# Data-centre server: Bill of materials

Source page: https://www.materialinsights.org/products/p_data_centre_server/prod-bom/

#### About these figures

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

Steel, aluminium and plastics are for a Lenovo ThinkSystem SR650 of 18.47 kg ^[4]; copper, cobalt and neodymium are from the JRC's 2015 example two-socket rack server of 27.8 kg ^[3]. Power supplies (1.713 kg), fans (0.899 kg) and drives are separate parts in the Lenovo list ^[4]. Gold sits in boards and chips, counted in those nodes. No source gave the make-up of an AI server beyond its 130.45 kg total. Not yet nodes: server power supply, server fan, hard disk drive, solid-state drive, GPU baseboard and heat sink.

as of 2026-11-16

#### Components

The products inside one unit, with quantity, mass share and cost share
 Variant 1 **General-purpose**

| Material | Role | Amount in one unit | Share of the unit's mass |
|---|---|---|---|
| [Processor chip](https://www.materialinsights.org/products/p_processor_chip/) | Component | 2 chips | – |
| Two CPUs per two-socket server, for example two Intel Xeon Silver 4410Y in the Dell R760 ^[1]. One CPU weighs 147.97 g in a Supermicro SYS-512B-WR server ^[5]. |  |  |  |
| [Memory chip](https://www.materialinsights.org/products/p_memory_chip/) | Component | Not given | – |
| A standard Lenovo SR630 V3 test configuration holds sixteen 16 GB memory modules ^[6]; no source counts the chips. |  |  |  |
| [Printed circuit board](https://www.materialinsights.org/products/p_printed_circuit_board/) | Component | Not given | – |
| The circuit-board node counts bare boards by area, and no source gives the board area of a server. Lenovo's SR650 list holds 1.940 kg of boards with their parts (motherboard and processors 0.9743 kg, RAM 0.088 kg, other boards 0.8777 kg; sum calculated), so that mass includes the processor and memory chips linked above and is not used as a mass share ^[4]. |  |  |  |

 Variant 2 **AI accelerator**

| Material | Role | Amount in one unit | Share of the unit's mass | Share of cost |
|---|---|---|---|---|
| [Processor chip](https://www.materialinsights.org/products/p_processor_chip/) | Component | 10 chips | – | 59.1% |
| Eight H100 GPUs and two Xeon CPUs in a DGX H100 ^[8]. Cost share is the 'GPU (COGS)' line of an analyst cost stack for a 2023 DGX H100, USD 26.6 thousand of 45.0 thousand (sum and share calculated); it is a manufacture cost, not a price ^[33]. |  |  |  |  |
| [Memory chip](https://www.materialinsights.org/products/p_memory_chip/) | Component | Not given | – | – |
| 2 TB of system memory in 32 modules ^[8], and 80 GB of HBM3 stacked memory on each GPU ^[9]; memory ICs were 42% of the GPU baseboard's making emissions ^[9]. |  |  |  |  |

as of 2026-11-16

#### Direct materials

The commodities and chemicals in one unit as bought, with amount, mass share, cost share and recycled content
 Variant **General-purpose**

| Material | Role | Amount in one unit | Share of the unit's mass |
|---|---|---|---|
| [St Steel](https://www.materialinsights.org/commodities/steel/) | Structural | 11.688 kgrange 11.688 to 12.265 kg | **63.3%** |
| Steel of the chassis, screws and other parts of a Lenovo ThinkSystem SR650 of 18.47 kg; Lenovo gives 63%, 63.3% calculated ^[4]. High is the 12.265 kg steel chassis of the JRC's example two-socket rack server of 27.8 kg ^[3]. |  |  |  |
| [Al Aluminium](https://www.materialinsights.org/commodities/aluminum/) | Functional | 1.327 kg | **7.2%** |
| Heat sinks and CPU brackets of a Lenovo ThinkSystem SR650; Lenovo gives 7%, 7.2% calculated ^[4]. |  |  |  |
| [Cu Copper](https://www.materialinsights.org/commodities/copper/) | Functional | 0.702 kgrange from 0.702 kg | – |
| Calculated lower bound from the JRC's example two-socket rack server: 0.179 kg in the chassis, 0.442 kg in the processor heat pipes and 0.081 kg in cables ^[3]. Copper in boards, power supplies and fans is not counted, and Lenovo does not list copper separately ^[4]. |  |  |  |
| [ABS ABS](https://www.materialinsights.org/chemicals/abs/) | Structural | 0.306 kgrange 0.306 to 0.348 kg | **1.7%** |
| Lenovo's 0.306 kg is the ABS plus polycarbonate of the air bezel and clips of an SR650 ^[4]; the JRC's example server has 0.348 kg of ABS and 0.282 kg of polycarbonate in its chassis ^[3]. |  |  |  |
| [REE Rare earths](https://www.materialinsights.org/commodities/rare_earths/) | Functional | 0.015 kg | – |
| 14.63 g of neodymium in the JRC's example two-socket rack server, 14.42 g of it in hard-disk magnets ^[3]; servers with solid-state drives hold less. |  |  |  |
| [Co Cobalt](https://www.materialinsights.org/commodities/cobalt/) | Functional | 0.009 kg | – |
| 9.27 g in the JRC's example two-socket rack server, 9.06 g of it in a lithium polymer battery ^[3]. Lenovo says no cobalt is expected in the battery types of its SR650 V3 ^[7]. |  |  |  |

as of 2026-11-16

#### Cost breakdown

The share of the unit's cost in each component or stage

| Part | Share of cost | Variant | Year | Source |
|---|---|---|---|---|
| Memory, 1 TB of DRAM (SemiAnalysis: 'nearly 40%') | 40% | General-purpose | 2023 | SemiAnalysis estimate for a two-socket CPU server costing about USD 10,424 to a large buyer ^[32] |
| NAND flash storage | 14.7% | General-purpose | 2023 | SemiAnalysis estimate, same server ^[32] |
| GPUs (cost of goods) | 59.1% | AI accelerator | 2023 | Calculated: 26.6 of 45.0 thousand USD, analyst cost stack for a 2023 DGX H100 (Silicon Analysts, from Epoch AI and SemiAnalysis estimates; manufacture cost before vendor margin, not a price) ^[33] |
| Networking | 17.8% | AI accelerator | 2023 | Calculated: 8 of 45.0 thousand USD, same analyst stack (an estimate) ^[33] |
| Power and cooling | 8.9% | AI accelerator | 2023 | Calculated: 4 of 45.0 thousand USD, same analyst stack (an estimate) ^[33] |
| Memory, storage and chassis | 14.2% | AI accelerator | 2023 | Calculated: 6.4 of 45.0 thousand USD, same analyst stack (Silicon Analysts' own estimate) ^[33] |

as of 2026-11-16

#### How it is put together

The main steps from parts and materials to the finished unit

| Step | What happens | Source |
|---|---|---|
| Make the parts | Steel chassis, motherboard, power supplies, fans, drives and heat sinks come from separate makers; in the JRC example server the steel chassis alone weighs 12.265 kg. | ^[3] |
| Build the barebones (L6) | The motherboard, chassis, power supply and cooling parts are fitted together into a barebones server. | ^[34] |
| Add processors, memory and storage | Processors, memory, storage devices, accelerators, network cards and other expansion parts are added in the stages after the barebones, to the buyer's configuration. | ^[34] |
| Configure and test the server (L10) | The barebones is configured as a complete, working server and validated. | ^[34] |
| Integrate into racks and clusters (L11, L12) | Many servers are joined into a rack with power, networking, cabling and cooling, and racks into a cluster, before shipping to the data centre. | ^[34] |

as of 2026-11-16

#### Technologies inside it

The technologies inside the product, with how much of each per unit
 Embodies 6

- [CPU Microprocessor](https://www.materialinsights.org/technologies/microprocessor/) 2 CPUs General-purpose Two CPUs in a two-socket server ^[1].
- [RAM Dynamic random-access memory](https://www.materialinsights.org/technologies/dram/) General-purpose Main memory modules; sixteen 16 GB modules in a standard Lenovo SR630 V3 test configuration ^[6].
- [Fl Flash memory](https://www.materialinsights.org/technologies/flash_memory/) 2 drives General-purpose Two 1.92 TB solid-state drives in the Dell R760 market configuration ^[1].
- [HDD Hard disk drive](https://www.materialinsights.org/technologies/hard_disk_drive/) 4 drives General-purpose Four hard disk drives (1.748 kg) in the JRC example rack server ^[3]; newer servers often use solid-state drives instead.
- [PE Transformers and power electronics](https://www.materialinsights.org/technologies/e_power_electronics/) 2 power supplies General-purpose Dual 1100 W power supplies in the R760 ^[1]; the converters reach about 96% efficiency at half load ^[6].
- [GPU Graphics processing unit](https://www.materialinsights.org/technologies/graphics_processing_unit/) 8 GPUs AI accelerator Eight H100 GPUs in a DGX H100 ^[8].

as of 2026-11-16

#### Sources

1. [Life Cycle Assessment of Dell Server PowerEdge R760 (report produced September 2025)](https://www.delltechnologies.com/asset/en-us/products/servers/briefs-summaries/poweredge-r760-life-cycle-assessment.pdf) Dell Technologies, 2025
2. [Analysis of material efficiency requirements of enterprise servers (EUR 27467 EN), Talens Peiró and Ardente, September 2015](https://publications.jrc.ec.europa.eu/repository/bitstream/JRC96944/lb-na-27467-en-n%20.pdf) European Commission Joint Research Centre, 2015
3. [Lenovo Server Product Recyclability Assessment (file dated December 2019; the document shows no date)](https://static.lenovo.com/ww/docs/sustainability/server-recyclability-assessment-dec-2019.pdf) Lenovo
4. [Recyclability Assessment of Supermicro Server Products Based on IEC TR 62635 Standards (SuperServer SYS-512B-WR)](https://www.supermicro.com/GreenComputing/EPEAT/Product_RecyclabilityAssessment_SYS-512B-WR.pdf) Supermicro
5. [Lenovo ErP Lot9 Information Sheet, Servers & Storage Products: ThinkSystem SR630 V3 (Annex B2, issued 10 January 2023, updated 1 October 2023)](https://cloud.ibm.com/media/docs/downloads/environment-info-datasheets/eco-thinksystem-sr630-v3.pdf) Lenovo, hosted by IBM Cloud, 2023
6. [Technical Documentation for EU Regulation 2019/424 laying down ecodesign requirements for servers and data storage product: ThinkSystem SR650 V3](https://cloud.ibm.com/media/docs/downloads/environment-info-datasheets/SR650v3-Lot9.pdf) IBM Corporation, hosted by IBM Cloud, 2023
7. [Introduction to NVIDIA DGX H100/H200 Systems (last updated 17 October 2026)](https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html) NVIDIA, 2026
8. [Product Carbon Footprint (PCF) Summary for NVIDIA HGX H100 (July 2025)](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf) NVIDIA, 2025
9. [AI Server Cost Analysis – Memory Is The Biggest Loser (Patel and Wong, 29 May 2023)](https://newsletter.semianalysis.com/p/ai-server-cost-analysis-memory-is) SemiAnalysis, 2023
10. [AI Server System BOM & Rack Cost: Historical Data & Chart (newest print March 2026)](https://siliconanalysts.com/market-data/ai-server-system-bom) Silicon Analysts, 2026
11. [What Are Server Manufacturing Levels? Understanding L6, L10, L11, and L12 Integration](https://www.gigabyte.com/glossary/server-manufacturing-levels) Gigabyte

## Other sections

- [Summary](https://www.materialinsights.org/products/p_data_centre_server/index.md)
- [Working](https://www.materialinsights.org/products/p_data_centre_server/prod-working/index.md)
- [Use case](https://www.materialinsights.org/products/p_data_centre_server/prod-use/index.md)
- [Supply chain](https://www.materialinsights.org/products/p_data_centre_server/prod-supply/index.md)
- [Footprint](https://www.materialinsights.org/products/p_data_centre_server/prod-footprint/index.md)
- [Evolution](https://www.materialinsights.org/products/p_data_centre_server/prod-evolution/index.md)
- [Interventions](https://www.materialinsights.org/products/p_data_centre_server/prod-interventions/index.md)
