What is AI made of?

Since 2025, Nvidia has published the manufacturing footprint of its HGX boards, which carry 8 chips, from raw material extraction to the factory gate. For the HGX H100, it is 1,312 kgCO2e : memory accounts for 42%, integrated circuits 25%, cooling components 18%. For the HGX B200, the next generation, it is 2,274 kgCO2e, 1.7 times more: memory accounts for 49%, integrated circuits 28%, cooling 12%. Assembly and transport account for less than 10% in both 2 cases.

The memory of these cards, known as HBM, is made of memory chips stacked on top of one another and placed next to the compute chip. From one generation to the next, its capacity has more than doubled, and so have its manufacturing emissions. The authors of the life cycle assessment of Google's TPU chips note the same movement: between the TPU v5e and the TPU v6e, the peak computing power of a chip was multiplied by 4.7, and its manufacturing emissions by about 1.8, with a larger chip, more memory around it and 3 times more working memory in the host server. Per computing operation, manufacturing emissions fell by 66%.

The whole server, with its processors, working memory and storage, weighs more: 8,341 kgCO2e for a server with 8 H100 according to a simplified GreenIT study, including manufacturing, distribution and end of life.

Nvidia and AMD design their AI chips but do not manufacture them. Nvidia's A100, H100, H200 and B200 chips and AMD's MI300X are fabricated by TSMC, in Taiwan, according to Greenpeace East Asia. Their HBM memory comes from SK hynix and Samsung, in South Korea, and from Micron; in its model, Greenpeace attributes to Micron the memory of the H200 and B200, which SK hynix also supplies. According to the American semiconductor manufacturers' association, the SIA, the United States had no production capacity for logic chips fabricated below 10 nanometres in 2022, the fineness of AI chips; it is expected to have 28% in 2032.

Greenpeace East Asia estimates the electricity used to manufacture Nvidia and AMD AI chips at 983.9 GWh in 2024, against 218 GWh in 2023 for Nvidia alone. Taiwan accounts for 375.8 GWh, about 38%, South Korea 315.2 GWh, about 32%, and Japan most of the rest, about 30%. The electricity of these 3 countries was 83.1%, 58.5% and 68.6% fossil-fuel based in 2023. Greenpeace derives 453,600 tCO2e from this in 2024. This is an estimate by model, the only one published for AI chips; it counts the electricity of the factories, excluding the fluorinated gases used in etching and the materials. 983.9 GWh is one fifth to one quarter of the electricity consumed in 2024 by data centres located in France on dedicated sites, 4 to 5 TWh according to RTE.

Hardware

Electricity for manufacturing AI chips

Electricity used to manufacture the AI chips of Nvidia (2023) then of Nvidia and AMD (2024), in GWh · 2023 and 2024

2023218GWh

2024984GWh

Source: Greenpeace East Asia, April 2025 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

Semiconductor factories use ultrapure water to rinse silicon wafers; our analysis of AI's water says what is known about it.

A chip is made of very pure silicon, etched, doped and connected by copper tracks; boards and servers add copper, aluminium, gold in the contacts and rarer metals. As of 14 November 2026, we have found no published composition of an AI server in metals, to the nearest kilogram.

Several semiconductor materials are produced in few countries. According to the United States Geological Survey, the USGS, China produced 99 % of the world's primary gallium in 2024, 750 tonnes out of 760, and remains the leading producer of germanium, which is used in the form of a gas to manufacture certain semiconductors. It banned its export to the United States in December 2024.

The life cycle assessment of Mistral Large 2, published in July 2025 with Carbone 4 and ADEME, is the only one to relate materials to an AI model. After 18 months, server hardware, covering manufacturing, transport and end of life, accounts for 11% of emissions but 61% of mineral resource depletion, the rest coming mainly from computing; the detail is in the chart in this analysis. Mistral AI specifies that the manufacturing of graphics cards had to be approximated there, for lack of manufacturers' data.

The AI Impact Database spreads the manufacturing of a server with 8 H100 over 5 years of service, used at 69%: about 35 gCO2e per graphics card hour. A graphics card hour uses about 0.99 kWh, data centre included.

The lower the carbon content of the electricity, the more manufacturing weighs. The Google authors give the same order of magnitude for their chips: about 10% of lifetime emissions if their purchases of carbon-free electricity are not counted, and half for manufacturing and half for electricity if the factories and data centres ran on 90% carbon-free electricity. Our analysis of AI decarbonisation draws out the levers, including the lifetime of the hardware.