Material Insights · Understanding the Material World

AI accelerator chip: Interventions

Try this in What if Switch on levers and see what changes in India by 2036.

Interventions

Changes that would cut emissions or material use, by lever and life stage, with effect per unit, readiness, adoption, evidence and who acts

LeverLife stageChangeEffect per unitReadinessAdoptionWho actsSource
Use lessUseKeep accelerators busy: raise the share of peak arithmetic that training achieves (model FLOPs utilisation)Raising utilisation from 20% to 60% cuts the GPUs needed by about 67% (A100 study) [15].Deployed (software and scheduling)Not foundModel labs, cloud operators[15]
CircularityUseRun each accelerator for longer before replacing itExtending lifespan from one to three years gives savings comparable to raising utilisation; both together cut GPU needs by up to 93% [15]. Assumed lifespans range from 3 years [17] to 6 years [3].DeployedNot foundCloud operators, buyers[15][17][3]
EfficiencyUseMove to newer generations that do more work for each wattGoogle says Ironwood gives 2x the performance per watt of its predecessor [40]; NVIDIA says the HGX B200 can be as much as 15x more energy efficient for AI inference (its own claim) [50]. Embodied carbon per exaflop of Google's TPUs fell from 114 to 38 gCO2e between v4i and v6e [3].DeployedRolling out with each generationChip designers, cloud operators[40][50][3]
Change routeUsePower the data centre with carbon-free electricityFor TPU v6e, operational intensity falls from 316 to 118 gCO2e per exaflop under market-based accounting, a 63% cut (calculated) [3].Deployed (Google procurement)Google's carbon-free energy procurementCloud operators, utilities[3]
DesignComponents designStack compute and memory in 3D rather than spreading chips out on a large interposerWhen compute units and HBM capacity are doubled, embodied emissions rise 89% with 3D integration against 141% with 2.5D (imec and VUB) [11].ResearchNot foundChip designers, packaging firms[11]
EfficiencyMakingAbate fluorinated gases in the fab and cut the gases used in etchingTSMC reports a company-wide abatement rate of 96% for fluorinated gases [30]; imec's TAP process cut CF4 use by 98% [12].Deployed (abatement); piloted (new etch)96% abatement at TSMC in 2025 [30]Foundries, equipment makers[30][12]
ElectrifySupply chainSupply fabs with renewable electricityElectricity can make up to 60% of the carbon footprint of an advanced N2 logic wafer [12].DeployedRenewable electricity was 11.4% of TSMC's Taiwan electricity in 2025 (2,940 GWh) [30]Foundries, utilities[12][30]
Business modelUsePool public GPUs and rent them out, so more users share fewer idle chipsNo saving per unit found; the IndiaAI portal has onboarded over 38,000 GPUs to rent at an affordable rate [23].Deployed38,000 GPUs on the Indian portal [23]Governments, cloud operators[23]

as of 2026-11-16

Sources

  1. Life-cycle emissions of AI hardware: a cradle-to-grave approach and generational trends Schneider et al. (Google), arXiv preprint 2502.01671, 2025
  2. 3D Integration Reduces Embodied Emissions Impact Of GPU Scaling (imec and VUB) Semiconductor Engineering, 2026
  3. How can we reduce environmental impact in chip manufacturing? imec (first published in Semiconductor Digest, July/August 2025 issue), 2025
  4. From FLOPs to Footprints: The Resource Cost of Artificial Intelligence (arXiv:2512.04142v1) arXiv preprint, 2025
  5. More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU (arXiv:2509.00093v3) arXiv preprint, 2025
  6. IndiaAI Mission expands AI ecosystem with affordable compute and startup support Digital India, Government of India, 2026
  7. TSMC 2025 Sustainability Report Taiwan Semiconductor Manufacturing Company (TSMC), 2026
  8. Google offers 7th-gen Ironwood TPUs for AI, with AI-inspired comparisons The Register, 2025
  9. NVIDIA HGX B200 Reduces Embodied Carbon Emissions Intensity (19 September 2025) NVIDIA Technical Blog, 2025