# AI accelerator chip: Interventions

Source page: https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-interventions/

[Try this in What if](https://www.materialinsights.org/analyse/what-if/?p=p_ai_accelerator_chip) 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

| Lever | Life stage | Change | Effect per unit | Readiness | Adoption | Who acts | Source |
|---|---|---|---|---|---|---|---|
| Use less | Use | Keep 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 found | Model labs, cloud operators | ^[15] |
| Circularity | Use | Run each accelerator for longer before replacing it | Extending 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]. | Deployed | Not found | Cloud operators, buyers | ^[15]^[17]^[3] |
| Efficiency | Use | Move to newer generations that do more work for each watt | Google 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]. | Deployed | Rolling out with each generation | Chip designers, cloud operators | ^[40]^[50]^[3] |
| Change route | Use | Power the data centre with carbon-free electricity | For 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 procurement | Cloud operators, utilities | ^[3] |
| Design | Components design | Stack compute and memory in 3D rather than spreading chips out on a large interposer | When compute units and HBM capacity are doubled, embodied emissions rise 89% with 3D integration against 141% with 2.5D (imec and VUB) ^[11]. | Research | Not found | Chip designers, packaging firms | ^[11] |
| Efficiency | Making | Abate fluorinated gases in the fab and cut the gases used in etching | TSMC 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] |
| Electrify | Supply chain | Supply fabs with renewable electricity | Electricity can make up to 60% of the carbon footprint of an advanced N2 logic wafer ^[12]. | Deployed | Renewable electricity was 11.4% of TSMC's Taiwan electricity in 2025 (2,940 GWh) ^[30] | Foundries, utilities | ^[12]^[30] |
| Business model | Use | Pool public GPUs and rent them out, so more users share fewer idle chips | No saving per unit found; the IndiaAI portal has onboarded over 38,000 GPUs to rent at an affordable rate ^[23]. | Deployed | 38,000 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](https://arxiv.org/pdf/2502.01671) Schneider et al. (Google), arXiv preprint 2502.01671, 2025
2. [3D Integration Reduces Embodied Emissions Impact Of GPU Scaling (imec and VUB)](https://semiengineering.com/3d-integration-reduces-embodied-emissions-impact-of-gpu-scaling-imec-vub/) Semiconductor Engineering, 2026
3. [How can we reduce environmental impact in chip manufacturing?](https://www.imec-int.com/en/articles/how-can-we-reduce-environmental-impact-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)](https://www.arxiv.org/pdf/2512.04142) arXiv preprint, 2025
5. [More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU (arXiv:2509.00093v3)](https://arxiv.org/pdf/2509.00093) arXiv preprint, 2025
6. [IndiaAI Mission expands AI ecosystem with affordable compute and startup support](https://www.digitalindia.gov.in/press_release/indiaai-mission-expands-ai-ecosystem-with-affordable-compute-and-startup-support/) Digital India, Government of India, 2026
7. [TSMC 2025 Sustainability Report](https://esg.tsmc.com/zh-Hant/file/public/2025-TSMC-Sustainability-Report-e.pdf) Taiwan Semiconductor Manufacturing Company (TSMC), 2026
8. [Google offers 7th-gen Ironwood TPUs for AI, with AI-inspired comparisons](https://www.theregister.com/2025/04/10/googles_7thgen_ironwood_tpus_debut/) The Register, 2025
9. [NVIDIA HGX B200 Reduces Embodied Carbon Emissions Intensity (19 September 2025)](https://developer.nvidia.com/blog/nvidia-hgx-b200-reduces-embodied-carbon-emissions-intensity) NVIDIA Technical Blog, 2025

## Other sections

- [Summary](https://www.materialinsights.org/products/p_ai_accelerator_chip/index.md)
- [Working](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-working/index.md)
- [Bill of materials](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-bom/index.md)
- [Use case](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-use/index.md)
- [Supply chain](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-supply/index.md)
- [Footprint](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-footprint/index.md)
- [Evolution](https://www.materialinsights.org/products/p_ai_accelerator_chip/prod-evolution/index.md)
