Mapping the AI Supply Chain:
An Analysis of the Complex Relationships in the AI Ecosystem

Motivation

While the creation and deployment of Artificial Intelligence (AI) spans nearly a century, there has been an explosion of AI products and services following the creation of LLMs. By enabling users to engage with AI through natural language, LLMs significantly lowered barriers to AI and broadened access, driving mainstream adoption.

The emergence of the AI supply chain is significant in its own right. Historically, AI and machine learning systems were largely developed in-house or by a few organizations. Teams collected their own data, trained their own models, and deployed their own infrastructure. The outsourcing of individual components (such as dataset collection) led to the rise of services like Amazon Mechanical Turk, Scale AI, and others. As the AI industry booms, specialized firms now handle distinct stages of the development pipeline. This fragmentation enables greater specialization and efficiency, signaling a maturing industry and creating new opportunities for innovation and competition.

Yet even as the AI supply chain becomes increasingly central to the AI ecosystem, it remains difficult to map and analyze. Market reports typically identify key organizations but do not describe the relationships and dependencies between them. This project addresses that gap by mapping the AI supply chain and examining how different actors interact. A clearer understanding of these relationships enables several critical analyses:

• Chokepoints and Industry Resilience: The AI sector depends on key inputs (such as compute, data, and talent). Chokepoints are bottlenecks in the supply chain (where are are high levels of dependence) that can introduce vulnerabilities in the AI industry.
• Innovation and Governance: Effective AI policy and governance require visibility into how knowledge, resources, and capabilities flow through the ecosystem. Understanding upstream and downstream dependencies is essential for designing interventions related to compute, disclosures, liability, and safety oversight.
• Market Concentration and Competition: Mapping the supply chain can reveal where market power is concentrated and how competition evolves across different layers of the ecosystem.
• Strategic Intelligence: A systematic view of the supply chain helps decision-makers track ecosystem dynamics, identify influential nodes, and anticipate shifts in strategic advantage. Tracing how capital, talent, and capabilities flow enables better forecasting and more effective intervention.

Project Overview

Our project seeks to fill this gap by systematically extracting information about the AI supply chain in real-time using public data sources, from late 2022 to the present. This website represents one artifact of our research: a platform for exploring our comprehensive and regularly updated dataset of the relationships between AI actors. We gather this dataset by collecting and processing SEC filings, press releases, news articles, and more approximately weekly. See the statistics page for real-time aggregate statistics of the data. Our methodology is summarized below and described in detail in our paper.

The second component of our research is a deep analysis of this dataset. This research is ongoing and will be released with our forthcoming research paper (see citation to use below). We are guided by research questions, such as: (1) How are the relationships in the AI ecosystem emerging, disappearing, and changing across time? (2) What industries gravitate toward AI? (3) Do different information sources reveal different information? (4) Where do we see unusual relationships? (5) What are key chokepoints?

If you are interested in collaborating, please reach out to the emails at the bottom of this page.

Data Sources

We collect data from two complementary sources, as described below. We plan to expand our data collection efforts to include more sources in the future. (Feel free to reach out if you have suggestions!)

Together SEC filings and news articles collected from our search provide complementary information and perspectives. SEC filings have some blindspots, largely because only publicly traded companies are subject to mandatory disclosures. As many AI companies are not yet publicly traded, the filings we collect reveal with other organizations interact with AI companies. On the other hand, our Google search results cover news articles, press releases, and more, allowing for more comprehensive coverage. At the same time, these documents are not audited and do not require certain disclosures, as SEC regulations do.

Methodology

Our data collection methodology systematically extracts relationship information from the two data sources described above: SEC filings and Google Search results. We begin with a curated list of 30+ frontier and foundation model providers, including OpenAI, Anthropic, Google, Microsoft, NVIDIA, and others. For each company, we generate targeted search queries that capture variations in company names, product names, and model names.

For detailed configuration parameters: See the Search Config Examples section below for complete configuration files specifying search keywords, data sources, and collection parameters for both SEC filings and Google Search.

{
  "sec_keywords": {
    "Bytedance AI": ["\"Bytedance AI\"", "\"Bytedance\" \"Doubao\"", "\"Bytedance\" \"Skylark\"", "\"Bytedance\" \"Seedance\""],
    "Huawei": ["\"Huawei AI\"", "\"Huawei\" \"Pangu\""],
    "IBM AI": ["\"IBM AI\"", "\"IBM\" \"watsonx\"", "\"IBM Granite\""],
    "Midjourney": ["\"Midjourney\""],
    "Snowflake": ["\"Snowflake AI\"", "\"Snowflake Arctic\"", "\"Snowflake\" \"Cortex\""],
    "Meta AI": ["\"Meta AI\"", "\"Meta\" Llama"],
    "xAI": ["\"x.AI\"", "\"xAI\" \"Grok\"", "\"x.AI\" \"Grok\""],
  "Anthropic": ["\"Anthropic\""],
    "Microsoft AI": ["\"Microsoft AI\"", "\"Microsoft\" Magma", "\"Microsoft\" \"Azure AI\"", "\"Microsoft\" \"Copilot\""],
    "Inflection": ["\"Inflection AI\"", "\"Inflection\" \"Pi\""],
    "AI21 Labs": ["\"AI21 Labs\"", "\"AI21\" \"Jamba\"", "\"AI21\" \"Jurassic\""],
    "Adept": ["\"Adept AI\"", "\"Adept\" \"Fuyu\"", "\"Adept\" \"ACT-1\"", "\"Adept\" \"ACT-2\""],
    "Krutrim": ["\"Krutrim\""],
    "Mistral AI": ["\"Mistral AI\"", "\"Mistral\" \"Mixtral\"", "\"Mistral\" \"Large\""],
    "Reka": ["\"Reka AI\""],
    "01.ai": ["\"01.ai\""],
  "Google AI": ["\"Google AI\"", "\"Google\" \"Gemini\""],
    "DeepMind": ["\"DeepMind\""],
    "LightOn": ["\"LightOn\""],
    "OpenAI": ["\"OpenAI\""],
    "Aleph Alpha": ["\"Aleph Alpha\""],
    "Numenta": ["\"Numenta\""],
    "4Paradigm": ["\"4Paradigm\""],
    "Contextual AI": ["\"Contextual AI\"", "\"Contextual Language Model\""],
    "Hugging Face": ["\"Hugging Face\""],
    "NVIDIA": ["\"NVIDIA AI\"", "\"NVIDIA\" \"Nemotron\""],
    "EleutherAI": ["\"EleutherAI\""],
    "Together AI": ["\"Together AI\"", "\"Together.AI\""],
    "Stability AI": ["\"Stability AI\"", "\"Stability\" \"Stable Diffusion\""],
    "AI2": ["\"AI2\"", "\"Allen Institute for AI\""],
    "Cohere": ["\"Cohere\""],
    "Databricks": ["\"Databricks\""],
    "Amazon": ["\"Amazon AI\"", "\"Amazon\" \"Titan\""],
    "Baidu": ["\"Baidu\" \"Ernie\""],
    "Alibaba": ["\"Alibaba AI\"", "\"Alibaba\" \"Qwen\""],
    "DeepSeek": ["\"DeepSeek\"", "\"DeepSeek\" \"DeepSeek-R1\""],
    "Cartesia": ["\"Cartesia\"", "\"Cartesia\" \"Cartesia-1\""]
  },
  "max_trials": 5,
  "num_chars": 1250
}

{
  "max_num_results_orgsite": 10,
  "max_num_results_newssite": 5,
  "max_num_results_gnews": 15,
  "time_duration": "1M",
  "org_sites": {
    "DeepMind": ["https://deepmind.google/discover/blog/"],
    "Meta AI": ["https://ai.meta.com/blog/"],
    "xAI": ["https://x.ai/news"],
    "Anthropic": ["https://www.anthropic.com/news"],
    "Microsoft AI": ["https://blogs.microsoft.com/ai/"],
    "Inflection": ["https://inflection.ai/blog"],
    "AI21 Labs": ["https://www.ai21.com/blog", "https://www.ai21.com/newsroom/"],
    "Adept": ["https://www.adept.ai/blog"],
    "Krutrim": ["https://blog.olakrutrim.com/"],
    "Mistral AI": ["https://mistral.ai/news/"],
    "Reka": ["https://reka.ai/news"],
    "01.ai": ["https://www.01.ai/#newsroom"],
    "Google AI": ["https://blog.google/technology/ai/"],
    "LightOn": ["https://www.lighton.ai/blog/"],
    "OpenAI": ["https://openai.com/news/"],
    "Aleph Alpha": ["https://www.aleph-alpha.com/newsroom"],
    "Numenta": ["https://numenta.com/blog/"],
    "Contextual AI": ["https://contextual.ai/news"],
    "Hugging Face": ["https://huggingface.co/blog"],
    "NVIDIA": ["https://blogs.nvidia.com/"],
    "EleutherAI": ["https://blog.eleuther.ai/"],
    "Together AI": ["https://www.together.ai/blog"],
    "Stability AI": ["https://stability.ai/news"],
    "AI2": ["https://allenai.org/blog"],
    "Cohere": ["https://cohere.com/blog"],
    "Databricks": ["https://www.databricks.com/blog", "https://www.databricks.com/company/newsroom"],
    "Amazon": ["https://www.aboutamazon.com/amazon-news-today", "https://www.aboutamazon.com/artificial-intelligence-ai-news"],
    "Baidu": ["http://research.baidu.com/Blog"],
    "Alibaba": ["https://www.alibabacloud.com/blog", "https://www.alibabacloud.com/en/news/"],
    "Cartesia": ["https://cartesia.ai/blog"],
    "IBM AI": ["https://newsroom.ibm.com/", "https://research.ibm.com/blog"],
    "Huawei": ["https://www.huawei.com/en/news"],
    "DeepSeek": ["https://deepseek.ai/blog"],
    "Snowflake": ["https://www.snowflake.com/en/news/"],
    "Midjourney": ["https://www.midjourney.com/updates"],
    "Bytedance": ["https://www.bytedance.com/en/news"]
  },
  "news_sites": ["techcrunch.com", "venturebeat.com"],
  "search_strings": ["AI partnership"],
  "search_locations": ["San Francisco, California, United States"]
}

Limitations

Our dataset and methodology have several limitations that should be considered when interpreting the results:

• SEC Filing Limitations: Note that one of our data sources is the SEC filings database. Because SEC filings are required only for companies publicly traded on U.S. markets, this data source captures relationships only when at least one party is a publicly traded U.S. company that submits the filing. We complement SEC filings with other sources, such as company websites, news articles, and more.
• Geographic Focus: Current data collection emphasizes US-based companies by nature of our methodology (see above).
• Relationship Types: Some relationship types (e.g., informal collaborations, research partnerships) may be under-represented if they are less likely to be present in SEC filings or announced online (e.g., in press releases and news articles).
• Temporal Coverage: Our data goes back to November 2022.
• Data Quality: We use LLMs to process and classify the documents in relationships in our data. This and other data filtering methods that we use may miss some important information or introduce mistakes.

How to Cite This Research

Cen, S. H., Gailmard, L., Bommasani, R., Ho, D. E., & Liang, P. (2025). Mapping the AI Supply Chain: An Analysis of the Complex Relationships in the AI Ecosystem. Forthcoming.

@misc{aisc_mapping_2025,
    title={{AI} {S}upply {C}hain {M}apping: An {A}nalysis of the {C}omplex {R}elationships in the {AI} {E}cosystem},
    author={Sarah H. Cen and Lindsey Gailmard and Rishi Bommasani and Daniel E. Ho and Percy Liang},
    year={2025},
    journal={Forthcoming}
}

Related Work

License

The data associated with this project and on this website (including relationship data, research findings, and website content) is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). This means the dataset may be used freely for academic research, education, and other non-commercial purposes, the dataset may not be used for commercial purposes without prior written permission, and proper attribution is required in all uses. Commercial use of the dataset, and commercial licensing arrangements for derivative products or services, are available by separate agreement. For commercial licensing inquiries, contact the authors above.

View Full License Information

Collaborations and Code/Data Access

This research is part of ongoing work at Stanford University. For code/data access, collaborations, or questions about methodology, please reach out to [email protected], [email protected], [email protected], [email protected], and [email protected].