For years, enterprise software customers have scrutinized the financial cost and latency of artificial intelligence, but a third dimension—environmental cost—has remained largely opaque. Salesforce is now pushing that calculus into the open, adding energy consumption and carbon emissions metrics to the model cards for several of its proprietary AI models. This move, announced in June, transforms what was essentially a technical documentation exercise into a strategic transparency play that could reshape how enterprises evaluate AI deployment.
Model cards, often described as AI’s nutritional labels, have been published by Salesforce since 2020. They typically detail training data, performance benchmarks, and demographic considerations. The addition of electricity use and carbon footprint data—starting with models for first name matching and account matching—represents a significant departure from industry norms. As Sunya Norman, Salesforce’s senior vice president of impact, noted, customers are increasingly asking how to deploy AI responsibly, not just efficiently. The feedback so far has been positive, with many hoping other providers will follow suit.
The environmental data is calculated using the AI Energy Score, an open-source tool from Hugging Face, which accounts for hardware, data center location, run times, and task type. While measuring the impact of AI training is relatively straightforward—a discrete process in a controlled environment—Salesforce has also included the more variable inference phase, where the model is applied to real-world tasks. This is a technically challenging but critical step, as inference often accounts for the majority of a model’s lifetime energy use.
Salesforce’s disclosure builds on a foundation laid at least two years ago with its Sustainable AI Policy, championed by the now-departed AI Sustainability team lead Boris Gamazaychikov, who left this spring to co-found a consultancy focused on managing AI’s environmental impacts. The policy called for tighter governance and transparency regarding AI’s social and environmental effects. The new model card metrics are a direct operationalization of that framework, but they also highlight a broader industry tension: as AI adoption accelerates, the environmental footprint of inference becomes an ecosystem-wide challenge that no single vendor can solve alone.
The implications for energy professionals and investors are twofold. First, enterprises that deploy AI at scale now have a tangible way to compare the environmental cost of different models, potentially influencing procurement decisions and data center location strategies. Second, the move pressures other major AI providers—including hyperscalers and foundation model developers—to standardize environmental disclosures. If this becomes a competitive differentiator, the energy intensity of AI could shift from an afterthought to a key performance metric, driving demand for more efficient hardware, greener data centers, and carbon-aware scheduling.
Read the full report at GreenBiz.