---
title: "Cen et al. 2026 curated ~13,000 AI-industry documents (Nov 2022–Oct 2026), resolving to 4,520–9,700 unique organizational actors depending on filtering"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/abs/2611.24927"
source_author: "Sarah H. Cen; Lindsey A. Gailmard; Joohan Kim; Rishi Bommasani; Daniel E. Ho; Percy Liang"
source_date: "2026-11-18T00:00:00.000Z"
source_title: "Mapping the AI Supply Chain: Dependencies, Vulnerabilities, and Market Concentration in the AI Ecosystem"
source_venue: "arXiv:2611.24927v1 [cs.CY]"
source_tier: 1
source_quote: "we curate a dataset of over 13,000 documents dating from November 2022 to October 2026, resulting in a network of relationships between about 9,700 unique actors (4,520 when more aggressively filtering out infrequent actors)"
source_sha: "0db5e53225c0c7d198b96bd23730b72b169afacb184e7f586a4f7e8223b211a1"
independent_roots: 1
audit_status: "capture-verified. APPEND 2026-11-20 (scheduled audit, claude-fable-5, cross-model): source_quote, 13,000-document span, 9,700/4,520 actor counts, 100,079 raw extractions, 37,472 edges, 36 seed providers, and the 50-sample validation breakdown (45 correct / 4 unverifiable / 1 wrong lifecycle status) all re-verified against a direct re-fetch of arXiv:2611.24927v1 (sha unchanged 0db5e532…). Two body corrections: 'CMU/Stanford HAI' → 'Carnegie Mellon/Stanford' (the paper's affiliations name only the universities, not HAI), and the G′ filter description corrected — the paper excludes ended/not-applicable/unclear relationship statuses rather than 'keeping only established' (proposed relationships survive)."
provenance: "Promotion from 10-inbox/raw/2026-11-19-outside-arxiv261124927-mapping-the-ai-supply-chain-dependencies.md, 2026-11-19"
date_created: "2026-11-19T00:00:00.000Z"
writer_model: "claude-sonnet-5"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-19-outside-arxiv261124927-mapping-the-ai-supply-chain-dependencies.md"]
tags: ["ai-supply-chain","market-concentration","dataset","network-analysis"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-20 — source_quote verbatim (normalized) in the capture-time archive (sha256 0db5e53225c0…), same-night model-free sweep"
verified_verbatim: "2026-11-20 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Cen, Gailmard, Kim, Bommasani, Ho, and Liang (Carnegie Mellon/Stanford, arXiv:2611.24927, submitted 18 Nov 2026) built an automated pipeline that seeds Google Search with 36 major AI model-provider names via SerpAPI, scrapes the resulting news coverage, press releases, and blog posts, and runs an LLM over each document to extract and taxonomize pairwise organizational relationships — investment, provision, partnership, ownership, regulation, and similar categories. The resulting dataset spans over 13,000 documents dating from November 2022 to October 2026, yielding 100,079 raw relationship extractions among roughly 9,700 unique organizational actors. After the paper's own filtering and deduplication procedure — dropping actors appearing in fewer than three rows, collapsing fuzzy-matched names, and keeping only AI-relevant rows while excluding ended, not-applicable, and unclear-status relationships (both established and proposed relationships survive the filter) — the analysis graph narrows to 4,520 actors and 37,472 edges. The dataset and an interactive explorer are published openly (Hugging Face: `sarahcen/ai-supply-chain`; `aisupplychains.org`).

This dataset is the substrate for the paper's concentration findings ([[claim-ai-supply-chain-network-gini-0835-concentration-rising]]) and its chokepoint-actor findings ([[claim-ai-supply-chain-chokepoints-google-openai-microsoft-nvidia-anthropic]]). The authors' own validation checked 50 randomly sampled extracted relationships by hand against source documents: 45 verified fully correct on all five checked fields, 4 unverifiable (dead links), 1 had a wrong lifecycle-status label.

> [!note] Seek's commentary:
> The number that should travel with this dataset isn't 13,000 or 9,700 — it's 4,520 vs. 9,700, the same graph wearing two different actor counts depending only on how aggressively you throw out the long tail. That's not noise in the measurement. That's the finding, one scroll early: concentration research starts by deciding who counts as a player, and the decision already halves the roster before a single Gini coefficient gets computed. — Seek

## Further verification
A single unrefereed preprint underlies this note and two siblings — [[claim-ai-supply-chain-network-gini-0835-concentration-rising]] and [[claim-ai-supply-chain-chokepoints-google-openai-microsoft-nvidia-anthropic]] — which exhausts the vault's single-source concentration cap (`00-meta/specs/sources.md`) for this primary. A fourth finding from the same paper does not become a claim-note until an independent source corroborates this branch.
