predict when equipment is likely to fail can prevent inefficient operation and extend equipment life.
Industrial Optimization
Manufacturing and industrial processes are major energy consumers. AI is being applied to:
- Process optimization: Machine learning can identify opportunities to reduce energy use in complex manufacturing processes — for example, optimizing the temperature profiles in industrial furnaces or the chemical reactions in materials processing.
- Supply chain efficiency: AI tools can optimize logistics and transportation, reducing fuel consumption and emissions.
💡 Intuition: Think of AI in energy optimization like a highly skilled building manager — one who never sleeps, monitors every system simultaneously, notices patterns that humans would miss, and makes thousands of micro-adjustments per day. No single adjustment saves much energy, but the cumulative effect can be substantial.
18.5 Green AI: Making Models More Efficient
The term "green AI" emerged from a 2019 paper by researchers at the Allen Institute for AI (AI2), who argued that the field's relentless pursuit of accuracy gains through ever-larger models was environmentally unsustainable and scientifically questionable. They proposed a shift in emphasis: from "Red AI" (pursuing state-of-the-art performance regardless of computational cost) to "Green AI" (achieving good performance with fewer resources).
Efficiency Techniques
Several technical approaches can reduce the environmental footprint of AI:
Model compression and distillation. A large model can be "distilled" into a smaller model that approximates its performance. The smaller model runs faster, uses less energy, and requires less powerful hardware. Knowledge distillation has produced models that retain 95%+ of the original's performance at a fraction of the computational cost.
Pruning. Removing unnecessary connections (weights) from a neural network after training can reduce its size and computational requirements without significantly affecting performance. Research has shown that many large models contain substantial redundancy — often 80–90% of weights can be removed with minimal accuracy loss.
Quantization. Reducing the numerical precision of a model's parameters (for example, from 32-bit to 8-bit or even 4-bit numbers) reduces memory requirements and speeds up computation. Modern hardware is increasingly optimized for low-precision arithmetic.
Efficient architectures. Designing model architectures that achieve strong performance with fewer parameters from the start. Techniques like mixture-of-experts (where only a subset of the model is activated for any given input) can dramatically reduce computation.
Transfer learning and fine-tuning. Instead of training a new model from scratch for every task, start with a pre-trained model and fine-tune it on a smaller dataset for the specific application. This approach is now standard practice and reduces training costs by orders of magnitude compared to training from scratch.
Reporting and Transparency
A growing movement advocates for standardized reporting of the environmental costs of AI research:
- Carbon reporting in papers: Some conferences and journals now encourage (or require) researchers to report the estimated carbon emissions associated with their experiments.
- Efficiency as a metric: Alongside accuracy and speed, some researchers now report "performance per watt" or "performance per ton of CO2" — treating efficiency as a first-class metric.
- Tools for estimation: Projects like the Machine Learning CO2 Impact Calculator and CodeCarbon provide tools for estimating the carbon footprint of training runs.
Comparison Table: Green AI Techniques
Technique How It Works Typical Energy Reduction Trade-off Knowledge distillation Train a small model to mimic a large one 10–100x less inference energy Slight accuracy loss Pruning Remove unnecessary network connections 2–10x faster inference Requires careful tuning Quantization Reduce numerical precision 2–4x less memory and computation Minor accuracy impact Efficient architectures Design smaller models from the start Varies widely Requires research investment Transfer learning Fine-tune pre-trained models instead of training from scratch 100–1000x less training energy Depends on task similarity
MedAssist AI: Efficiency in Healthcare
Consider MedAssist AI, the diagnostic tool from our anchor examples. The hospital deployed a model originally trained on millions of medical images. If MedAssist were retrained from scratch every time the hospital wanted to add a new type of scan or update its performance, the training costs (environmental and financial) would be substantial.
Instead, MedAssist uses transfer learning: the base model was trained once on a large dataset, and the hospital fine-tunes it on its own patient data. This approach reduces training energy by a factor of 100 or more compared to training from scratch. It is also faster and often produces better results on the specific task, because the base model has already learned general visual features.
This is green AI in practice — not a theoretical concept but a standard engineering approach that happens to be both more efficient and more effective.
🔄 Check Your Understanding: Name three techniques for reducing the environmental footprint of AI models. For each, explain the basic mechanism and the main trade-off involved.
18.6 The Rebound Effect: When Efficiency Doesn't Save
Here is a cautionary tale that applies far beyond AI.
In 1865, the English economist William Stanley Jevons observed that improvements in the efficiency of coal-burning steam engines did not reduce total coal consumption. Instead, more efficient engines made coal-powered technology cheaper and more attractive, which led to more use of coal, not less. Total coal consumption increased despite — and in part because of — the efficiency gains.
This phenomenon is called the rebound effect, sometimes known as Jevons paradox. And it is directly relevant to the environmental impact of AI.
How the Rebound Effect Works in AI
Imagine a research lab develops a technique that makes training large language models 50% more energy-efficient. Good news for the environment, right? Perhaps — but consider what happens next:
- Training is now cheaper, so more organizations can afford to train large models
- Companies that were already training models can now train even larger ones within the same budget
- More efficient inference means generative AI can be deployed in more applications and served to more users
- The total energy consumed by AI increases, even though each individual computation is more efficient
This is not hypothetical. Over the past decade, hardware efficiency for AI computation has improved dramatically — each new generation of GPUs is significantly more efficient per operation than the last. Yet total energy consumption by AI has increased even faster, because the growth in AI usage has outpaced the efficiency gains.
💡 Intuition: The rebound effect works like a highway expansion. Building more lanes is supposed to reduce traffic congestion. But wider highways attract more drivers (induced demand), and within a few years, the new lanes are just as congested as the old ones. The efficiency gain was real, but the behavioral response swallowed it.
Is the Rebound Effect Inevitable?
Not necessarily. The rebound effect is a tendency, not a law of nature. Policy interventions can counteract it:
- Carbon pricing: If the carbon cost of computation is priced into the system (through carbon taxes or cap-and-trade schemes), efficiency gains reduce costs but the price signal keeps total consumption in check.
- Efficiency standards: Mandatory minimum efficiency standards for data centers, hardware, and algorithms can ensure that efficiency gains are not entirely consumed by growth.
- Renewable energy mandates: Requiring data centers to run on renewable energy changes the equation: more computation does not necessarily mean more emissions if the power source is clean.
- Intentional restraint: Organizations can choose to use efficiency gains to do the same work with less energy, rather than to do more work with the same energy. This requires conscious decisions, not market defaults.
The Scale Question
The rebound effect raises a deeper question about the relationship between AI and the environment: Is it possible to have ever-growing AI capabilities without ever-growing environmental impact?
The technology industry's answer has generally been: yes, if we transition to renewable energy. Build enough solar panels and wind turbines, and data centers can run on clean power indefinitely. There is truth in this — data centers powered by 100% renewable energy produce zero operational carbon emissions (though embodied carbon, water use, and e-waste remain concerns).
But the transition to renewable energy is not yet complete, and the pace of AI growth is straining even the most ambitious renewable energy plans. Several major technology companies, after years of progress toward carbon neutrality, have reported increasing emissions driven by AI-related data center expansion. The demand for electricity from AI is growing faster than the supply of new renewable capacity.
⚖️ Evidence Evaluation: "Carbon Neutral" Claims
When a technology company claims that its AI operations are "carbon neutral," examine the claim carefully:
- Offsets vs. reductions: Does the company actually reduce emissions, or does it buy carbon offsets (paying someone else to plant trees or avoid deforestation)? Offsets are controversial — some deliver real benefits, others do not.
- Scope: Does the claim cover only operational emissions (Scope 1 and 2 in accounting terms), or does it include the full supply chain (Scope 3), including hardware manufacturing, mining, and e-waste?
- Renewable energy accounting: Does the company actually consume renewable energy, or does it purchase renewable energy certificates (RECs) that allow it to claim renewable energy while the data center actually runs on the local grid mix?
- Timeframe: Does the company match its energy consumption with renewable generation on an hourly basis, or only on an annual average? Annual matching can mean the data center runs on fossil fuels during nights and winters while claiming "100% renewable" on paper.
🔄 Check Your Understanding: Explain the rebound effect (Jevons paradox) in your own words. Give one example of how it could apply to AI energy efficiency improvements. Then describe one policy intervention that could counteract it.
18.7 Chapter Summary
This chapter has asked you to confront a genuine paradox: AI is both an environmental problem and an environmental solution. Let us consolidate what we have learned.
AI has a significant and growing environmental footprint. Training large models consumes substantial electricity and produces meaningful carbon emissions. Inference — using trained models — is individually cheap but collectively enormous due to scale. Data centers consume vast quantities of water for cooling. AI hardware requires mining environmentally destructive minerals and produces toxic e-waste. The costs are disproportionately borne by communities far from the benefits.
AI is also a powerful environmental tool. Climate modeling, weather prediction, environmental monitoring, energy optimization, materials discovery — AI is contributing to environmental solutions in ways that are real, measurable, and growing. Some of these applications (satellite-based deforestation monitoring, smart grid optimization) are already operational and producing results.
The trade-off is real but not fixed. Whether AI's environmental benefits outweigh its costs depends on decisions that humans make: which models to build and how large; where to locate data centers and how to power them; whether to invest in efficiency or pursue ever-larger scale; and whether to price the environmental externalities of computation.
Green AI is a growing movement but not yet the default. Techniques like model compression, pruning, quantization, and transfer learning can dramatically reduce AI's environmental footprint. Standardized carbon reporting is gaining traction. But the competitive incentives in the AI industry still favor scale over efficiency.
The rebound effect means efficiency alone is not enough. Making AI more efficient per computation does not automatically reduce total environmental impact if efficiency gains are consumed by increased usage. Policy interventions — carbon pricing, efficiency standards, renewable energy mandates — are necessary to ensure that efficiency gains translate into environmental gains.
📋 Key Concepts Introduced in This Chapter
Concept Definition Carbon footprint (of AI) The total greenhouse gas emissions associated with developing, training, and deploying AI systems Training vs. inference energy Training is a one-time computational cost; inference is the ongoing cost of using a trained model Power usage effectiveness (PUE) Ratio of total data center energy to computing energy; measures infrastructure efficiency Embodied carbon Carbon emissions from manufacturing hardware, distinct from operational emissions E-waste Discarded electronic equipment, including AI hardware, containing toxic materials Green AI A movement advocating for computationally efficient AI research and deployment Model efficiency Achieving strong AI performance with minimal computational resources Rebound effect (Jevons paradox) When efficiency gains lead to increased total consumption rather than reduced resource use Life cycle assessment (LCA) A method for evaluating environmental impacts across a product's entire lifespan Scope 1/2/3 emissions Categories of carbon emissions: direct (Scope 1), electricity (Scope 2), supply chain (Scope 3)
Optional Python Exercise: Simple Carbon Footprint Estimator
If you are comfortable with basic Python, try this exercise. It is not required — the conceptual content of this chapter does not depend on it.
# Simple AI carbon footprint estimator
# Estimates CO2 emissions from training based on GPU hours and grid carbon intensity
def estimate_carbon(gpu_hours, gpu_power_watts=300, pue=1.2, carbon_intensity_g_per_kwh=400):
"""Estimate CO2 emissions from an AI training run.
Args:
gpu_hours: Total GPU-hours of training
gpu_power_watts: Power consumption per GPU (default: 300W for modern AI GPU)
pue: Power Usage Effectiveness of data center (default: 1.2)
carbon_intensity_g_per_kwh: Grid carbon intensity in grams CO2 per kWh
(default: 400, roughly the global average; varies widely by region)
Returns:
Dictionary with energy and emissions estimates
"""
energy_kwh = (gpu_hours * gpu_power_watts / 1000) * pue
carbon_kg = energy_kwh * carbon_intensity_g_per_kwh / 1000
car_miles_equivalent = carbon_kg / 0.411 # avg US car emits 411g CO2 per mile
return {
"energy_kwh": round(energy_kwh, 1),
"carbon_kg": round(carbon_kg, 1),
"car_miles_equivalent": round(car_miles_equivalent, 0)
}
# Example: Estimate emissions for a small training run (100 GPU-hours)
small_run = estimate_carbon(gpu_hours=100)
print(f"Small training run: {small_run['energy_kwh']} kWh, "
f"{small_run['carbon_kg']} kg CO2, "
f"equivalent to driving {small_run['car_miles_equivalent']} miles")
# Compare: large training run (50,000 GPU-hours) on a clean grid (50 g/kWh)
large_clean = estimate_carbon(gpu_hours=50000, carbon_intensity_g_per_kwh=50)
print(f"Large run (clean grid): {large_clean['carbon_kg']} kg CO2")
# Compare: same large run on a coal-heavy grid (900 g/kWh)
large_dirty = estimate_carbon(gpu_hours=50000, carbon_intensity_g_per_kwh=900)
print(f"Large run (coal grid): {large_dirty['carbon_kg']} kg CO2")
Try it: Modify the parameters to estimate the carbon footprint of different scenarios. How much does the energy source matter? How much does model size (measured in GPU-hours) matter?
Spaced Review
These questions revisit concepts from earlier chapters. Try answering them before checking.
From Chapter 3 (How Machines Learn): When we say a model is "trained," what is actually happening computationally? Why does training require so much more computation for larger models?
Review
Training involves processing data through the model, comparing the model's outputs to the correct answers, calculating the error, and adjusting the model's parameters to reduce that error — repeated millions or billions of times. Larger models have more parameters to adjust, require more data to train effectively, and need more iterations to converge, all of which multiply the computational requirements.From Chapter 5 (Large Language Models): How has the scale of language models changed over the past five years? Why does this matter for environmental impact?
Review
Language models have grown from hundreds of millions of parameters (GPT-2, 2019) to hundreds of billions (GPT-3, 2020) to reportedly over a trillion (recent models). Each order-of-magnitude increase in parameters requires roughly proportional increases in training computation, and often more than proportional increases in data. This means the energy cost of training frontier models has grown dramatically.From Chapter 10 (AI and Work): How do economic incentives shape AI deployment decisions? How might these incentives conflict with environmental sustainability?
Review
Companies adopt AI when it increases productivity, reduces costs, or provides competitive advantages. These economic incentives drive rapid scaling — more models, more users, more applications — which increases environmental impact. The environmental costs are often externalities (not borne by the company or user), so they do not naturally factor into deployment decisions unless policy interventions (carbon pricing, regulation) make them visible.🎯 Project Checkpoint: AI Audit Report — Step 18
Your task: Estimate the environmental footprint of your AI system and propose improvements.
For this chapter, complete the following:
-
Energy profile. Research the computational infrastructure behind your AI system. Is it cloud-hosted? On-premise? What kind of hardware does it likely use? If the system involves a large language model or neural network, estimate (even roughly) the scale of training and inference.
-
Carbon estimate. Using the information in this chapter (and optionally the Python estimator), produce a rough estimate of your system's carbon footprint. Include both training (one-time) and inference (ongoing) components. Note your assumptions and uncertainties — precision is less important than demonstrating the reasoning process.
-
Full lifecycle. Consider the broader environmental footprint: water use for data center cooling, hardware manufacturing (embodied carbon), and end-of-life disposal. Which of these is likely the most significant for your system?
-
Benefit assessment. Does your AI system provide any environmental benefits (directly or indirectly)? If so, describe them. If not, that is a valid finding.
-
Improvement recommendations. Propose at least two specific, actionable recommendations for reducing your system's environmental footprint. Consider: green AI techniques (model compression, efficient architectures), infrastructure choices (renewable energy, efficient data centers), and usage patterns (reducing unnecessary inference, optimizing query frequency).
Deliverable: 1–2 pages. Add to your AI Audit portfolio.
What's Next
In Chapter 19: Global Perspectives on AI — Governance, Culture, and Power, we zoom out to the international level. Different countries and cultures are developing, deploying, and regulating AI in very different ways — shaped by their political systems, economic priorities, cultural values, and historical experiences. You will examine how the EU, the United States, China, and the Global South are navigating AI governance, and why a technology developed primarily in a handful of wealthy nations raises profound questions about global equity and power.