The most powerful AI models emit up to 50 times more carbon dioxide than smaller ones

The artificial intelligence (AI) boom has sharply increased energy consumption and carbon emissions of major tech companies, as AI systems must process massive amounts of data using high-powered machines. But not all generative AI models consume the same amount of energy. A group of researchers set out to determine which models are the most polluting and which are the most efficient. Their conclusion: smaller models (those trained with fewer parameters) can produce up to 50 times less pollution than larger ones — but their answers tend to be less accurate, and vice versa. The greater the precision, the greater the pollution.

The findings of this study, published on Thursday in the journal Frontiers, are based on a comparison of 14 open-source generative AI models, meaning researchers are able to explore and manipulate their inner workings. The selection included models of various sizes, including some with built-in “reasoning” capabilities: that is, models that go beyond simply predicting the next word in a sentence (as large language models, or LLMs, typically do), and are able to analyze and infer.

Another key takeaway: for solving a simple task, it’s better to use a smaller model. The result will usually be comparable to that of a larger model — but with roughly four times less environmental impact.

What’s the value of the information provided by this study? Ideally, it can help users choose the right model based on the complexity of the task at hand. Just as we walk or ride a bike for short trips but opt for public transport or a car for longer journeys, powerful AI models could be reserved for advanced tasks — like complex programming — while smaller models might suffice for simpler jobs, like basic translations.

But selecting the best model for each task is far from simple. That’s why Dauner and his team are developing an automated tool to guide this choice. “We’re working on a planner that selects the most appropriate model based on the user’s request, thereby minimizing CO₂ equivalent emissions,” the researcher notes.

For example, using the Chinese model DeepSeek R1 to answer 600,000 questions would generate as much CO₂ as a round-trip flight from London to New York. Meanwhile, Qwen 2.5 — a model of similar size — can handle over three times as many questions (around 1.9 million) with similar accuracy and the same carbon footprint.

If users always chose the least polluting model, it would help reduce AI’s growing environmental footprint. The latest report from the International Telecommunication Union (ITU), the U.N. agency specializing in digital technologies, reveals that energy consumption by data centers (which power AI models) grew by 12% annually between 2017 and 2023 — four times faster than global energy use overall. The four biggest companies developing AI (Amazon, Microsoft, Alphabet, and Meta) have seen their emissions increase by an average of 150% since 2020.

That’s according to the U.N. agency, which analyzed public data from 2023 — the most recent year with full information — submitted by the world’s 200 largest tech firms regarding greenhouse gas emissions, energy use, and environmental targets.

“Despite the progress made, greenhouse gas emissions continue to rise, confirming that​​ the need for digital companies to adopt science-aligned, transparent, and accountable climate strategies has never been greater,” said Cosmas Luckyson Zavazava, one of the report’s authors, during its presentation.

According to the data, the 10 tech firms with the highest energy use (China Mobile, Amazon, Samsung, China Telecom, Alphabet, Microsoft, TSMC, China Unicom, SK Hynix, and Meta) consume more electricity in a single year than the entire country of Spain.

Less polluting models

“Our results reveal strong correlations between LLM size, reasoning behavior, token generation, and emissions,” write the study’s authors, Maximilian Dauner and Gudrun Socher, from the Munich University of Applied Sciences. “While larger and reasoning-enabled models achieve higher accuracy, up to 84.9%, they also incur substantially higher emissions, driven largely by increased token output. Subject-level analysis further shows that symbolic and abstract domains such as Abstract Algebra consistently demand more computation and yield lower accuracy.”

Researchers tested three versions of Meta’s Llama model, ranging from 8 billion to 70 billion parameters; four versions of Alibaba’s Qwen, between 7 billion and 72 billion parameters; three models from Deep Cogito, between 8 billion and 70 billion parameters; and three more from DeepSeek, also ranging from 7 billion to 70 billion parameters. The authors clarify upfront that the results cannot be extrapolated to more well-known models like GPT, Gemini, or Copilot, which were not included in the experiment because they are not open source.

All 14 models were asked to answer the same 500 questions across different subjects. Each model received 100 questions in each of the following five areas: philosophy, world history, international law, abstract algebra, and mathematics — at a difficulty level comparable to university entrance exams. The test was carried out in two phases. In the first, a multiple-choice format was used, with four possible answers provided. In the second, the models had to respond to open-ended questions without any prompt guidance. To grade the answers, the researchers used OpenAI’s o4-mini model, “which is faster and smaller than other GPT models, meaning it emits less CO₂,” notes Dauner.

All experiments were carried out on an Nvidia A100 GPU with 80 GB of memory, which allowed for accurate measurement of energy use, memory consumption, and response time. “We ruled out considering water consumption in the study because we couldn’t measure it directly, but only through estimates,” Dauner explains. Water is used to cool the high-density processors needed for AI computations.

The larger models performed better on both open-ended and multiple-choice questions.

How to measure AI’s environmental footprint

The researchers considered the full lifecycle of AI models — from the mining of materials used to build GPUs, to the resources consumed during manufacturing, database creation and processing, model design and training, and eventual deployment.

“Due to limited transparency across these phases, existing studies often rely on estimates of material and manufacturing impacts […] or focus on directly measurable quantities, notably energy consumption during training and inference,” the study states.

Most analyses, they add, focus on emissions produced after the tool’s launch. This is not the case in this study.

To perform the calculations, all greenhouse gases (carbon dioxide, methane, and nitrous oxide) were converted into carbon dioxide equivalents, using the global warming potential (GWP) of each gas relative to CO₂ as a measure.

“The study is very interesting for better understanding the carbon footprint of LLMs,” says Shaolei Ren, associate professor of electrical and computer engineering at the University of California, Riverside and a specialist in AI sustainability. “But it would have been even more interesting if the authors had used country- or region-specific carbon intensities, as there are significant differences.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition