Salesforce adds lifecycle energy and carbon estimates to AI model cards

Salesforce said it is expanding the “model cards” it publishes for AI systems with standardized environmental impact metrics, giving customers an additional lens on energy use and estimated carbon emissions tied to each model’s lifecycle.

In a release dated June 8, 2026, Salesforce described model cards as “nutrition labels” for AI models, documenting intended use cases, evaluation results, and guardrails. The company said the new Environmental Impact section extends that transparency beyond performance and safety by adding sustainability disclosures across the model lifecycle.

Salesforce said AI workloads depend on physical infrastructure, particularly data centers, and that training and running models can require substantial compute. It argued that as AI adoption accelerates, customer expectations for transparency should include how systems are built and operated, including their environmental footprint.

The company said the Environmental Impact disclosures estimate energy consumption and carbon emissions across three phases: pre-training, post-training, and inference. Salesforce said it uses the AI Energy Score methodology to calculate the figures, which it described as an emerging framework that standardizes energy reporting by considering factors such as hardware type, GPU utilization, runtime, and data center region.

Salesforce added that the new section is available for select models, naming First Name Match, Account Match, and TextEval. The release positioned the workflow update as a way to embed environmental reporting into model evaluation processes so model builders can treat sustainability metrics alongside performance and risk information.

Executives and researchers at Salesforce framed the change as a practical step toward its broader “trusted AI” approach. The release said the update reinforces Salesforce’s trusted AI principles and ISO 42001-certified governance standards, and it cited a “transparent” pillar within those principles, describing model cards as part of how the company communicates why and how its AI systems are built and evaluated.

Salesforce also said the disclosure work reflects collaboration between its AI Research and Impact teams and its responsible AI organization to make sustainability metrics part of the same review pipeline customers may rely on for other model evaluation outputs. In particular, the company quoted Orlando Lugo, Senior Product Manager, Responsible AI, and Sarah Tan, Principal Research Scientist, Responsible AI, characterizing the model-card approach as a way to operationalize sustainability as a measurable component of trusted AI.

The AI Energy Score documentation describes a methodology intended to support more comparable comparisons by standardizing benchmarking conditions, including focus on GPU energy. It also notes that carbon emissions estimation depends on location and grid carbon intensity, and that additional energy components may require further approximations when translating GPU energy into broader energy and emissions estimates. Salesforce’s release did not specify whether customers will see the underlying assumptions for each model, or provide any example values, so the degree of technical auditability may depend on what Salesforce exposes in the model cards themselves.