Is AI Bad for the Environment?

If you’re wondering “how is AI bad for the environment,” the honest answer is that today’s boom comes with real costs—electricity, water, and embodied emissions in the chips and buildings that make modern AI possible. But it’s not all doom: smarter design and cleaner power can shrink that footprint even as AI usage grows. Here’s the grounded view, with real numbers and what they mean.

 

The footprint: where it comes from

Training giant models concentrates energy use over weeks or months on clusters packed with GPUs/TPUs. Inference (running the model to answer queries) can dwarf training over time because it happens continuously at global scale. Both phases live in data centers whose environmental impact is driven by:

• Electricity (and the fuel mix behind it),

• Cooling water (especially where evaporative systems are used), and

• Supply chains (cement, steel, and especially semiconductor manufacturing).

Put simply: big models + lots of users = big, ongoing resource draw.

 

What the data shows (info table)

Bottom line: AI’s environmental footprint is material and growing, especially for electricity and water—but it’s shaped by design choices that we control.

 

So…how is AI bad for the environment?

1. It can drive up grid demand. Hyperscale AI clusters pull hundreds of megawatts at a site; aggregated worldwide, data‑center electricity could roughly double by 2030. Where power is fossil‑heavy, operational emissions rise.

2. It’s “thirsty.” Popular cooling designs evaporate large amounts of water. In places already under water stress, that’s a local environmental risk as AI scales.

3. Upstream emissions are big. Cement and steel for new buildings, plus chip fabrication, create Scope 3 spikes that many companies now report.

4. Usage never sleeps. Even if a training run is efficient, inference at global scale can dominate lifetime energy—inefficient model choices multiply the impact.

But “bad” isn’t destiny: what actually works

• Right‑size the model. Use smaller or distilled models, quantized weights, or Mixture‑of‑Experts so heavy compute only wakes up when needed. That slashes inference energy without killing quality for many workloads. For developers looking to optimize further, an AI tool that generates efficient code from plain English can streamline the process of creating energy-efficient applications.

• Code and system efficiency. Kernel‑level optimizations, batching, caching, and token‑efficient prompting often cut compute by double digits—for free.

• 24/7 clean power, not just offsets. Locating and time‑shifting AI workloads to high-renewables grids (and contracting for 24/7 carbon‑free energy) reduces real‑world emissions, not just accounting lines.

• Low‑water or water‑free cooling. Rear‑door heat exchangers, refrigerant‑based or direct‑to‑chip liquid cooling, and recirculated/treated water all shrink onsite withdrawals.

• Measure what matters. Track per‑request energy/water and surface it to product teams. If you can’t see it, you can’t improve it.

A pragmatic answer

Is AI bad for the environment? **It can be—**especially when we deploy oversized models on fossil‑heavy grids with water‑intensive cooling and ignore upstream supply‑chain impacts. But the same technology can optimize grids, buildings, logistics, and materials in ways that avoid more emissions than it creates—if the industry embraces efficiency‑first design and verifiable clean power.

The next two or three years will determine whether the curve bends toward “sustainable AI” or locks in a larger, harder‑to‑abate footprint.

If you build or buy AI, demand this checklist: smaller where possible, measured per‑query footprint, cleaner siting (24/7 CFE), cooling with less water, and transparent Scope 3 reporting. That’s how we make “how is AI bad for the environment” a question with a shrinking list of answers.

---

Sources (annotated)

• S1. International Energy Agency (IEA), Energy and AI (2025): global data‑center electricity use (~415 TWh in 2024; ~945 TWh base‑case by 2030) and AI‑driven accelerator growth. IEA

• S2. Microsoft, 2024 Environmental Sustainability Report (FY2023 data): +29.1% total emissions vs 2020 baseline; Scope 3 +30.9%, with drivers including data‑center construction and hardware. The Official Microsoft Blog

• S3. Google 2024 Environmental Report coverage: 14.3 MtCO₂e in 2023; +48% vs 2019 as AI infrastructure expanded. (News summary referencing the official report.) Data Center Dynamics

• S4. Google water use (2023): 6.1 billion gallons; +17% YoY—linked to data‑center cooling as AI services grew. (Industry report summarizing the Google 2024 Environmental Report.) Aquatech Trade

• S5. Li, Yang, Islam, Ren (2023–2025), Making AI Less “Thirsty”: training GPT‑3 estimated ~700,000 L on‑site (cooling) and ~5.4 million L including off‑site power water; ~500 mL of water per ~10–50 generative responses. arXivar5iv

Note: The IEA also concludes that while data‑center demand rises, AI can help cut emissions elsewhere if adopted wisely—another reason to focus on efficiency + clean energy simultaneously. IEA

Key takeaways to remember: right‑size models, run on cleaner grids, cool with less water, and measure per‑query footprint. That’s how we keep AI’s benefits—and rein in its environmental bill.