Aggregated view of the AI compute supply chain — fabs, chips, packaging, datacenters, and market signals.
All market size estimates, capacity figures, and bottleneck assessments are illustrative demo data from public sources (earnings calls, analyst reports, company presentations). Actual figures are commercially sensitive and may differ materially. Confidence levels shown per data point. Not for investment, trading, or business decisions.
2,950 MW
combined AI datacenter capacity [Demo est.]
$31.1B
announced datacenter investment [Demo est.]
$204.7B
announced fab investment [Demo est.]
Est. AI Accelerator Market (2025)
MED$115–145B
FP8 AI training GPU market. Not confirmed.
[Composite analyst estimates — DEMO only]TSMC AI Revenue Share (2024 est.)
MED~25–30%
% of TSMC revenue from AI/HPC. Estimates only.
[TSMC Q4 2024 Earnings / Analyst]CoWoS Capacity Growth Target
LOW~2.5–3x
2023→2025 capacity expansion target. Not confirmed.
[TSMC capacity expansion plans — analyst est.]Global AI DC Power Demand (2025 est.)
LOW~50–80 GW
Total AI datacenter power. Highly uncertain.
[IEA / Goldman Sachs estimates — DEMO]Taiwan
HIGH CONCENTRATIONTSMC's leading-edge fabs (N3, N2) are concentrated in Taiwan. ~90% of sub-5nm global capacity [est.]. Geopolitical concentration risk.
[Demo analysis — illustrative only]South Korea
MEDIUM CONCENTRATIONSamsung Foundry advanced nodes, plus SK Hynix HBM3e production. HBM supply concentration for AI GPUs.
[Demo analysis — illustrative only]United States
LOW CONCENTRATIONCHIPS Act investments building out Intel 18A, TSMC Arizona, Samsung Taylor. Diversifying from Asia over 2025–2030.
[Demo analysis — illustrative only]China
MEDIUM CONCENTRATIONSMIC and domestic fabs constrained to ~7nm-class DUV. US export controls limiting EUV access. DRAM/NAND capacity significant.
[Demo analysis — illustrative only]| Supply Chain Layer | Key Item | Severity | Intelligence Note |
|---|---|---|---|
| Process Node | TSMC N3/N2 Capacity | HIGH | N3E and N2 demand from Apple, NVIDIA, AMD, and hyperscaler custom silicon competing for limited leading-edge wafers. [est.] |
| Advanced Packaging | CoWoS-L (TSMC) | CRITICAL | Key bottleneck for H100/H200/Blackwell GPU production 2023–2025. Single-source dependency (TSMC only). Capacity ramping. [est.] |
| Memory | HBM3e (SK Hynix) | HIGH | HBM3e supply allocated primarily to NVIDIA Blackwell. Micron and Samsung ramping but SK Hynix dominant. [est.] |
| Substrates | ABF Substrate Supply | MEDIUM | Advanced ABF (Ajinomoto Build-up Film) substrates for AI GPU packaging. Cyclical supply-demand dynamic. [est.] |
| Power | Datacenter Power Capacity | MEDIUM | AI datacenter power demand growing rapidly. Grid connectivity, cooling, and permitting becoming constraints in key US/EU regions. [est.] |
Supply Chain Dependency Stack
AI Chip Designers
NVIDIA, AMD, Google, Amazon, Tesla, Cerebras, Groq
Foundry (Dominant)
TSMC (N4P, N5, N3E, N2) — ~90% of AI accelerators [est.]
Advanced Packaging
TSMC CoWoS-L — primary AI GPU packaging. SK Hynix HBM3e.
Hyperscale Deployers
Microsoft, Google, AWS, Meta, xAI, Oracle
Key Relationships
NVIDIA Blackwell → TSMC N4P → CoWoS-L → SK Hynix HBM3e
AMD MI300X → TSMC N5/N6 → CoWoS-S → HBM3
Google TPU v5 → TSMC N4/N5 → Custom packaging
Amazon Trainium2 → TSMC N4 → Custom AWS packaging
Intel 18A → Intel Fab 52/62 (Chandler AZ) → Foveros
[All chains are illustrative est. based on public reports]
* All figures are illustrative estimates from public sources. Market sizes, capacity figures, and bottleneck assessments are approximated for demo purposes. Confidence levels indicate data quality. Not for investment, financial, or business decisions.