14 Mar 2025 ~ 3 min read

The Carbon Cost of AI Training

As someone who spent several years working in the conservation field before transitioning into software engineering, I’ve always been mindful of the environmental impact of technology. Recently, while checking out Hugging Face’s NLP course, I was struck by the section on AI’s carbon cost. It got me thinking about how the energy demands of training AI models could affect our planet in the long run.

Back in 2019, a study by Strubell et al. highlighted just how energy-hungry AI training could be. They estimated that training a single large transformer model could emit as much carbon as five American cars over their lifetime.

It’s been over five years since this paper sparked widespread discussion and concern. Let’s take a quick look at how the conversation around AI’s environmental footprint has evolved—what progress, if any, have we made in reducing energy consumption, and where are we headed now?

How AI’s Carbon Footprint is Evolving

Energy Use is Soaring

AI’s hunger for computational power has only grown. More complex models, larger datasets, and expanding infrastructure all require vast amounts of energy. Google’s most recent environmental report revealed a 48% increase in emissions over five years, primarily due to the rising demand for AI services and data center expansions. As cloud computing and AI become more integrated into everyday applications, maintaining sustainable energy practices is becoming a significant challenge.

Google’s 2024 Environmental Report

New Approaches to Training

Some AI researchers and companies are working on more efficient training methods. For example, DeepSeek has proposed techniques that aim to reduce computational costs, such as auxiliary-loss-free strategies designed to streamline learning without unnecessary redundancies. However, while these approaches improve training efficiency, it remains unclear whether they lead to a net reduction in carbon emissions. Some studies suggest that efficiency gains might be offset by the increasing scale of AI training.

DeepSeek Research | MIT Technology Review Analysis

Growing Environmental Concerns

A recent 2025 AI Safety Report flagged the rising emissions and water usage associated with AI infrastructure, particularly as data centers expand at an unprecedented rate. Cooling these massive facilities requires significant water and electricity, creating further environmental strain. Meanwhile, companies like xAI, founded by Elon Musk, have built large-scale AI data centers in record time—one in just 19 days—but often rely on high-pollution energy sources like gas turbines. The rapid expansion of AI infrastructure without sufficient investment in renewable energy exacerbates the environmental impact.

UK Government AI Safety Report | Time: Musk’s AI Data Center

What Can We Do?

AI isn’t going anywhere, but we can make it greener:

  • Use Pre-trained Models – Instead of starting from scratch, leverage existing models that have already been trained. This avoids duplicating energy-intensive computations and significantly reduces emissions.
  • Optimize Training – Smarter training techniques like gradient accumulation, mixed precision, and early stopping can reduce power usage without sacrificing performance.
  • Use Efficient Hardware – Choosing energy-efficient GPUs and TPUs can cut emissions while maintaining AI performance.
  • Power AI with Renewables – Data centers running on solar, wind, or hydroelectric power can significantly lower carbon impact.
  • Track Carbon Footprint – Tools like ML CO₂ Impact and CodeCarbon help monitor emissions and guide greener AI decisions.

Resources