Material Insights · Understanding the Material World

AI accelerator chip: Footprint

Raw materials in one unit

Every commodity and chemical inside one unit, traced through its components by the site, with its emission factor and emissions

VariantMaterialAmount (kg)ThroughEmission factor (kg CO2e per kg)Emissions (kg CO2e)
DATACopper0.006Memory chip3.9650
DATASilicon0.005Memory chip– –
DATAEpoxy resins0.004Memory chip– –
DATATin0.002– – –
DATAAluminium0– 14.80
DATASilver0– – –
DATAGold0– – –
CustomCopper0.001Memory chip3.9650
CustomEpoxy resins0.001Memory chip– –
CustomSilicon0Memory chip– –

Worked out from data

The site's estimate

Emissions of one unit worked out by the site from its materials and energy, for the world and India, with the share of mass covered by emission factors

VariantStagePlaceEmissions (kg CO2e)Mass covered by factors
DATAMakingWorld17.638%
DATAUseWorld18,873100%
DATAMakingIndia25.438%
DATAUseIndia27,609100%
CustomMakingWorld11.756%
CustomUseWorld– –
CustomMakingIndia17.156%
CustomUseIndia– –

Worked out from data

Energy to make it

Energy used to make one unit, by carrier and stage

Energy to make
  • El Electricity 22.4 kWh · Manufacturing Calculated, logic die only: 2.75 kWh/cm2 of fab energy for a 5 nm node in the ACT model [42] x 8.14 cm2 for the 814 mm2 H100 die [13] = 22.4 kWh (80.6 MJ), before yield losses, which ACT applies as a separate factor [42]. TSMC's 4N is a derivative of the 5 nm N5 family [13]. HBM, the interposer and packaging are not included: no source opened gives their energy, only carbon (see the stages note).

Estimate · as of 2026-11-16. No source gives the energy to make a whole accelerator package; treat the figure as a floor for the logic die alone.

Energy in use

Energy one unit uses per year of use, by carrier

Energy in use
  • El Electricity 6868 kWh per year · lifetime 6 years Data-centre GPU Calculated: 700 W (H100 SXM design power [4]) x 8,760 h x 0.80 (operational time LBNL assumes for training servers [22]) x 1.4 (US average PUE in 2023 [22]) = 6,868 kWh a year at the meter; 4,906 kWh at the chip. At the 40% operational time for inference it is 3,434 kWh. With Uptime's 2023 industry average PUE of 1.58 [31] the training case is 7,751 kWh. Lifetime is Google's six-year assumption for AI machines [3]; the A100 study assumes three years [17].

Estimate · as of 2026-11-16. Design power overstates real draw: LBNL measured an 8-H100 node in training at 7.9 kW, about 78% of rated power [22], and Google found TDP exceeds actual average power for TPUs by 2 to 6 times [3]. At 78% the training case falls to 5,357 kWh a year (calculated). NVIDIA rates the DGX H100 at 10.2 kW maximum [21] and the DGX B200 at about 14.3 kW maximum [41]; whole-server use is on the data-centre server page. No power figure was found for custom accelerators.

Emissions by stage

Published life-cycle emissions of one unit in four stages (making, logistics, use, end of life), with their split, place and variant

StageValueRangeUnitShareVariantPlaceYearSource
Making105– kg CO2e per A100 GPU chip with its memory (82.3% of the 127.6 kg module figure, calculated)– Data-centre GPUWorld2025[17]
Making127.6– kg CO2e per A100 SXM 40 GB module (chip, board, heat sink and casing), cradle to gate– Data-centre GPUWorld2025[17]
Making164– kg CO2e per GPU place on an HGX H100 baseboard (1,312 kg divided by eight, calculated); materials and components are 91% of the baseboard figure– Data-centre GPUWorld2025[2]
Logistics0.66– kg CO2e per GPU place (0.4% of the 1,312 kg baseboard figure, divided by eight, calculated)– Data-centre GPUWorld2025[2]
Use– – kg CO2e per GPU over a lifetime87Data-centre GPUWorld2025[17]
Making692386 to 1101 kg CO2e per TPU chip, embodied including a share of the host machine (TPU v6e; range v4i to v5p)kg CO2e per TPU chip, embodied including a share of the host machine (TPU v6e; range v4i to v5p)11Custom acceleratorWorld2025[3]
Use57552149 to 5755 kg CO2e per TPU v6e chip over six years: location-based high, market-based low (calculated from the study's carbon intensities)kg CO2e per TPU v6e chip over six years: location-based high, market-based low (calculated from the study's carbon intensities)89Custom acceleratorWorld2025[3]

Estimate · as of 2026-11-16. One product per row. The A100 figures are for one A100 SXM 40 GB: 127.6 kg for the module, of which the GPU chip and its memory are 82.3% [17]. The NVIDIA baseboard is cradle to gate only and excludes use and end of life [2]; its 1,312 kg includes the switch chips and heat sinks as well as the eight GPUs. By the chart's labels, memory is 546 kg of it, about 68 kg per 80 GB GPU (calculated) [2]. The ACT model gives 0.15 kg CO2 for packaging one IC [42]; no source opened gives the interposer's footprint. The ADEME range of 60 to 150 kg CO2e embedded per card [16] covers eight cards. Google's TPU figures include a share of the host machine [3]: embodied 692 kg and carbon intensities of 38 (embodied) and 118 (market-based) or 316 (location-based) gCO2e per exaflop for v6e [3]; the use value is 692 x 316/38 (calculated). TechInsights expects a 2029 average AI GPU above 1 tonne, nearly seven times an H100 [9]. The 87% use share is for training BLOOM on the A100 [17]. No end-of-life figures were found.

End of life

What happens to a unit when it is thrown away: collection, reuse, recycling and disposal shares, by place

RouteSharePlaceYearSource
All e-waste documented as formally collected and recycled (not specific to accelerators)22.3World2022[29]

as of 2026-11-16. No source gives an end-of-life split for accelerators, nor for any figure in India. The world generated 62 million tonnes of e-waste in 2022, with metals worth USD 91 billion in it [29].

Recovered as

The commodities and chemicals recovered at end of life, with recovery rates

Recovered as
  • Cu Copper Recycling of the heat sink and boards with the chip An A100 module holds 1,374 g of copper, 1,300 g of it in the heat sink and 0.08 g in the GPU chip [15]; no recovery rate for accelerators was found.
  • Au Gold E-waste recycling 0.0426 g of gold in an A100 module, 0.00115 g of it in the GPU chip [15]; no recovery rate found.
  • Sn Tin Recycling of solder 20.3 g of tin in an A100 module, 2.17 g of it in the GPU chip [15]; no recovery rate found.

as of 2026-11-16

End of life in India

Collection and recycling in India, including the informal sector

No source gives how many AI accelerators are recovered, in India or elsewhere. The world made 62 million tonnes of e-waste in 2022, and only 22.3% was documented as properly collected and recycled [29]. In India the E-Waste (Management) Rules 2022 bring IT and telecommunication equipment under extended producer responsibility from 1 April 2023, with certificates valid for two years [35]. Its schedule lists mainframes and minicomputers, but no entry for AI accelerators was found, and no India-specific figure for informal recycling was found.

as of 2026-11-16

Water and waste

Water use and waste per unit where a source gives them

MeasureValueUnitStageYearSource
TSMC company-wide abatement rate for fluorinated greenhouse gases96%Manufacturing2025[30]
Direct process emissions (fluorinated gases and nitrous oxide) as a share of TSMC's greenhouse gas emissions4%Manufacturing2025[30]
Share of the carbon footprint of an advanced logic wafer that can come from electricity (scope 2), up to60%Manufacturing2025[12]
Share of chemical use per wafer that goes to wet processing50% (about)Manufacturing2025[12]
Manufacturing emissions of AI GPU production, TechInsights estimate for 2024 and forecast for 20301.8 then 21.6million tonnes CO2Manufacturing2030[9]

as of 2026-11-16. No source gives water or waste per accelerator package. Wet processing takes about half the chemicals and a major share of the ultra-pure water used per wafer, and on an N2 node dry etch is the largest source of direct emissions [12]. TSMC's water use per wafer layer and imec's footprint per N2 wafer are on the processor chip page.

Sources

  1. Product Carbon Footprint (PCF) Summary for NVIDIA HGX H100 (July 2025) NVIDIA, 2025
  2. Life-cycle emissions of AI hardware: a cradle-to-grave approach and generational trends Schneider et al. (Google), arXiv preprint 2502.01671, 2025
  3. NVIDIA H100 Tensor Core GPU: specifications NVIDIA, 2026
  4. The AI Carbon Challenge: Why Your Next GPU's Memory Stack Is the New Emissions Hotspot TechInsights, 2025
  5. How can we reduce environmental impact in chip manufacturing? imec (first published in Semiconductor Digest, July/August 2025 issue), 2025
  6. Nvidia Hopper GPU architecture revealed. 4nm die & 18432 shaders HWCooling.net, 2022
  7. From FLOPs to Footprints: The Resource Cost of Artificial Intelligence (arXiv:2512.04142v1) arXiv preprint, 2025
  8. More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU (arXiv:2509.00093v3) arXiv preprint, 2025
  9. 2024 United States Data Center Energy Usage Report (LBNL-2001637) Lawrence Berkeley National Laboratory, 2024
  10. Global e-Waste Monitor 2024: Electronic Waste Rising Five Times Faster than Documented E-waste Recycling UNITAR, 2024
  11. TSMC 2025 Sustainability Report Taiwan Semiconductor Manufacturing Company (TSMC), 2026
  12. Large data centers are mostly more efficient, analysis confirms Uptime Institute Journal, 2024
  13. The Gazette of India: Extraordinary, G.S.R. 801(E), 2 November 2022: E-Waste (Management) Rules, 2022 (copy on the Andaman and Nicobar Administration site) Ministry of Environment, Forest and Climate Change, Government of India, 2022
  14. ACT: Designing Sustainable Computer Systems With An Architectural Carbon Modeling Tool Gupta et al., ISCA '22, 2022