In data analysis, dealing with highly skewed data often requires statistical techniques to improve model fit and explanatory power. Two common methods are log transformation and log link functions, both of which can make data more amenable to modeling but have distinct differences in their approach and implications. This study, conducted by a data analyst using Epoch AI's data, compares these two methods in the context of analyzing energy consumption in AI model training. The original energy consumption data (Energy_kWh) was heavily right-skewed and contained significant outliers. To address this, the analyst first applied a log transformation to the response variable. The transformed data appeared more normally distributed, with a Shapiro-Wilk test p-value of approximately 0.5. However, further modeling revealed some practical issues with this approach. Four different generalized linear models (GLMs) were constructed: a log-transformed Gaussian model, a log-linked Gaussian model, a log-transformed Gamma model, and a log-linked Gamma model. The Akaike Information Criterion (AIC) was used to evaluate these models. The AIC values for the log-transformed Gaussian and log-linked Gaussian models were 311.5963 and 2005.8263, respectively. For the Gamma distribution-based models, the AIC values were 352.5450 and 1780.8524 for the log-transformed and log-linked models, respectively. Although the log-transformed Gaussian model had a lower AIC value, its coefficients were problematic. Continuous variables had nearly zero or slightly negative slopes, and the intercept was around 1 kWh, which contradicts the expected high energy consumption of AI models. The analyst then switched to the log-linked Gamma model, which, despite having a slightly higher AIC value, provided better fitting and more interpretable results. In this model, each additional hour of training time increased the total energy consumption by 0.18%, and each additional hardware unit increased it by 0.07%. Moreover, the interaction term between training time and hardware units showed a negative effect, reducing energy consumption by 2 × 10⁵%. These findings align more closely with the real-world data, where energy consumption should logically increase with more training time and hardware units but may be offset by efficiency gains from better resource utilization. To visually compare the two approaches, the analyst created prediction plots. The left plot, based on the log-transformed Gamma model, showed a near-zero prediction line that diverged significantly from the actual data. In contrast, the right plot, derived from the log-linked Gamma model, produced a prediction line that closely matched the observed values, demonstrating superior fit and reliability. Ultimately, the analyst chose a log-linked Gamma GLM that included an interaction term between training time and hardware units, as well as different types of hardware. This model had an AIC value of 1775, balancing good fit with meaningful coefficient interpretation. It effectively captured the multiplicative effects of the predictor variables on energy consumption, maintaining the integrity of the original data scale. Industry experts emphasize that selecting the right modeling approach for skewed data is crucial. While log transformation can improve data distribution, it may distort the relationships between variables. On the other hand, log link functions preserve the original properties of the response variable and are thus better suited for interpreting real-world changes. Understanding the underlying mathematical principles is essential for accurate data analysis and model selection. The importance of this work lies in its practical application. The analyst, working for a startup focused on AI energy efficiency, has demonstrated a method that can lead to more reliable predictions and better resource management in AI training processes. This is particularly relevant as AI models become increasingly complex and resource-intensive, and optimizing energy usage is a growing concern for both environmental and economic reasons. By choosing the appropriate statistical methods, researchers and practitioners can make more informed decisions, ultimately contributing to more sustainable and efficient AI practices. The study highlights how correct model choice and statistical methods enhance predictive accuracy and research credibility. This work can serve as a valuable reference for data analysts and AI researchers looking to improve their energy efficiency analysis.