Amid concerns that data centers could consume up to 12 percent of total U.S. electricity by 2028, researchers from MIT and the MIT-IBM Watson AI Lab have developed a rapid tool to estimate artificial intelligence power consumption. The new method, presented at the IEEE International Symposium on Performance Analysis of Systems and Software, enables operators to predict energy usage in seconds, a dramatic improvement over traditional modeling techniques that require hours or days. Led by MIT postdoc Kyungmi Lee, the team created a lightweight estimation model named EnergAIzer. Traditional approaches often involve emulating every step of a complex AI workload, a process that is too slow for practical use when comparing algorithms or configurations. EnergAIzer circumvents this by identifying repeatable patterns in software optimizations, such as how code is distributed across parallel processing cores, to generate quick and reliable power estimates. While speed was a primary goal, the researchers ensured accuracy by accounting for real-world variables. The model incorporates correction terms derived from actual GPU measurements to include fixed setup costs, operational energy charges, and hardware fluctuations like data bandwidth conflicts. When tested against real AI workloads on actual graphics processing units, EnergAIzer achieved an estimation error of only 8 percent, comparable to much slower traditional methods but delivered almost instantaneously. This tool offers significant benefits for the AI industry. Data center operators can use the rapid feedback to allocate limited resources more effectively across multiple models and processors, thereby improving overall energy efficiency. Similarly, algorithm developers can assess the potential energy footprint of new models before deployment, allowing them to optimize for sustainability early in the design process. The system is also flexible enough to predict power consumption for emerging hardware designs, provided the fundamental architecture does not change drastically. The research team includes Zhiye Song, an EECS graduate student, research managers Eun Kyung Lee and Xin Zhang from IBM Research, IBM Fellow Tamar Eilam, and senior author Anantha P. Chandrakasan, the MIT provost. Their work aims to bridge the gap between hardware capability and environmental responsibility by providing a fast, accessible solution for the entire AI stack. Currently funded in part by the MIT-IBM Watson AI Lab, the team plans to test EnergAIzer on newer GPU configurations and scale the model to handle scenarios where multiple GPUs collaborate on a single workload. By making energy estimation accessible to hardware designers, data center operators, and software developers, this innovation represents a crucial step toward a more sustainable artificial intelligence infrastructure.