Introduction Since the invention of the first digital computers in the 1940s, scientists have aspired to create human-like intelligence in machines, a vision that has now materialized into Artificial Intelligence (AI). The release of ChatGPT in November 2022 marked a significant milestone, showcasing an AI model capable of instantaneously responding to user queries. Since then, the rapid adoption of AI has seen over 1 billion messages being processed daily through ChatGPT alone (OpenAI Newsroom, 2024). However, the environmental costs associated with this technological advancement are often overlooked. Training AI models is an energy-intensive process, requiring massive computational resources and specialized hardware, which primarily rely on non-renewable energy sources like coal and natural gas. This study aims to increase awareness of these hidden costs and investigates two main research questions: Is there a significant relationship between AI models' architectural and hardware choices and their resource consumption during training? Has AI training become more energy-efficient over time? Methods The study utilized the "Notable AI Models" dataset from Epoch AI, which includes key training information such as the number of parameters, dataset size, total compute, hardware type, and quantity. Only models with a confidence level of "Confident" or "Likely" were analyzed. The dataset provided estimated Total Power Draw (W) and training times (hours), which were used to compute total energy consumption (kWh). Water and carbon footprints were derived using fixed conversion rates: Water Usage Effectiveness (WUE, 0.36 L/kWh, Lawrence Berkeley National Laboratory, 2024) and Carbon Intensity (CI, 0.548 kg CO₂e/kWh, Guidi et al., 2024). To assess energy efficiency, the Floating-Point Operations Per Second per Watt (FLOPS/W) metric was used, which measures computing performance relative to power consumption. The study employed Generalized Linear Models (GLMs) with log transformations to analyze the relationships between predictors and response variables, specifically energy consumption and efficiency. Results RQ1: Architectural and Hardware Choices vs Resource Consumption A Gamma GLM with a log link was selected to model resource consumption due to its lower Akaike Information Criterion (AIC) value (1780.85) and better predictive accuracy compared to other models. The study found that while architecture-related variables (Parameters, Training Compute, Dataset Size) did not significantly affect resource consumption, hardware-related factors (Hardware Quantity and Type) and Training Time did. Specifically, each additional hour of training increased energy consumption by 0.18%, and each extra hardware unit by 0.07%. The interaction term between Training Time and Hardware Quantity was significantly negative, indicating that as training time increases, the energy consumption growth slows down when more hardware units are used. Different hardware types exhibited significant variations in energy consumption: - NVIDIA A100, Tesla V100, and Google TPU v3 were among the most energy-efficient. - Older models, such as NVIDIA GeForce GTX 285, were significantly less efficient. RQ2: Energy Efficiency Over Time A log-linked Gamma model was used to examine the relationship between energy efficiency (FLOPS/W) and publication year. The study found a positive but slight improvement in energy efficiency of 0.13% per year (GLM: t = 8.005, p < 0.001). Notably, longer training times decreased energy efficiency by 0.03% per hour (GLM: t = 8.581, p < 0.001). Variations in efficiency were observed among models using the same hardware, suggesting that factors beyond hardware, such as training practices, may influence energy consumption. Discussion The study revealed that hardware choices, particularly the type and quantity of hardware, and training duration significantly impact the environmental costs of AI training. For example, a single hour of training on an NVIDIA A100 chip consumes 29,213 kWh of energy, 10,521 liters of water, and emits 16,009 kg of CO₂, far exceeding the yearly usage of an average U.S. household. These findings underscore the substantial environmental footprint of AI training. While there has been a slight improvement in energy efficiency over time, the progress is minimal, and newer, more efficient hardware has not been widely adopted. The study's small dataset (70 valid models) and limited sample size across 18 hardware types may have constrained the statistical power of the analysis. Future research should explore these interactions with larger and more diverse datasets. Industry Insights and Company Profiles Industry experts agree that the environmental impact of AI is a critical concern. Calvert (2024) highlights that current AI energy usage is comparable to that of a small country and emphasizes the need for sustainable solutions. OpenAI and other tech companies are increasingly aware of these issues but face challenges in balancing performance and environmental sustainability. The lack of comprehensive and transparent data on resource consumption remains a barrier to systematic change. Epoch AI, known for its research into AI trends, supports the findings by providing detailed datasets on notable AI models, helping researchers and practitioners understand the environmental implications of different hardware and training practices. Companies like NVIDIA and Google are investing heavily in developing more efficient hardware, but widespread adoption and optimized training practices are crucial for achieving significant environmental benefits. This study's insights can guide future efforts in creating more sustainable AI technologies.