Material Insights · Understanding the Material World

AI accelerator chip: Bill of materials

About these figures

Which unit, study and year the composition figures describe, and how they vary

One product only: amounts are for the GPU chip, with its memory, of one NVIDIA A100 SXM 40 GB taken apart and analysed in a 2025 study [15], the only per-chip composition found. The module's heat sink and board are left out. The A100 has older, smaller memory than an H100. Custom accelerators have no composition data. Not yet product nodes: the silicon interposer, the package substrate, the baseboard and the heat sink.

as of 2026-11-16

Components

The products inside one unit, with quantity, mass share and cost share

Variant 1 Data-centre GPU
MaterialRoleAmount in one unitShare of the unit's mass
Memory chipComponent5 HBM stacks–
An H100 SXM5 package carries five active HBM3 stacks for its 80 GB [13]. An HBM stack is a pile of DRAM chips; the memory chip node covers them.
Variant 2 Custom accelerator
MaterialRoleAmount in one unitShare of the unit's mass
Memory chipComponentNot given –
Google's Ironwood carries 192 GB of HBM3e [39]; no source gives the number of stacks.

as of 2026-11-16. The silicon interposer and the package substrate are not yet product nodes (see new_products). The memory is 42% of the embodied carbon of NVIDIA's HGX H100 board [2]; nearly 60% of the manufacturing emissions of the Gaudi accelerator come from its HBM dies [10].

Direct materials

The commodities and chemicals in one unit as bought, with amount, mass share, cost share and recycled content

Variant Data-centre GPU
MaterialRoleAmount in one unitShare of the unit's mass
Si SiliconFunctional0.004 kg–
3.70 g in the GPU chip group [15]: the logic die, the memory dies and the interposer, cut from wafers of ultra-pure silicon; 13.4 g in the whole module.
Sn TinFunctional0.002 kg–
2.17 g in the GPU chip group [15], mainly as solder; 20.3 g in the whole module.
Al AluminiumFunctional0 kg–
0.418 g in the GPU chip group [15]; 6.9 g in the whole module.
Cu CopperFunctional0 kg–
0.0814 g in the GPU chip group of one A100 SXM 40 GB [15]. The whole module held 1,374 g, 1,300 g of it in the heat sink and 68.7 g in the circuit board, which are not part of the package [15].
Ag SilverFunctional0 kg–
0.0113 g in the GPU chip group [15]; 0.55 g in the whole module, the most abundant precious metal found.
Au GoldFunctional0 kg–
0.00115 g in the GPU chip group [15]; 0.0426 g in the whole module. Gold is used as bonding wire and in contacts and interconnects [19].

Indicative · as of 2026-11-16. One product only: amounts are the 'GPU chip' column (GPU die and its memory) of one NVIDIA A100 SXM 40 GB taken apart and analysed in a 2025 study [15]. The heat sink, circuit board and power parts of the module are left out, as they are not part of the package. The A100 is older than the H100 and has less memory. Tantalum was not reported. No source gives a composition for a custom accelerator.

How it is put together

The main steps from parts and materials to the finished unit

StepWhat happensTechnologySource
Print the logic dieA foundry prints the processor on a silicon wafer layer by layer; NVIDIA names TSMC and Samsung as its foundries [44].Planar process and photolithography[44]
Make the HBM stacksSK hynix, Micron or Samsung make the DRAM chips and stack them into high-bandwidth memory [44].Dynamic random-access memory[44]
Build the interposer package (CoWoS)The logic die and HBM stacks are placed on a passive silicon interposer, which links them [20]; NVIDIA uses CoWoS for packaging [44].– [20][44]
Mount on the substrateThe interposer and chips are attached to a substrate that carries the connections to the circuit board [20].– [20]
Assemble and test final productsContract manufacturers such as Hon Hai, Wistron and Fabrinet assemble, test and package NVIDIA's final products [44].– [44]
Fit to a baseboard and serverEight H100 GPUs go on a baseboard linked by NVLink and NVSwitch [2]; the DGX H100 server holds 8 GPUs with 640 GB of GPU memory [21].– [2][21]

as of 2026-11-16

Technologies used to make it

The technologies used to make it, at which step

Made with
  • Lit Planar process and photolithography Print the logic die layer by layer on silicon wafers The H100 logic die is made on TSMC's 4N process, customised for NVIDIA [13]; the Blackwell dies on 4NP [14].
  • EUV Extreme ultraviolet lithography Print the finest layers of the newest nodes Leading-edge nodes print their finest layers with EUV; no source opened says which layers of NVIDIA's chips use it. Resolution figures are on the processor chip page.
  • RAM Dynamic random-access memory Make the DRAM chips that are stacked into HBM NVIDIA buys memory from SK hynix, Micron and Samsung [44].

as of 2026-11-16

Technologies inside it

The technologies inside the product, with how much of each per unit

Embodies 4
  • GPU Graphics processing unit 1 GPU Data-centre GPU A data-centre GPU is a graphics processing unit scaled up and given tensor units; the twin technology page tells its history.
  • Tsr Transistor 80 billion transistors Data-centre GPU The H100 logic die has 80 billion transistors [13]; HBM adds more.
  • RAM Dynamic random-access memory 5 HBM stacks Data-centre GPU Five active HBM3 stacks of DRAM chips in an H100 SXM5 [13].
  • DL Deep learning The chip is built to run deep neural networks; the twin technology page explains the method.

as of 2026-11-16

Sources

  1. Product Carbon Footprint (PCF) Summary for NVIDIA HGX H100 (July 2025) NVIDIA, 2025
  2. Nvidia Hopper GPU architecture revealed. 4nm die & 18432 shaders HWCooling.net, 2022
  3. NVIDIA Blackwell Platform Arrives to Power a New Era of Computing NVIDIA investor relations, 2024
  4. From FLOPs to Footprints: The Resource Cost of Artificial Intelligence (arXiv:2512.04142v1) arXiv preprint, 2025
  5. Gold Demand Trends Full Year 2025: Technology World Gold Council, 2026
  6. AI Expansion - Supply Chain Analysis For CoWoS And HBM SemiAnalysis, 2023
  7. Introduction to NVIDIA DGX H100/H200 Systems (last updated 17 October 2026) NVIDIA, 2026
  8. Google's Ironwood TPUs represent a bigger threat than Nvidia would have you believe The Register, 2025
  9. NVIDIA 2026 Annual Report (Form 10-K, Item 1 Manufacturing) NVIDIA, 2026