03 · The Footprint of AI and Software2 min read
Digital decarbonization: tools worth knowing
A field guide to measuring and reducing the footprint of software, websites and AI: open standards, calculators, data sources and where each fits.
Theme 03
The Footprint of AI and Software
We have energy ratings and LCAs for physical products, but software is still figuring out how to account for itself, especially AI. This theme gathers notes on measuring Software Carbon Intensity, the plateau in data center efficiency, carbon-aware scheduling and the many hidden layers beneath a single AI agent.
Entries
03 · The Footprint of AI and Software2 min read
The AI agent infrastructure iceberg, through an LCA lens
An AI agent looks like one chat box. Underneath sit a dozen layers of infrastructure, each with its own footprint. How LCA thinking helps map them.
03 · The Footprint of AI and Software2 min read
The carbon cost everyone is ignoring in AI
Token optimization is everywhere. Carbon optimization is not. Why data center efficiency has plateaued while AI demand keeps climbing.
03 · The Footprint of AI and Software2 min read
Measuring software carbon intensity for AI
Notes on the SCI specification, what it asks product teams to measure, and the product strategies that lower the footprint of AI features.
Common questions
- What is Software Carbon Intensity (SCI)?
- SCI is a Green Software Foundation specification, now ISO/IEC 21031, that expresses the carbon emitted per functional unit of software, such as per query or per user, combining energy, grid intensity and embodied hardware emissions.
- What is PUE?
- Power usage effectiveness is total data center facility power divided by the power that reaches IT equipment. Industry averages fell quickly after 2007 and have been close to flat at around 1.58 for several years.
- How can a product team lower AI emissions?
- Add energy impact to acceptance criteria, set carbon budgets next to performance budgets, choose cleaner grid regions, schedule flexible workloads for low-carbon hours, cache aggressively and right-size models.