Embodied Energy of AI Training Hardware

The embodied energy of an AI accelerator is the total primary energy spent on extracting raw materials, fabricating the die, packaging the board, and shipping it to a data center. It is conventionally measured as cradle-to-gate kg CO2e via life cycle assessment and sits alongside operational carbon (electricity drawn during training and inference) in the total footprint of a GPU (Graphics Processing Unit): From Rendering Pixels to Training AI. Advanced-node semiconductor fabrication dominates the manufacturing share. NVIDIA's H100 and A100 are built on TSMC's 4N and 7nm processes, and TSMC's own disclosures put per-12-inch-wafer-mask-layer power at roughly 28 kWh in 2022, rising past 40 kWh as 3 nm scaled in 2023. EUV photolithography steps are especially power-hungry, and TSMC's Taiwan operations are projected to consume a quarter of the island's electricity by 2030. Fabs also draw millions of litres of ultrapure water per day, use thousands of process chemicals (including PFAS), and rely on high-purity neon, argon, and specialty gases. Materials add their own footprint. A modern accelerator package depends on copper, tin, tungsten, hafnium, gallium, and small quantities of rare earth elements, plus tantalum capacitors whose ore (coltan) is heavily sourced from the Democratic Republic of the Congo with documented labour and conflict concerns. HBM memory stacks, substrates, and thermal components contribute the bulk of the per-board embodied carbon: NVIDIA's published Product Carbon Footprint for an HGX H100 baseboard estimates roughly 1,300 kg CO2e of embodied emissions for eight SXM cards, with memory near 42 percent, integrated circuits 25 percent, and thermal hardware 18 percent. Over a typical 3-to-5-year service life, operational electricity usually still outweighs embodied carbon for heavily utilized training GPUs. Patterson et al. (2021) showed that grid mix and accelerator efficiency swing training emissions by orders of magnitude. Wu et al.'s "Sustainable AI" (Meta, 2022) reported that embodied emissions approach roughly half of total impact once data centers run on low-carbon power, and Gupta et al.'s "Chasing Carbon" (HPCA 2021) made the broader case that hardware manufacturing has become the limiting term in computing's footprint. Falk et al.'s 2025 cradle-to-grave LCA of the A100 added that the use phase dominates climate-change impacts but manufacturing dominates human toxicity, ozone depletion, and minerals depletion categories. Caveats: vendor PCFs rely on supplier-reported data and disclosed Scope 3 emissions that are notoriously incomplete; end-of-life recycling rates for high-density server boards are low; and figures scale strongly with the carbon intensity of the fab's local grid, which for Taiwan and South Korea remains coal- and gas-heavy.