In short
Embodied carbon is the emissions from making, shipping and disposing of hardware. To account for it, take the hardware's total embodied emissions, share them out by how long and how much of the kit your AI work uses, and add that share to the operational figure, as the SCI specification's M term does. Because the data is often thin, state your assumptions; and remember that keeping hardware longer spreads the same emissions over more work.
Steps
- Ask your hardware or cloud supplier for embodied emissions data for the equipment you use.
- Allocate a share to your AI work by the time and fraction of the hardware it uses.
- Add that share to your operational emissions and say which you included.
- State the assumptions, including expected hardware lifetime.
- Extend hardware lifetimes where you control them, so embodied emissions cover more useful work.
Count the hardware as well as the electricity, because the cleaner the grid gets, the larger the manufacturing share becomes. Wu and colleagues at Meta reported in 2022 that embodied carbon was roughly 30% of the footprint of Meta's large-scale machine learning, and argued it was becoming the dominant source as operational energy gets cleaner.
How do I allocate it?
The Green Software Foundation's SCI specification includes embodied emissions as the M term. A simple approach:
- Take the device's total embodied emissions from the manufacturer or supplier.
- Divide by its expected lifetime to get emissions per year or per hour.
- Multiply by the time your work reserved it, and by the share of the machine you used.
For cloud services you can't see the hardware, so ask the provider whether its carbon reporting includes manufacturing (Scope 3) and how it is allocated. See your cloud provider's carbon data.
What about e-waste?
Wang and colleagues estimated in 2024 that large language models could generate between 1.2 and 5.0 million tonnes of electronic waste in total between 2020 and 2030, depending on how fast their use grows, and that reuse and recycling could cut this substantially.
Common mistakes
- Leaving embodied carbon out without saying so. Put it in the boundary statement either way.
- Assuming a short life by default. Replacing kit early raises the yearly share.
- Buying new hardware for a small efficiency gain without checking the manufacturing cost it adds.
See how to measure AI energy for boundaries.
Sources
- Sustainable AI: environmental implications, challenges and opportunities, Wu et al., MLSys (2022)
- E-waste challenges of generative artificial intelligence, Wang et al., Nature Computational Science (2024)
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
Related
How do I calculate a Software Carbon Intensity score for an AI feature?
Work out an SCI score for an AI feature: energy times grid intensity, plus embodied emissions, divided by a functional unit such as per request.
How do I avoid replacing my devices just to use AI?
Why you rarely need a new phone or laptop to use AI, how to keep older devices useful, and why hardware life matters for AI's footprint.
How do I choose energy-efficient hardware for AI?
Pick AI hardware on measured energy per task at your real batch sizes, compare runtimes and accelerators, and weigh new kit against embodied carbon.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.