Capex Flow Map
Two-stage Sankey of AI-attributable hyperscaler capex routed through infrastructure layers to specific suppliers. Layer mix uses an industry-research baseline; supplier shares within each layer reflect disclosed customer concentration. Every number on this page is a modeled estimate — see Methodology for the math and assumptions.
Total AI-attributable capex in scope: $370.4B
| Layer | Mix | $B | Rationale |
|---|---|---|---|
| Compute Silicon | 45.0% | $154.8B | GPUs + custom AI ASICs (~45% of AI capex per SemiAnalysis composition) |
| HBM & Memory | 10.0% | $11.1B | HBM3E/HBM4 stacks (typically bundled in GPU BoM but allocated here) |
| Networking & Optical | 8.0% | $29.6B | Switches + optical (back-end fabric ~7–9% of cluster cost) |
| Power & Electrical | 15.0% | $55.6B | Switchgear, UPS, busway, transformers — rising with density |
| Datacenter Build | 12.0% | $44.4B | DC shell + electrical EPC + land prep |
| Foundry & Equipment | 7.0% | $25.9B | Imputed via foundry/wafer cost back to AMAT/LRCX/ASML pass-through |
| Neoclouds | 3.0% | $10.2B | Hyperscaler take-or-pay to GPU clouds (MSFT↔CRWV style deals) |
Modeled dollar flow from Big 5 AI capex
| Supplier | Primary layer | $B | Share |
|---|---|---|---|
| NVDA NVIDIA | Compute Silicon | $111.8B | 30.2% |
| AVGO Broadcom | Compute Silicon | $39.0B | 10.5% |
| VRT Vertiv Holdings | Power & Electrical | $36.1B | 9.7% |
| ETN Eaton | Power & Electrical | $25.6B | 6.9% |
| PWR Quanta Services | Datacenter Build | $24.4B | 6.6% |
| ANET Arista Networks | Networking & Optical | $16.3B | 4.4% |
| GEV GE Vernova | Power & Electrical | $13.9B | 3.8% |
| AMD Advanced Micro Devices | Compute Silicon | $12.9B | 3.5% |
| TSM TSMC | Foundry & Equipment | $11.7B | 3.2% |
| MU Micron Technology | HBM & Memory | $11.1B | 3.0% |
| CRWV CoreWeave | Neoclouds | $10.2B | 2.7% |
| ASML ASML Holding | Foundry & Equipment | $5.2B | 1.4% |
| AMAT Applied Materials | Foundry & Equipment | $5.2B | 1.4% |
| CIEN Ciena | Networking & Optical | $4.4B | 1.2% |
| LRCX Lam Research | Foundry & Equipment | $3.9B | 1.1% |
It is not a revenue model. Hyperscaler capex is the spend; supplier revenue depends on shipment timing, mix, customer concentration, and competitive substitution. The flow map gives an upper-bound proxy for the addressable dollar base.
Internal accelerators are partially excluded. Google TPU is captured via AVGO (designer); Amazon Trainium and Microsoft MAIA are not in the supplier roster, so the "sum of shares" for AMZN and MSFT compute_silicon intentionally is <100% — the residual stays internal.
HBM is partial. Micron is the only HBM name in scope; SK Hynix and Samsung dominate. The shown $B for memory understates industry-wide HBM TAM by ~3×.