Embodied Carbon
Embodied Carbon
When discussing AI’s environmental impact, most attention goes to operational energy: the electricity consumed during training and inference. But before a single query runs, significant environmental costs have already been incurred.
The Pre-Operational Footprint
Mining and extraction:
- Rare earth elements for chip components
- Copper, gold, silver for circuitry
- Silicon for wafers
- Water and energy for mining operations
- Land disruption and local pollution at mining sites
Manufacturing:
- Chip fabrication (fabs) are enormously energy-intensive
- Ultra-pure water in huge quantities
- Toxic chemicals for etching and processing
- Clean rooms requiring constant environmental control
- Multiple fabrication steps, each with its own footprint
Transportation:
- Global supply chains (mining in one country, processing in another, assembly in a third)
- Components shipped multiple times before final product
- Data center equipment shipped to deployment location
Construction:
- Data center buildings (concrete, steel, embodied carbon in construction)
- Cooling systems
- Power infrastructure
- Networking equipment
All of this happens before the AI answers its first question.
The Hidden Accounting
Embodied carbon is harder to track than operational carbon:
- Supply chains are complex and often opaque
- Extraction happens in many countries with varying standards
- Manufacturing emissions aren’t always attributed to end products
- Providers don’t disclose full lifecycle assessments
A provider might truthfully say “our data centers run on renewable energy” while ignoring that the hardware itself required enormous carbon emissions to produce.
Hardware Turnover
The problem compounds with hardware turnover:
- AI chips become obsolete quickly (performance improvements, efficiency gains)
- Data centers upgrade hardware frequently
- “Old” hardware may be only 2-3 years old
- Each upgrade cycle incurs new embodied carbon
The faster the upgrade cycle, the more embodied carbon per unit of compute delivered.
E-Waste
At end of life:
- Specialized AI chips have limited secondary uses
- Recycling recovers only some materials
- Some components are hazardous
- Disposal often happens in developing countries with weak environmental protections
The embodied carbon at the beginning of lifecycle and the waste at the end are both externalized from the AI cost structure.
Implications
- Full lifecycle accounting dramatically changes AI’s environmental footprint
- Operational energy is the visible tip of a larger iceberg
- Hardware efficiency improvements may be offset by faster turnover
- “Clean energy” claims are incomplete without lifecycle assessment
Open Questions
- What’s the true full-lifecycle carbon cost of AI inference?
- How should embodied carbon factor into AI environmental claims?
- Can hardware lifespans be extended without sacrificing capability?
- Who should be responsible for embodied carbon in global supply chains?
See Also
- Training vs Inference Footprint — the operational energy that gets more attention
- Geographic Inequality of Compute — where extraction and manufacturing impacts land
- The Irony of AI for Climate — whether full lifecycle still nets positive
- The Nuclear Renaissance Question — nuclear plants carry enormous embodied carbon in concrete, steel, and reactor materials
- Stranded Assets Risk — stranded infrastructure means its embodied carbon was spent for nothing
- The One More Query Problem — aggregate demand justifies the hardware whose embodied carbon we’re counting