AI Supply Chain

A directed graph of ~160 publicly-traded companies across the AI supply chain, in 20 stages. Flow runs left → right, supplier to customer. Each node is sized and coloured by its gross margin; the slider scrubs a monthly timeline. Hover a node to trace its links; click to pin its detail and capacity proxy. Nodes with a gold dashed ring are non-SEC foreign filers whose margin is sourced externally (see methodology). The graph is wide — scroll it sideways on a narrow screen, or use the searchable table below.

Methodology

• Nodes: ~160 public companies, each placed by role into one of 20 stages (equipment & materials, EDA & IP, foundry, accelerators & custom silicon, analog/power/broad semis, memory & storage, optics & interconnect, connectors & cabling, networking, ODM/assembly, servers & systems, power & cooling, power generation & energy, datacenter real estate, cloud/hyperscalers, AI software).
• Encoding: node size = gross margin (%). Default colour = gross margin; the toggle recolours by capacity growth (year-over-year change in net PP&E).
• Gross margin: GrossProfit / Revenue, or (Revenue − CostOfRevenue) / Revenue where GrossProfit isn't tagged. Primary source: SEC sec_company_facts, reduced in ClickHouse and served from /api/aichain.
• Cadence: US filers contribute quarterly points (10-Q); 20-F/40-F filers (TSMC, ASML, GlobalFoundries, UMC, Nova, Camtek, Nebius, Cameco, Alibaba…) don't tag three-month figures, so they're annual. The client linearly interpolates each company between its own filings along a continuous monthly axis.
• External (non-SEC) nodes — gold dashed ring: ~35 companies across Japan, Korea, Taiwan, Europe and Australia that don't file with the SEC (SK Hynix, Samsung, Tokyo Electron, Lasertec, Shin-Etsu, SUMCO, Disco, Ibiden, Murata, TDK, Infineon, BE Semi, Foxconn, Quanta, Lynas, …). Their gross margin is hand-sourced from company IR / stockanalysis.com as a static snapshot; each node's tooltip/detail shows its "data as of" date. No SEC capacity (PP&E/capex) data is available for these.
• Concept coalescing: revenue and cost tags are merged across variants (IFRS Revenue/CostOfSales; the modern finance-lease PP&E tag; Amazon's PaymentsToAcquireProductiveAssets).
• Capacity proxy: net property, plant & equipment. Expected capacity: capex (annual).
• Edges (supplier → customer): ~420 links, rebuilt from real disclosures. Each company's latest SEC 10-K / 20-F was parsed for its customer-concentration, major-customers and business sections — about 100 edges are named directly in filings (e.g. Arista's >10% customers, Fabrinet's customer base, CoreWeave's concentration). These are supplemented with well-documented public relationships (ASML→TSMC, TSMC→Nvidia, the memory makers→Nvidia/AMD, hyperscaler custom silicon fabbed at TSMC), and a small remainder are sector-level structural links. Direction is the real trade flow, so memory, optics, power and packaging-equipment legitimately point "back" upstream into their accelerator/foundry customers.

Assumptions & caveats

• Edges are grounded in SEC filings and public disclosures, but are not a complete procurement/contract ledger: dollar amounts aren't shown, anonymised "Customer A" disclosures can't be resolved to a name, and a minority of links are sector-level rather than a specific named account.
• Gross margin is GAAP/IFRS and currency-free (a ratio), so it is comparable across €/NT$/¥/₩ filers; PP&E and capex are in the filer's reporting currency and are compared cross-company only as growth %.
• External (gold-ring) figures are a manually-sourced static snapshot as of the date shown on each node — not live, and not from SEC filings.
• Capacity figures are financial proxies (PP&E, capex), not unit capacity (wafers, megawatts).
• Gray nodes have no gross-margin data: utilities and REITs (Constellation, Vistra, Talen, Digital Realty) don't report a gross margin; Cadence / HPE don't tag a clean total cost-of-revenue; Oklo is pre-revenue. They still carry their chain links (and, where available, capacity).
• Gross margin means different things across sectors — a fab-equipment maker, a thin-margin assembler, and a software company aren't directly comparable on it.
• Values reflect each company's latest filing; between and after reports they are interpolated or held flat (flagged when extrapolated past the last filing).
• Per-stage figure in the graph is the median gross margin of that stage's companies at the selected month.