The Climate Cost of AI: Examining the Environmental Cost of Large Language Models

In the past couple of years, we have witnessed a rapid increase in scope and size of large language AI models (LLM). Tools like Open AI’s o1 or Google DeepMind’s Gemini model are pushing boundaries in what machines can do, from generating music to writing code to helping scientists accelerate drug discovery. But alongside these advances comes a growing concern: What is the environmental impact of all this computing power?  

It’s true that modern AI systems require a lot of energy. Training and running them at scale demands powerful hardware and vast data centers. But here’s a common misconception: AI isn’t uniquely wasteful, and swearing off it entirely won’t “fix” climate change—not any more than turning off your lights or recycling will. The real issue lies in the broader infrastructure behind the entire digital world, and AI is just the newest, most visible contributor.  

By 2030, global data center energy usage is projected to exceed 1,000 terawatt-hours – comparable to the annual energy consumption of Japan. While a growing slice of that demand is from AI, it also powers everything from Netflix to online banking. So the question isn’t whether AI should exist or be developed, but rather: How we can leverage the visibility of AI to promote and build more sustainable methods of energy consumption. 

Why Does AI Use So Much Energy? 

At the core of the energy usage issue are the enormous computational demands of training and running large-scale AI models. Training a model like GPT-3 involves running quadrillions to quintillion of calculations, often spread across thousands of processors running in parallel over several weeks. Each individual operation might require only a tiny amount of electricity, but scaled to trillions of parameters and massive datasets, the total energy use becomes staggering. 

The most commonly used processors for these tasks are GPUs (Graphics Processing Units), originally designed for gaming and graphics rendering. Because of their ability to handle many calculations at once, they’ve become standard in machine learning. More recently, companies have started using TPUs (Tensor Processing Units)—custom chips specially optimized for AI workloads—which can be much more energy-efficient. However, TPUs are still expensive, hard to access, and often reserved for internal use by companies like Google. 

A landmark 2019 study from the University of Massachusetts Amherst found that training just one large NLP model can emit over 284 metric tons of CO₂—roughly equivalent to the lifetime emissions of five American cars. Even then, training is only part of the process. 

Even after a model is trained, running it (or “inference”) on user queries, which happen constantly, also consumes energy. In fact, according to a recent Berkeley study, running just one day’s worth of ChatGPT queries resulted in 50 pounds of CO2 or 8.4 tons of CO2.

But again, context matters: by comparison, the aviation industry produced over 1 billion metric tons of CO2 annually. The emissions from AI are real, but they’re still a small part of a much larger climate equation. 

Where AI’s Emission Come From 

What really determines AI’s climate cost isn’t just how much energy it uses–-it’s where the energy comes from. If a data center is powered by fossil fuels, the carbon footprint is high. If it runs on wind, solar, or nuclear, the impact is much lower. 

Today, most tech companies use a mix. And as demand for AI continues to rise, new energy sources are being discovered and researched. In 2024, Google partnered with Kairos Power to explore small nuclear reactors for data center power. Similarly, Microsoft signed a 20-year deal to source energy from the Three Mile Island nuclear plant. Others are experimenting with more efficient cooling systems, including placing data centers underwater to reduce heat naturally. 

Still, many AI systems today rely on electricity grids that use coal or natural gas, especially in regions where renewables aren’t yet dominant. The results? Between 2019 and 2023, Google’s emissions rose 48% and Microsoft’s by 29%. Despite sustainability pledges, clean energy infrastructure is struggling to keep up with digital demand. 

It’s Bigger Than Just AI 

Here’s the broader truth: AI is not the biggest driver of our carbon footprint. Not even close. 

The global fashion industry, for instance, emits more greenhouse gases than international flights and maritime shipping combined. The meat industry? A massive contributor to deforestation and methane emissions. Even cryptocurrency mining, in some cases, burns more electricity than many small countries. 

What AI highlights is a more systemic issue: as our digital lives grow, so does our need for electricity. AI is simply the newest, and most demanding, use case. It’s a symbol of the larger challenge we face in decarbonizing the grid, making energy use more efficient, and building systems that scale without hurting the planet. 

A Smarter Path Forward

So what can be done? Actually, quite a lot! And the good news is that there is already progress. 

1. Designing Leaner Models 

Not all AI models need to be massive. Techniques like pruning, model distillation, and sparse attention allow developers to build efficient models that perform just as well but consume far less energy. 

2. Smarter Hardware 

Specialized processors (like TPUs, upcoming neuromorphic chips, and even optical computing) are built to handle AI tasks more efficiently than general-purpose GPUs, reducing energy computation dramatically. 

3. Green Energy + Better Cooling 

Transitioning data centers to 100% renewable energy can slash emissions. Improving cooling systems alone (which make up 30-40% of data center energy use), and innovations in heat reuse and natural cooling are showing real promise. 

4. Carbon Transparency 

Projects like Hugging Face’s Emissions Calculator are pioneering carbon transparency in AI. 

Making carbon accounting part of AI development and including emissions disclosures in model papers and tech announcements can push the field toward responsibility and accountability.  

AI Can Help Fight Climate Change, Too

AI isn’t just a climate cost – it can be part of the solution too. AI models are already being used to:

• Forecast extreme weather more accurately 

• Optimize renewable energy distribution 

• Design next-generation solar materials 

• Identity methane leaks from satellites 

If we can use AI responsibly, it can help solve the climate problem it is part of. 

So no, quitting AI won’t save the planet. Neither will skipping a plastic straw or switching off a light once. But that doesn’t mean small actions don’t matter, they just remind us that real, visible change comes from transitioning to smarter systems. 

The future of AI should be powerful, accessible, and efficient. Not just in what it can do for us, but also its effects on the Earth.