SCI for AI — Deep Dive

A deep-dive episode on SCI for AI — the standard I’m proud we shaped at the Green Software Foundation, and one of the most consequential pieces of work in green software today. The full episode is above.

Why AI needs its own measurement standard

The Software Carbon Intensity (SCI) specification gave the industry an ISO-aligned way to measure the carbon impact of software. But AI systems stress that framework in ways conventional software does not: a lifecycle split between training and inference with very different consumption profiles, specialized accelerators with their own embodied footprint, and usage that scales with every prompt rather than every deployment. Measuring AI’s carbon honestly requires a methodology built for those realities.

What SCI for AI provides

SCI for AI is the industry’s first consensus-based standard for measuring the carbon footprint of AI systems — built on the ISO-certified SCI methodology and designed for the full lifecycle of AI, from training through inference. The episode walks through how the specification approaches that lifecycle, and the principle underneath it, consistent with everything we do at the Foundation: the emphasis is on measurement and reduction, not offsets. You cannot engineer down what you have not measured — and once you can measure it, reduction becomes an engineering discipline like any other.

From specification to practice

A ratified specification is the beginning, not the end. The work now is operational: helping teams calculate their first SCI for AI scores, folding the measurement into design decisions — model selection, inference architecture, accelerator choice, the cost-carbon-latency trade-off — and building the best practices and reference architectures that make green AI systems the default rather than the exception.

If you build or operate AI systems, this is a standard worth knowing early: the disclosure landscape is moving toward it, and the teams measuring today will be the ones ahead of it tomorrow. The specification itself is open at sci.greensoftware.foundation.