Digital

Servers and cloud hardware

Data center compute, storage, and networking hardware with advanced chip and power supply dependence.

Why it matters · Cloud service cost increases, AI infrastructure delays, and enterprise IT procurement backlogs.

6

Inputs

14

Companies

11

Facilities

10

Source countries

How it's made

The process

  1. 01

    Silicon Wafer Fab (CPU/GPU Die)

    TSMC N3/N4/N5 extreme ultraviolet (EUV) lithography; ASML EUV scanner required; 800+ process steps; 3-month cycle time per wafer lot; TSMC ~55% global advanced logic foundry market

  2. 02

    DRAM/HBM Wafer Fabrication

    Samsung, SK Hynix, Micron DRAM fabs; HBM adds through-silicon via (TSV) drilling and die stacking (4-12 dies per HBM stack); yield management for TSV is critical challenge

  3. 03

    Advanced Packaging (CoWoS/InFO)

    TSMC CoWoS (Chip-on-Wafer-on-Substrate) places GPU die and HBM stacks on silicon interposer; TSMC holds near-monopoly on CoWoS for AI chips; 18-24 month capacity ramp; ABF substrate at this stage

  4. 04

    Board and Module Assembly

    PCB fabrication (Taiwan, China); component SMT placement; DRAM modules assembled in Korea, Taiwan; NIC, SSD, power supply manufactured and tested

  5. 05

    Server System Integration (ODM)

    Quanta, Foxconn Industrial, Supermicro integrate CPU, DRAM, storage, NIC, GPU into server chassis; software/firmware loading; Taiwan and US assembly; hyperscalers use custom designs (ODM direct)

  6. 06

    Burn-in and Quality Testing

    Full server burn-in (24-72 hours under thermal stress); diagnostic testing; GPU cluster validation; NVIDIA NVLink mesh connectivity tested; failure rate identification

  7. 07

    Data Center Deployment

    Racked and stacked in data center; liquid cooling CDU connections (for AI GPU servers); power distribution, networking commissioned; hyperscalers deploy in 500-1,000 GPU clusters

Where it comes from

Country dependencies

Weighted share of upstream inputs sourced from each country.

CountryWeighted shareInputs supplied to servers and cloud hardware
TWTaiwan55%ABF Substrate (Ajinomoto Build-up Film) · AI Training GPU (H100/H200/B200 class) · Server CPU (x86 Xeon/EPYC)
KRSouth Korea41%ABF Substrate (Ajinomoto Build-up Film) · AI Training GPU (H100/H200/B200 class) · DDR5 DRAM Server Memory +1
JPJapan27%ABF Substrate (Ajinomoto Build-up Film) · DDR5 DRAM Server Memory · High-Bandwidth Memory (HBM3/HBM3E)
IEIreland15%Server CPU (x86 Xeon/EPYC)
USUnited States12%AI Training GPU (H100/H200/B200 class) · DDR5 DRAM Server Memory · High-Bandwidth Memory (HBM3/HBM3E) +1
SGSingapore6%DDR5 DRAM Server Memory
ILIsrael5%Server CPU (x86 Xeon/EPYC)
MYMalaysia5%Server CPU (x86 Xeon/EPYC)
ATAustria5%ABF Substrate (Ajinomoto Build-up Film)
CNChina2%DDR5 DRAM Server Memory

Who makes it

Supplier companies

14 companies in this supply chain, sorted by market share.

Taiwan Semiconductor Manufacturing Company (TWSE: 2330; founded 1987 by Morris Chang); the world's dominant pure-play foundry and the most critical single company in the global technology supply chain. TSMC manufactures approximately 90% of all sub-5nm logic chips globally, including every NVIDIA AI GPU (H100 on N4, H200 on N4X, B200 on N3E), every Apple processor (A18 on N3E, M4 on N3E), and every AMD data center chip. TSMC's N3 and N5 fabs are located in Taiwan — 100 miles from mainland China. Revenue: $93.7B (2024, +34% YoY). AI/HPC share of revenue: 57% (Q3 2025), up from 18% in 2019. N2 volume production began Q4 2025 at Fab 22 Kaohsiung — world's first 2nm logic node in production. TSMC Arizona Fab 21 Phase 1 (N4P) reached 92% yield and profitability in 2025. CHIPS Act recipient: $6.6B grant + $5B loan.

Supplies these inputs

AI Training GPU (H100/H200/B200 class)

Replaceability

Substitutability 2% · 60 mo to replace

Business segments

  • Advanced Logic Foundry (<7nm)65% rev
  • Specialty Technology (Mature Nodes)20% rev
  • Advanced Packaging (InFO, CoWoS)10% rev
  • Global Expansion Fabs5% rev
Ajinomoto Fine-Techno
HQ JP95% shareSOLE SUPPLIER

Subsidiary of Ajinomoto Group; sole commercial producer of ABF (Ajinomoto Build-up Film) used in ~95% of CPU and GPU substrates worldwide. ABF was invented from waste byproducts of MSG manufacturing.

Supplies these inputs

ABF Substrate (Ajinomoto Build-up Film)

Replaceability

Substitutability 3% · 60 mo to replace

Business segments

  • ABF (Ajinomoto Build-up Film) — Global Monopoly85% rev
  • Functional Materials (Electronic + Biomedical)15% rev
Ajinomoto Co.
HQ JP90% share

World's largest MSG and amino acid producer AND sole-source supplier of ABF (Ajinomoto Build-up Film) — the insulating substrate used in advanced CPU/GPU packaging for all leading chips

Supplies these inputs

ABF Substrate (Ajinomoto Build-up Film)

Business segments

  • Seasonings & Foods (MSG, Amino Acid)40% rev
  • Amino Science (Feed & Industrial)25% rev
  • Pharmaceutical Amino Acids15% rev
  • Electronics Materials (ABF)10% rev

NVIDIA Corporation (Nasdaq: NVDA; founded 1993 by Jensen Huang, Chris Malachowsky, Curtis Priem) is a fabless semiconductor company and the dominant supplier of AI accelerators. NVIDIA designs all chips in-house but outsources 100% of fabrication to TSMC Taiwan. H100 (Hopper) on TSMC N4; H200 on N4X; B100/B200/GB200 (Blackwell) on N3E. NVIDIA surpassed Apple as TSMC's largest customer in 2025 ($23.4B, 19% of TSMC revenue, +62% YoY). Revenue: $130B (FY2026 guidance). AI data center segment: ~88% of total revenue. NVIDIA's GPU monopoly in AI training (95%+ share of AI accelerator market) means that global AI infrastructure buildout is physically dependent on TSMC Taiwan fabs. A Taiwan Strait crisis that halted TSMC production would immediately freeze global AI infrastructure expansion.

Supplies these inputs

AI Training GPU (H100/H200/B200 class)

Replaceability

Substitutability 8% · 24 mo to replace

Business segments

  • Data Center (AI Accelerators)88% rev
  • Gaming GPU7% rev
  • Automotive & Robotics3% rev
  • OEM & Professional Visualization2% rev
Intel Corporation
HQ US63% share

Dominant x86 server CPU maker (Xeon Scalable); owns its own fabs (IDM model); also makes Gaudi AI accelerators; losing server CPU share to AMD EPYC; transitioning to hybrid IDM+foundry model

Supplies these inputs

Server CPU (x86 Xeon/EPYC)

Replaceability

Substitutability 30% · 12 mo to replace

Business segments

SK Hynix (HXSCL)
HQ KR62% share

SK Hynix Inc. (KRX: 000660; ~₩66T revenue) is the world's #3 NAND flash manufacturer with ~18% global market share — a position significantly strengthened by its $9B acquisition of Intel's NAND business (completed December 2021), which created the Solidigm subsidiary (formerly Intel NAND) and added the Dalian, China fab. SK Hynix's primary domestic NAND production is at the Cheongju M15X fab (North Chungcheong Province, Korea). SK Hynix is also the world's dominant HBM (High Bandwidth Memory) manufacturer, supplying all of NVIDIA's H100/H200/B200 memory.

Supplies these inputs

AI Training GPU (H100/H200/B200 class) · DDR5 DRAM Server Memory · High-Bandwidth Memory (HBM3/HBM3E)

Replaceability

Substitutability 12% · 12 mo to replace

Business segments

  • DRAM (World #2)45% rev
  • High Bandwidth Memory — HBM (Dominant Supplier)25% rev
  • NAND Flash (World #3)20% rev
  • Advanced Packaging + AI Infrastructure10% rev

Samsung Electronics Co., Ltd. (KRX: 005930; ~$200B revenue) is the world's largest NAND flash memory manufacturer with ~33% global market share. Samsung's NAND operations span the Pyeongtaek P3 complex (South Korea) and — critically — the Xi'an, Shaanxi, China facility, which is Samsung's largest single NAND fab by capacity. Samsung received a one-year BIS export control exemption (later extended) to continue upgrading Xi'an equipment under US restrictions. Samsung pioneered 3D NAND V-NAND technology in 2013 (first commercial 3D NAND) and now produces 200+ layer V-NAND. Also produces DRAM (world's largest) and logic chips (Exynos, foundry).

Supplies these inputs

DDR5 DRAM Server Memory · High-Bandwidth Memory (HBM3/HBM3E)

Replaceability

Substitutability 20% · 12 mo to replace

Business segments

  • NAND Flash Memory (World #1)28% rev
  • DRAM Memory (World #1)25% rev
  • Samsung Foundry (Logic Chips) + Exynos20% rev
  • Consumer Electronics (Smartphones, TVs, HVAC)20% rev

Rapidly gaining server CPU share with EPYC series (Milan/Genoa/Turin); fabless — manufactured by TSMC; also makes Instinct AI GPUs (MI300X); ~35% server CPU market share in 2024

Supplies these inputs

AI Training GPU (H100/H200/B200 class) · Server CPU (x86 Xeon/EPYC)

Replaceability

Substitutability 20% · 12 mo to replace

Business segments

Ibiden (IBIDF)
HQ JP25% share

Japan's leading manufacturer of high-end ABF flip-chip BGA substrates for Intel, AMD, and Nvidia CPUs and GPUs. Dominant player alongside Shinko Electric; both received Japanese government subsidies to expand capacity.

Supplies these inputs

ABF Substrate (Ajinomoto Build-up Film)

Replaceability

Substitutability 20% · 24 mo to replace

Business segments

  • FC-BGA Semiconductor Package Substrates (World #1)55% rev
  • Diesel Particulate Filters (DPF)20% rev
  • Graphite and Carbon Products15% rev
  • Printed Circuit Boards10% rev
Micron Technology
HQ US24% share

Only US DRAM and NAND flash manufacturer; Boise ID; ~22% global DRAM market share; also makes HBM3E for AI accelerators; critical US semiconductor company with CHIPS Act funding

Supplies these inputs

DDR5 DRAM Server Memory · High-Bandwidth Memory (HBM3/HBM3E)

Replaceability

Substitutability 18% · 12 mo to replace

Business segments

  • DRAM (Compute & Server)55% rev
  • NAND Flash & SSDs25% rev
  • High Bandwidth Memory (HBM)12% rev
  • Embedded & Automotive5% rev

Taiwan's largest IC substrate manufacturer; major supplier of ABF substrates for AI GPU packaging alongside Ibiden (Japan). Expanding capacity for AI chip demand.

Supplies these inputs

ABF Substrate (Ajinomoto Build-up Film)

Replaceability

Substitutability 20% · 24 mo to replace

Business segments

  • IC Substrates (ABF, Flip-Chip)55% rev
  • Standard PCB (Consumer, Automotive, Telecom)35% rev
  • Advanced Packaging Substrates10% rev

Major Japanese ABF substrate manufacturer; key supplier for Intel processors. Fuji Electric holds a significant stake. Alongside Ibiden, forms a near-duopoly on high-end CPU package substrates.

Supplies these inputs

ABF Substrate (Ajinomoto Build-up Film)

Replaceability

Substitutability 20% · 24 mo to replace

Business segments

ChangXin Memory Technologies (CXMT; Chinese: 长鑫存储; founded 2016; headquartered Hefei, Anhui, China) is China's primary state-backed DRAM manufacturer and the most significant attempt to break the Samsung/SK Hynix/Micron oligopoly. CXMT has received an estimated $15B in direct and indirect Chinese government subsidies as part of China's drive toward semiconductor self-sufficiency. As of 2025, CXMT is limited to 17nm DDR4 and LPDDR4 technology — approximately 2-3 process generations behind Samsung/Hynix/Micron's 1a/1b-nm class DRAM. CXMT cannot manufacture HBM of any generation: HBM requires Through-Silicon Via (TSV) stacking and advanced packaging techniques that require equipment and process knowledge CXMT does not possess. US export controls (October 2022, October 2023) target the semiconductor manufacturing equipment CXMT would need to advance to 10nm-class DRAM. CXMT's production primarily serves the domestic Chinese market for consumer electronics and is strategically significant as a pressure valve on global commodity DRAM pricing, but it does not threaten the HBM or advanced DRAM segments where the oligopoly's value is most concentrated.

Supplies these inputs

DDR5 DRAM Server Memory

Replaceability

Substitutability 30% · 24 mo to replace

Business segments

  • DDR4 DRAM (Desktop/Server)55% rev
  • LPDDR4/4X Mobile DRAM35% rev
  • Specialty & Emerging DRAM10% rev

ARM-based server CPU designer; founded 2017 by Renée James (former Intel President); Altra and AmpereOne processors used by Microsoft Azure, Google Cloud, Oracle Cloud; not yet profitable but growing rapidly as hyperscalers seek x86 alternatives.

Supplies these inputs

Server CPU (x86 Xeon/EPYC)

Replaceability

Substitutability 30% · 18 mo to replace

Business segments

  • Altra & AmpereOne ARM Server CPUs90% rev
  • Software & Ecosystem10% rev

Where it's made

Facilities

11 facilities producing inputs that feed servers and cloud hardware.

Ajinomoto Fine-Techno Kawasaki Plant

JP

Ajinomoto Fine-Techno · Kawasaki, Kanagawa, Japan · materials-manufacturing

Primary ABF film production facility. ~90-95% of all advanced CPU/GPU substrate film produced here. Capacity expansions announced: JPY 25B investment by 2030 to meet AI chip demand.

CXMT Hefei Fab — Domestic Chinese DRAM

CN

CXMT (ChangXin Memory Technologies) · Hefei, Anhui · fab

CXMT (ChangXin Memory Technologies) Hefei fab — China's primary DRAM manufacturing facility, producing approximately 100,000-120,000 wafer starts per month as of 2024-2025 (compared to ~1.2M wspm total global DRAM capacity). CXMT is limited to 17nm DDR4 and LPDDR4 — approximately 2-3 generations behind Samsung/Hynix/Micron's 1a/1b-nm class DRAM. The Hefei fab has received an estimated $15B in Chinese state subsidies through direct grants, tax incentives, and below-market equipment loans. US export controls (October 2022 and October 2023 rounds) specifically targeted semiconductor manufacturing equipment that would allow CXMT to advance beyond its current 17nm capability. CXMT cannot produce HBM of any generation. China's domestic DRAM represents approximately 5-8% of global commodity DRAM capacity by volume, but zero HBM capacity. Source: https://www.trendforce.com/news/2024/06/05/news-cxmt-expands-dram-output-but-remains-generationally-behind/

Ibiden Ogaki/Gifu Substrate Fabs

JP

Ibiden · Gifu · manufacturing

Ibiden's primary IC substrate manufacturing cluster in Ogaki and surrounding Gifu Prefecture; sole high-end ABF substrate supplier for NVIDIA AI GPU lineup.

Intel Hillsboro Oregon — Xeon Scalable CPU Production (Intel 4/Intel 3)

US

Intel Corporation · Oregon · manufacturing

Intel's primary US CPU manufacturing campus; Ronler Acres and D1X fabs; produces Xeon Scalable (Sapphire Rapids, Granite Rapids, Clearwater Forest) at Intel 4/Intel 3 nodes. Intel 18A process ramping 2025-2026 for next-gen Xeon.

Micron Boise (MU-1) — DRAM R&D and Production

US

Micron Technology · Boise, Idaho · fab

Micron's Boise, Idaho headquarters and primary US DRAM manufacturing and R&D facility — the only advanced DRAM fab on US soil operated by a US company. Micron's Boise fab (MU-1) focuses on leading-edge DRAM process development (1a/1b-nm class) and some volume production, but the majority of Micron's high-volume commodity DRAM production occurs in Singapore. Boise is strategically critical as the home of Micron's DRAM process technology IP. CHIPS Act: Micron received $6.1B in grants (April 2024) for domestic expansion at Boise and at a new greenfield fab in Clay, New York. The Clay NY fab (announced 2022) will be the largest DRAM fab ever built in the US, targeting 1b-nm class DRAM and HBM when fully ramped by 2030. Source: https://investors.micron.com/news-releases/news-release-details/biden-harris-administration-announces-6-billion-chips-act

Micron Hiroshima Fab — HBM / Advanced DRAM

JP

Micron Technology · Hiroshima · manufacturing

Primary Micron HBM production site; formerly Elpida Memory (Micron acquired Elpida in 2013 bankruptcy). Micron scaled HBM3E production here for NVIDIA H200; Japan CHIPS-equivalent subsidies supporting further investment.

SK Hynix Icheon M16 Fab — HBM and Advanced DRAM

KR

SK Hynix · Icheon, Gyeonggi-do · fab

SK Hynix's Icheon M16 fab — the world's primary HBM3e production facility and the most consequential DRAM fab for global AI infrastructure. M16 is SK Hynix's most advanced DRAM production line (1a/1b-nm class) and the site where HBM3e — the memory variant required by NVIDIA H100 and H200 AI accelerators — was first produced at volume. In 2023, all HBM3e chips in the world came from this facility; NVIDIA was on a multi-year waitlist. M16 also produces leading-edge DDR5 and LPDDR5X. Icheon campus capacity: M14 + M16 together represent approximately 280,000 wafer starts per month. SK Hynix is building M15X expansion and a new fab at Cheongju (NAND). Source: https://news.skhynix.com/sk-hynix-completes-the-construction-of-m16-fab-the-worlds-most-advanced-dram-production-facility/

Samsung Pyeongtaek P3 Fab — DRAM / HBM

KR

Samsung Electronics · Pyeongtaek, Gyeonggi-do · fab

Samsung Pyeongtaek P3 fab — the newest and most advanced DRAM/HBM production facility at Samsung's Pyeongtaek campus, the world's largest semiconductor manufacturing complex. P3 produces 1a-nm class DRAM (approximately 12-14nm effective node) and is the primary HBM3/HBM3e stacking line for Samsung's AI memory products. The Pyeongtaek campus covers approximately 2.89 million square meters — larger than some small cities. P4 construction began 2023 and targets HBM4 and next-generation DRAM for 2026 ramp. Samsung is investing approximately $230B in the Pyeongtaek campus over 2023-2042 under a government-supported K-Chips plan. Source: https://semiconductor.samsung.com/us/newsroom/news/samsung-electronics-and-the-government-of-south-korea-announce-300-trillion-won-semiconductor-investment-plan/

Shinko Nagano Substrate Plant

JP

Shinko Electric Industries · Nagano Prefecture · substrate_manufacturing

Primary manufacturing site for CPU/GPU flip-chip BGA substrates. Supplies Intel, AMD, and others. Key competitor to Ibiden in high-end substrate market.

TSMC Fab 18 (Tainan) — AMD EPYC CPU Production

TW

Taiwan Semiconductor Manufacturing Company (TSMC) · Tainan · manufacturing

AMD EPYC Genoa (N5) and Turin (N4P/N3) server CPUs fabbed exclusively at TSMC Fab 18; same fab as NVIDIA H100/H200 GPUs — AMD server CPU production competes with AI GPU production for TSMC allocation.

TSMC Fab 18 (Tainan) — N3/N4 Gigafab

TW

Taiwan Semiconductor Manufacturing Company (TSMC) · Tainan · fab

TSMC Fab 18, Southern Taiwan Science Park, Tainan — an 8-phase gigafab covering approximately 160,000 square meters (22 soccer fields). Primary production: N3 (3nm) and N4 (4nm EUV) nodes. Capacity: 150,000–160,000 wafer starts per month (wspm). This is the single most consequential industrial building on Earth for AI hardware — every NVIDIA H100, H200, and Blackwell GPU compute die was manufactured here. Fab 18 is located in an active seismic zone. April 3, 2024: M7.4 earthquake (Hualien) caused temporary disruptions to cleanroom operations; estimated 2–3 weeks partial recovery. December 2025: M7.0 earthquake — Taiwan's strongest in 27 years — Fab 18 escaped with minimal disruption due to extraordinary engineering safeguards. Source: https://www.usgs.gov/news/national-news-release/major-79-earthquake-strikes-taiwan