The AI Energy Score attempts to quantify the environmental harm of AI models. Here’s how to use it.

Artificial Intelligence is notoriously energy-hungry. Now, a new tool from Salesforce is attempting to quantify the toll AI models are taking on the planet. “AI models vary dramatically in terms of their environmental impact, from big general models to small domain-specific models,” says Boris Gamazaychikov, head of AI sustainability at Salesforce. “What the AI Energy Score is aiming for is to develop a standardized way to actually start measuring these things.” AI Energy Score, launched Monday by Salesforce in collaboration with Hugging Face, Cohere, and Carnegie Mellon University, assigns a star rating to assess the energy efficiency of various AI models relative to other models. The ratings are available for 10 distinct tasks, from text and image generation to summarization and automatic speech recognition. A five-star rating is the most energy-efficient, whereas a one-star rating is least. It also shares a number estimate of the quantity of GPU energy consumed per 1,000 search queries in Watt hours. As of Feb. 10, DistilGPT2, a model developed by Hugging Face that is meant to be a “faster, lighter version of GPT-2,” was the most energy efficient, consuming 1.31 Wh of energy per 1,000 text generation queries. Alongside an interactive landing page with more information, the site also has a label generator that Gamazaychikov hopes developers will use to showcase their scores and drive awareness. For now, the tool only rates open-source models hosted on Hugging Face, a platform for data scientists, AI developers, and enthusiasts. That includes some models from Meta, Microsoft, Mistral, and StabilityAI. Salesforce also submitted its own open- and closed-source models for rating. Closed-source, or proprietary models, which include many big names such as OpenAI’s GPT-3 and 4, Anthropic’s Claude, and Google’s Gemini, are not readily available. But Gamazaychikov says there is a way for developers of these proprietary models to conduct the analysis securely and then submit it if they so choose. “By making a really easy to use, clear, standardized approach, we’re reducing any kind of friction between those leading companies being able to not do this,” Gamazaychikov says, adding that the tool was developed using feedback from academia, government officials, and AI companies. “We’re also hoping to increase some pressure, showing that it’s important to disclose this type of information.” Although Salesforce welcomes individuals to experiment with the tool to get a sense of how their use of AI models could take a toll on the environment, Gamazaychikov says “the primary audience is probably enterprises, or those that are integrating these types of models within their products.” Furthermore, a tool like AI Energy Score could help companies calculate indirect greenhouse gas emissions throughout the value chain. The launch of AI Energy Score follows Salesforce’s debut last fall of Agentforce, which allows customers to build or customize AI-powered agents that automate tasks or augment employee labor. Gamazaychikov says Salesforce also hopes to inspire greater change by giving regulators a jumping-off point for assessing the emissions-load of AI. He also envisions promoting the use of smaller, local models at a time when large reasoning models, which could consume even more energy than current large language models, are in development. A 2024 study from Goldman Sachs found that AI could drive up the energy demand of data centers by 160 percent by the end of the decade, even as the growing global population exerts additional pressures on the world’s energy needs. This comes at a time when leaders are contending with the worsening effects of climate change, prompting big tech companies to look to renewable energy and innovations in nuclear power to satisfy the growing demand. Want to try it out? Head on over to huggingface.co/AIEnergyScore and give it a whirl. BY CHLOE AIELLO @CHLOBO_ILO