Environmental Footprint & Carbon Accounting

Measuring the environmental footprint of AI systems is harder than it might seem. The direct emissions from running hardware (Scope 2 emissions from electricity) are relatively straightforward to estimate if you know the power consumption and the carbon intensity of the local electricity grid. But the full picture includes the embodied carbon in manufacturing the chips and servers (Scope 3), the emissions from cooling, the impact of building and maintaining data centres, and the upstream effects of mining the rare materials used in electronics. Tools like ML CO2 Impact, CodeCarbon, and cloud provider dashboards can estimate the carbon footprint of training runs and inference workloads. However, these estimates involve significant uncertainty, particularly around Scope 3 emissions and the actual energy mix powering a given data centre at a given time. For organisations wanting to report AI-related emissions accurately, the methodology is still evolving. The GHG Protocol provides general guidance, but AI-specific carbon accounting standards are in their early stages. What you can do today is estimate your AI energy consumption, understand the carbon intensity of your infrastructure providers, and factor environmental cost into your technology decisions. Perfect measurement shouldn't be the enemy of directionally correct decision-making.