Salesforce adds carbon, energy metrics to AI model cards

Salesforce has expanded its model card disclosures to include energy consumption and carbon emissions metrics, a step the company describes as bringing "more transparency and structure to AI sustainability." Model cards, sometimes called AI "nutritional labels," are fact sheets published alongside machine learning models to provide developers with information about training methods, performance, potential applications, and demographic factors. While many enterprise software companies publish model cards, Salesforce is unique in its inclusion of carbon and energy data as of June.

The environmental metrics now appear on model cards for Salesforce-trained models, including first name match and account match models. Sunya Norman, senior vice president of impact at Salesforce, explained the motivation: "As organizations scale AI, they're asking broader questions about how to deploy it responsibly, including how to better understand its environmental impact alongside performance, cost and business value." Norman noted that Salesforce has received positive feedback and that many customers are hopeful other model providers will adopt a similar approach.

Salesforce calculated the environmental metrics using the AI Energy Score, a resource published by open source software firm Hugging Face, which analyzes hardware specifications, data center location, run times, and other factors. A key innovation in Salesforce's approach is the inclusion of inference impact — the environmental cost when AI is applied to specific tasks — alongside training impact. While training measurements are relatively straightforward because they involve a discrete process and fixed computing environment, inference impact is harder to estimate because it varies depending on where the model runs. Salesforce's inclusion of this phase provides a more complete picture of the model's lifecycle footprint.

This initiative builds on Salesforce's longer-term commitment to AI sustainability. The company has published AI environmental data for at least two years as part of its "Sustainable AI Policy," which emphasizes governance and transparency around AI's social and environmental impacts. The strategy was championed by Salesforce's AI Sustainability team, led by Boris Gamazaychikov, who left the company in spring to co-found a consulting and research firm dedicated to managing AI's environmental impacts. Looking forward, Salesforce plans to add environmental metrics to additional models as measurement techniques become more straightforward. The company's sustainability team worked with its office of ethical and humane AI use to develop the model card data, framing AI sustainability as an ecosystem-wide challenge that will require shared methodologies, collaboration, and continued innovation.

Salesforce's decision to publish carbon and energy data alongside its AI models reflects a shift in how enterprise software companies approach AI governance. While model cards themselves are a standard practice among many vendors, Salesforce stands alone in adding environmental metrics as of June — a disclosure that signals both organizational commitment and response to customer demand. The company has supported this direction for at least two years through its "Sustainable AI Policy," which called for tighter governance and transparency around AI's social and environmental impacts. This commitment was led by the AI Sustainability team under Boris Gamazaychikov, who departed in spring to establish a consulting and research firm focused on managing AI's environmental footprint.

The technical challenge of measuring AI's environmental impact lies in the distinction between training and inference. Training involves a discrete, controlled process in a fixed computing environment, making measurement relatively straightforward. Inference — the phase where a deployed model actually processes tasks for end users — presents greater complexity because its environmental cost depends on where and how often it runs. By including inference calculations in its model cards, Salesforce acknowledges that a complete environmental picture must account for the full AI lifecycle, not just the initial training phase. The use of Hugging Face's AI Energy Score as the underlying methodology suggests an effort toward standardization, though Salesforce notes that shared methodologies and continued innovation across the ecosystem remain necessary.

No comments yet. Be the first to share your thoughts!