By Aparna Hanumantu
One day, you find yourself chatting with a machine that not only listens to your words but understands exactly what you mean, responding just like a human would. This is the magic of Large Language Models (LLMs), a type of artificial intelligence that acts as the brain behind many tools we use.
Five years ago, AI was still a futuristic concept, with tools like Siri and Alexa handling basic tasks. Today, Large Language Models (LLMs) have made technology more intuitive and efficient. For example, John, an HR employee, trains his LLM assistant by feeding it company documents which helps him to effectively screen job applications thus saving hours of manual screening of applicants. Over time, the assistant improves, learns and recognizes patterns in data, saving John time on routine tasks thus helping him work more efficiently. Similarly, bakery owner Sarah uses an LLM chatbot to automate customer inquiries, saving time and boosting her business. This shows us how LLMs have “read” billions of books, articles, and websites, learning how language works and how to respond in a natural way by transforming daily tasks, enabling people to work smarter.
The mechanics of LLMs
Training a Large Language Model (LLM) involves teaching a computer to understand and generate text by processing vast amounts of data, like books and websites. Then it uses a neural network model, adjusting its parameters to predict words or phrases based on initial learnings, and through multiple training rounds with the help of powerful computing units, it learns more and reports fine-tuned text for specific tasks. The more data and computing power available, the better the model becomes at producing human-like language.
Early Large Language Models (LLMs) like GPT-1 and Google’s BERT were simpler, but with GPT-2 in 2019, the real impact began. GPT-3 in 2020 came up with enhanced capabilities for tasks like writing and summarizing, and with the launch of ChatGPT in November 2022, the platform quickly gained popularity, reaching over 100 million active users who utilized it for a wide range of tasks. As AI advances, models like GPT-4 and Google’s PaLM deliver improved text generation, fueled by increased computational power and access to vast datasets. By 2025, ChatGPT’s user base could reach anywhere between 500 million to 1 billion, with a projected growth of 2 to 3 billion by 2030 and possibly surpassing 5 billion by 2035, as AI becomes essential across industries.
While Large Language Models (LLMs) have transformed how we interact with technology, they also pose significant environmental challenges. These models require enormous computational power, driving up energy consumption in data centers, many of which still rely on fossil fuels, leading to high carbon emissions. A total of 8,000 global data centers are in operation, with more than 600 new facilities added in 2023 alone. In the United States, there are 3,000 data centers, including over 1,000 newly established facilities in 2023. The training of large AI models in recent years (like GPT models) can result in emissions between 200 to 500 metric tons of CO2-equivalent (CO2e) for a single training, as much power as a small city, with its carbon emissions equivalent to thousands of cars over their lifetime.
The chart above illustrates the growth of global data centers alongside the corresponding increase in carbon emissions, measured in tons. It is evident that as the number of data centers has expanded over the years demanding more power, carbon emissions have slowly risen over the years 2018 to 2022. Notably, there is an exponential surge in emissions projected for the years 2030 and beyond, highlighting the growing energy demands and raising concerns about sustainability. For context, 100 tons of CO₂ in 2018 is equivalent to the emissions of approximately 20 cars driven for one year, or the energy consumption of 100-200 households for one month. For the years 2025 and beyond, 300,000 tons of CO₂ equates to the emissions from 60,000 cars driving for a year, or the energy used by 30 million homes for one month.
The above chart illustrates the cost components of running LLM models, including training costs for development, inference costs for production, and operational costs such as maintenance, energy consumption, and infrastructure. We observe a linear increase in costs from 2022 to 2024, with projected costs rising exponentially from 2023 to 2035.
Looking Ahead: Initiatives for Sustainable AI
Companies today are increasingly prioritizing the automation versus non-automation decision as part of their strategy. While AI has advanced in automating tasks, not all applications are necessary or efficient. For example, AI chatbots for simple inquiries like business hours or order status can often be replaced by basic FAQ pages or automated emails, offering a more cost-effective solution. Corporate responsibility in automation is essential, and many organizations establish AI Governance Councils to oversee ethical and sustainable AI practices, ensuring AI is used only where it adds significant value, thus reducing resource consumption, costs, and environmental impact while focusing AI resources on high-value areas.
While Large Language Models (LLMs) have transformed how we interact with technology, they demand enormous computational power, high energy consumption in data centers, often relying on fossil fuels thus leading to high levels of carbon emissions thus making them not so efficient models. Exploring efficient AI models like MobileNet, EfficientNet, and DistilBERT— which are smaller, faster, and less resource-intensive—offers a better alternative to LLM models. These models are optimized to deliver high performance with minimal computational resources, including processing power, memory, and energy consumption.
References:
Energy and Policy Considerations for Deep Learning in NLP (Strubell et al., 2019)
This paper discusses the computational costs and energy consumption associated with training large models like BERT and other NLP models, highlighting environmental concerns and offering a detailed look at energy usage and carbon emissions.
AI and Carbon Emissions: The Real Cost of Deep Learning (VentureBeat)
VentureBeat discusses how the increasing energy demand for LLMs and AI applications is contributing to carbon emissions, exploring potential solutions for more sustainable AI practices.
https://www.learningtree.com/blog/carbon-footprint-ai-deep-learning/
There’s greater cost of deploying AI and ML models in production – the AI carbon footprint
