Machine Intelligence Infrastructure
The global physical supply chain behind artificial intelligence
View
Follow the Chain
Pick a material and watch it travel from origin to your screen.
Chokepoint Atlas
Five lenses on supply-chain fragility — pick one to highlight every node in that category.
Layers
Click to solo · Shift-click to combine
Filter by country
About this map
Every time you ask an AI chatbot a question, you set in motion a chain of dependencies that spans 15 countries, more than 100 companies, and stretches from 380-million-year-old quartz deposits to fiber-optic cables on the ocean floor.
This is a living atlas of the complete physical supply chain behind artificial intelligence — 124 nodes, 228 typed dependencies, 384 cited sources, and 392 real-world sites. Each dot is a critical stage. Each arc is a dependency. Click any node to read its story, sources, and confidence rating — or zoom in, since every node now sits at its real headquarters or facility.
What you're seeing
Ten layers stack from the ground up: raw materials at the bottom, AI services at the top. Each layer feeds the one above it. The map currently covers 124 nodes across 15 countries, with 35 chokepoints — stages where a single company, country, or site controls most of global supply.
Most nodes sit at a real headquarters or facility — zoom in to land on the rooftop. A few are deliberately not a single place: AI Assistants (where the whole chain converges), the frontier-lab and cloud groupings, the power grid, and submarine cables. These wear a soft cyan glow to show they stand for many locations at once.
Every claim is sourced. Hover any node in the panel to see its citations, last-verified date (YYYY-MM), and confidence rating (high / medium / low) based on how many independent sources back it up. The full methodology — data schema, source rubric, risk model, edit workflow — is covered in the sections below.
How to use
- Play — animates the supply chain building up layer by layer, from raw materials to AI, with connections drawing in as each stage appears.
- Labels — toggle node names on the map.
- Connections — links stay hidden by default to keep the map legible; hover or click any node to reveal just its connections, or toggle Connections to show every dependency at once.
- Flow — animates link directions to show the direction of material and data flow.
- Chokepoints — pulsing red rings mark stages where a single company or region controls most of global supply.
- Top 20 Risks — highlights the twenty highest-risk nodes by composite fragility score (see the Risk model section below).
- Cables — submarine fiber-optic cables carrying ~99% of international internet traffic.
- Data Centers — the 96 frontier AI data centres in Epoch AI’s Frontier Data Centers hub (CC BY 4.0), the same set as the Build-out view and refreshed with it every month. A diamond’s area is proportional to its IT power today; paler diamonds are not yet running. Click one to open it on the Build-out timeline.
- Stories — eight amber pins on places where a single dot tells a big chunk of the chain (Carajás, Spruce Pine, Mariupol+Odesa, Bayan Obo, Dammam, Abilene Stargate, Kolwezi, Veldhoven).
- Crowded places — zoomed out, hubs that share a region gather in a small flower, with raw materials at its heart and AI on its rim. Hover to open it up and pick a hub. As you zoom in, each dot glides to its true location and stays there; hubs at the same address remain side by side, tied to their shared spot by a hairline.
- Sites — ~400 real-world locations the chain actually runs on (fabs, plants, mines, R&D offices, HQs), colored by chain layer. Click any site to open its parent node.
- Chip Tiers — the three country tiers of the US Framework for Artificial Intelligence Diffusion (90 FR 4544, 15 January 2025). Tier 1 is the United States and the 18 partners in Supplement No. 5 to Part 740; Tier 3 is Country Group D:5 plus Macau; Tier 2 is everyone else. The Department of Commerce withdrew the plan on 13 May 2025, two days before compliance was due, so the Tier 2 quotas never applied. Its text was never formally removed and still sits in the regulations (15 CFR 740.27). The Tier 3 bans were not new: they come from the October 2022 and October 2023 advanced-computing rules (87 FR 62186, 88 FR 73458) and remain in force. Hover a country for what has changed since, such as the UAE’s licence-free access from July 2026 and Cambodia leaving the arms-embargo group in February 2026. Tier sizes in chips are RAND’s conversions (Heim, PE-A3776-1).
- Diversion Cases — routes of AI chips allegedly sent to China without a US licence, drawn only where a court filing names the route, as published by the US Department of Justice: one guilty plea (Hao Global, Houston, 2025), two indictments (Tampa via Malaysia and Thailand, 2025; City of Industry via Malaysia and Singapore, 2026) and one criminal complaint (El Monte via Singapore and Malaysia, 2025). Line weight shows how far the case has gone. An indictment or complaint is an allegation, and every defendant is presumed innocent unless proven guilty; case status is as published on the date checked (11 November 2026). Where a filing names a country but no city, the line ends at the capital. We do not draw press reports, investigations without charges or allegations by private parties. Singapore’s February 2025 server case is left out because it was charged as fraud against suppliers, and the Singapore government said it is unrelated to US export controls. Since 14 July 2025 Malaysia requires its own Strategic Trade Permit to export or transship US AI chips. Renting computing power abroad is a different matter from moving chips, and is often legal; the Data Centers layer shows who Epoch AI identifies as each site’s users.
- Export Controls — highlights nodes subject to US, China, Japan, or Dutch export controls.
- Edge Types — recolors links by what flows through them: material, equipment, design-IP, power, data, or service.
- Search — type any node name in the search box at the top of the panel.
- Dotted terms — underlined acronyms in node descriptions have inline definitions; hover to read them.
The ten layers
The chain is stratified vertically. Toggle any layer in the side panel to isolate it; use the × next to Layers to clear them all.
- −3 · Raw Materials — quartz, iron ore, copper, germanium, rare earths, sugarcane, tin.
- −2 · Energy & Water — grids, nuclear, gas turbines, transformers, cooling, ultra-pure water.
- −1 · Chemistry & Refining — polysilicon, neon gas, fluorochemistry, photoresists, industrial gases.
- 0 · Processed Materials — silicon wafers, photoresists, ABF film, copper foil, fiber cable.
- 1 · Parts & Design — substrates, photomasks, EDA software, ARM, laser chips.
- 2 · Manufacturing Machines — ASML EUV scanners, Zeiss mirrors, TRUMPF lasers, etch & deposition tools.
- 3 · Chip Making & Assembly — TSMC, Samsung, Intel foundries; CoWoS advanced packaging; HBM stacking.
- 4 · Chips & Networking — NVIDIA, AMD, Broadcom, custom ASICs, switches, submarine cables.
- 5 · Cloud & Data Centers — Microsoft Azure, AWS, Google Cloud, Oracle, CoreWeave.
- 6 · AI Services — OpenAI, Anthropic, Google, Meta, xAI — and the assistants you use.
A parallel Chinese chain (SMIC, YMTC, CXMT, Naura, AMEC, Huawei HiSilicon …) is included — toggle Chinese Chain under Follow the Chain to isolate it.
Risk model & Top 20
Each node carries a composite risk score from 0 to 100, computed transparently from a fixed rubric. The Top 20 Risks view highlights the twenty most fragile nodes; every detail panel shows the breakdown so any score is fully auditable.
| Signal | Weight |
|---|---|
| Bottleneck flag | +30 |
| Sole supplier (atlas) | +25 |
| Geographic concentration | +15 |
| SME oligopoly | +10 |
| Capacity-constrained | +10 |
| Trade-war flashpoint | +10 |
| Subject to export controls | +5 |
| Downstream fan-out (log-scaled) | 0 to +15 |
| Low confidence | −5 |
Signals are not mutually exclusive — a node like ASML EUV picks up bottleneck + sole + geographic + trade, then gets a downstream-fan-out bonus because almost every advanced chip depends on it.
Chokepoint Atlas
Five lenses on supply-chain fragility — pick one to highlight every node in that category:
- Sole Supplier — if this one company stops, the chain breaks. No alternative exists (ASML EUV, Zeiss EUV optics, Shin-Etsu photoresist…).
- Geographic — one country or one site controls supply (Taiwan for leading-edge logic, China for refined rare earths, Spruce Pine NC for high-purity quartz).
- SME Oligopoly — four companies (Applied Materials, Lam, TEL, KLA) control most non-EUV semiconductor equipment.
- Capacity — demand outpaces production; backlogs measured in years (HBM, CoWoS, large transformers).
- Trade Wars — subject to export controls between the US, China, Japan, and the Netherlands.
Follow the chain
Pick a material and watch it travel from origin to your screen. The map dims everything else and highlights only the nodes and links on that single chain.
- Full AI Chain — the entire path from raw materials to AI services.
- Quartz → Chips — Spruce Pine, NC quartz becomes silicon ingots and then wafers.
- Sugar → GPUs — Brazilian sugarcane becomes ABF film inside NVIDIA GPUs.
- Neon → Chips — Ukrainian steel mills supply the neon gas inside EUV lasers.
- Tin → Light — molten tin droplets, vaporized 50,000 times per second, create the EUV photons that print chip circuits.
- Water → Chips — 10 million gallons a day, purified to parts-per-quadrillion.
- Germanium → Internet, MSG → AI, Sand → Internet, Copper → Cloud, Steel → Power, Chinese Chain.
Stories — eight places
Eight curated geographic pins surface the cartography in narrative form. Each pin is a place where a single dot tells a big chunk of the chain. Toggle Stories in the View controls to drop the pins, then click any to read.
- Carajás, Brazil — the world's largest iron mine, in the eastern Amazon.
- Spruce Pine, North Carolina — the quartz monopoly almost no one knows about.
- Mariupol & Odesa, Ukraine — the steel mills that were the world's neon supply.
- Bayan Obo, Inner Mongolia — the mine behind China's rare-earth dominance.
- Dammam, Saudi Arabia — the oil that fuels the grid that fuels the chips.
- Abilene, Texas (Stargate) — the new gigawatt-class AI data-center build-out.
- Kolwezi, DR Congo — the cobalt belt powering every laptop and EV battery.
- Veldhoven, Netherlands — the small Dutch town that builds every EUV scanner on Earth.
Sites overlay — 394 real locations
Each chain node is an aggregate. Toggle Sites to see the 394 individual fabs, plants, mines, R&D offices, and HQs that aggregate up. Sources: SemiAnalysis Global Fabs Database (161 fabs), GLOBAÏA-curated manual sites (214), and HQ fallback for the rest. Dots are colored by their parent layer.
This is the view to switch on if a node like NVIDIA or ASML feels too “Silicon Valley.” You'll see the actual geography — Israel, India, Singapore, Taiwan, Germany — that the headline name hides.
Data & methodology
The map's data is in JSON, not code — three canonical files in public/data/ai-supply-chain/: nodes.json, edges.json, sources.json. Each node carries layer, country, coordinates, a sourced description, a list of source IDs, a last-verified date, and a confidence rating.
Confidence rubric
Auto-assigned by source count: high (4+ independent sources), medium (2–3), low (0–1). The detail panel surfaces this as a colored dot next to the verified date.
Source quality
Sources are prioritized in order: primary company filings → government statistical (USGS, EIA, IEA) → peer-reviewed papers → think tanks (CSET, RAND, CSIS, OECD) → specialist trade press (SemiAnalysis, TeleGeography, TrendForce) → secondary news. Wikipedia is avoided except for stable definitional content.
Edge types
Every connection has a type: material (default), equipment, design-IP, power, data, service. Toggle Edge Types in the View controls to recolor links by what flows through them.
Sites & cartographic exports
The site overlay is built from a 1,213-fab SemiAnalysis dataset filtered to the chipmakers covered, plus 214 GLOBAÏA-curated manual sites and HQ fallbacks. A GIS bundle (GeoJSON + Esri shapefile) is produced for downstream projects but kept private to the staging folder.
References
- Chris Miller, Chip War: The Fight for the World's Most Critical Technology, Scribner, 2022
- International Energy Agency, Energy and AI, IEA, 2025
- OECD, Competition in the AI Infrastructure Stack, OECD, 2025
- OECD, Mapping the Semiconductor Value Chain, OECD Science, Technology and Industry Policy Papers, 2024
- International Energy Agency, Global Critical Minerals Outlook, IEA, 2024
- Semiconductor Industry Association (SIA), various reports and data, semiconductors.org
- Epoch AI, Nvidia B200 Cost Breakdown, 2025
- TrendForce, HBM market share and semiconductor industry reports, 2025–2026
- SemiAnalysis, Global Fabs Database, 2025 (1,213 fabs worldwide)
- ASML, annual reports and investor presentations, 2024–2025
- TSMC, corporate social responsibility and earnings reports, 2023–2025
- Wood Mackenzie, Power Transformer Supply Outlook, 2025
- Vertiv, CoreWeave, NVIDIA, AMD, Broadcom — quarterly earnings reports and investor filings, 2025–2026
- TeleGeography, Submarine Cable Map, 2025
- U.S. Geological Survey, Mineral Commodity Summaries: Germanium, 2025
- John VerWey, Tracing the Emergence of Extreme Ultraviolet Lithography, CSET, 2024
- Pilz, Mahmood, Heim, AI Power Requirements, RAND, 2025
- Lennart Heim, US AI Diffusion Framework, RAND, 2025
- Scaling Limits to AI Chip Manufacturing, ICML, 2025
- Andre Barbe and Will Hunt, Preserving the Chokepoints, CSET, 2022