Embodied Carbon in Computing Hardware
Embodied carbon is the greenhouse-gas emissions released to manufacture, transport, and eventually dispose of a piece of hardware — distinct from {{operational carbon}} burned while the device runs. For modern {{GPU}}s and AI accelerators, embodied carbon can be a large fraction of lifetime emissions.
**The two halves of a hardware footprint.** Lifecycle accounting splits emissions into operational carbon (the electricity a device consumes while running, multiplied by the grid's carbon intensity) and embodied carbon (everything required to bring the device into existence and retire it: mining and refining of materials, semiconductor fabrication, assembly, packaging, shipping, and end-of-life processing). Apportioning embodied carbon to a single workload requires amortizing the manufacturing footprint across the hardware's useful life and utilization rate. **Why semiconductor manufacturing is carbon-intensive.** Leading-edge fabs consume enormous amounts of electricity (often on grids still dominated by fossil generation in Taiwan, Korea, and parts of the US), use process gases with high global-warming potentials (SF6, NF3, fluorinated chemicals for etching), and require ultrapure chemicals and water. High-Bandwidth Memory (HBM), now standard on AI accelerators, adds further footprint through 3D die stacking and advanced packaging (CoWoS), which reduce yield and increase rework. **Numbers to keep in mind.** Luccioni et al.'s BLOOM lifecycle analysis attributed roughly 22% of total training-phase emissions to embodied carbon from the GPU cluster and supporting infrastructure. Independent analyses of NVIDIA A100 and H100 cards have estimated several hundred kg CO2eq of embodied carbon per accelerator. As operational energy gets cleaner (more renewable grid, more efficient cooling, lower PUE), embodied carbon becomes a larger share of total impact — which is why hardware longevity, high utilization, and refurbishment matter as much as choosing a clean grid.