A practical playbook for becoming a sustainable AI company, and the standard MABS Digital builds AI software to: measured, right-sized, and run where the grid is cleaner.
How we measure it
ISO/IEC 21031:2024SCI = ((E × I) + M) per R
Software Carbon Intensity · Green Software Foundation
The case
AI’s energy appetite is real, and so is the room to cut it. Four facts that frame the work.
Electricity · IEA
Source: IEA, Energy and AI. The 2030 figure is an IEA projection.
Efficiency · Google
less energy for a median Gemini text prompt in twelve months: 0.24 Wh in May 2025, with a 44× lower carbon footprint.
Figures reported by Google for its own services. Not independently verified.
Standard
The Software Carbon Intensity specification is now an international standard. It gives teams one agreed way to score the carbon intensity of software.
Regulation · EU AI Act
Providers of general-purpose AI models must document their models’ energy consumption, and estimates are allowed when exact figures are unknown. As a buyer, ask your vendors what energy data they can share.
The roadmap
Start with a baseline, end with a habit. Each step leaves you with something you can keep.
Baseline
You walk away with
A footprint baseline: one number per workload, with gaps flagged.
Targets
You walk away with
A one-page AI sustainability policy.
Right-size
You walk away with
A ranked list of efficiency changes, each with expected gain and risk.
Run clean
You walk away with
A deployment and scheduling plan with the trade-offs written down.
Report and improve
You walk away with
A release-by-release scorecard and a quarterly review.
Step 04 in one picture
Carbon intensity varies by region and by hour. Work that can wait can move. Live inference can’t, so right-size it instead.
Schematic. Illustrates the idea; not real grid data.
Our standard
Six habits, applied across the lifecycle of every build, from first idea to switch-off.
Design
01 / 06Every large-model choice comes with a written reason and an agreed quality bar. If a smaller option clears the bar, we use it.
Build
02 / 06Retrieval before generation, caching and token budgets from the first commit. Distilled or quantized models where quality holds, and batching where users can wait.
Measure
03 / 06We score each build with SCI per functional unit, label estimates as estimates, and ask model providers for their energy data.
Deploy
04 / 06Batch work and training run in lower-carbon windows where your data rules allow it. Idle services scale to zero.
Operate
05 / 06Alerts on regressions, and a quarterly review of the KPIs we agreed with you.
Retire
06 / 06An inventory with an owner for every model and pipeline, and a retirement date set at launch.
FAQ
Building and running AI so that its energy use, carbon emissions and hardware footprint are measured, proportionate to the value it delivers, and reduced over time. It covers the model, the code around it and where it runs.
With the Software Carbon Intensity score, an ISO/IEC standard: SCI = ((E × I) + M) per R. E is energy, I is the carbon intensity of the grid, M is embodied hardware emissions and R is the unit you choose, such as one request. Where exact data is missing, we estimate and label the number as an estimate.
Not automatically. Many tasks don’t need the largest model. We test quality before and after every change, so you keep a smaller model only when it clears the quality bar you set.
Mostly not for live inference, because users are waiting and latency matters. It suits training, fine-tuning and batch jobs, which can run later or in a cleaner region. For chatbots, the better levers are right-sizing, caching and shorter prompts.
Providers of general-purpose AI models must document their models’ energy consumption, and estimates are allowed when exact figures are unknown. Standardization work and voluntary codes of conduct on energy efficiency are also planned. This is a general summary, not legal advice.
Next step
Bring one AI workload to a free 30-minute call, and we’ll talk through how to measure and shrink its footprint.
Last updated .