Stanford’s Ecosystem Graphs Maps AI Model Dependencies and Risks

Stanford’s Center for Research on Foundation Models (CRFM), housed within the Stanford Institute for Human-Centered Artificial Intelligence (HAI), has released Ecosystem Graphs, a centralized knowledge graph framework designed to systematically map the dependencies, ownership, and technical lineage of foundation models (FMs) across the AI landscape. The initiative arrives at a pivotal moment: as major AI labs increasingly withhold critical details about model architecture, training data, and compute, the industry faces a growing transparency crisis. OpenAI’s GPT-4 technical report, which omitted key specifications such as dataset construction and training compute, underscores this trend toward opacity.

Ecosystem Graphs directly addresses this challenge by providing a structured, community-maintained resource—akin to a Wikipedia for foundation models—that documents the interconnected web of datasets, models, and applications underpinning modern AI. For executives, investors, developers, and policymakers, the tool offers an unprecedented view into the concentration of power, risk, and opportunity within the AI supply chain.

A Comprehensive Map of the AI Ecosystem

The Ecosystem Graphs framework currently documents over 200 assets, spanning datasets, models, and applications, across 60 organizations and 9 modalities, including text, images, video, code, music, and genomic sequences. Among the documented assets are foundational datasets such as The Pile and LAION-5B; models like BLOOM (Meta), Stable Diffusion (Stability AI), PaLM (Google/DeepMind), Mochi 1 (released under Apache 2.0), and Gemini 2.0 (Google); and widely deployed applications including GitHub Copilot X, Notion AI, Microsoft 365 Copilot, and the ChatGPT API.

What sets Ecosystem Graphs apart from prior attempts to catalog AI systems is its emphasis on dependencies. The framework meticulously traces how assets interrelate—for example, how LAION-5B enabled the development of Stable Diffusion, which in turn spawned derivatives like Stable Diffusion Reimagine. Beyond dependency mapping, the graph annotates critical metadata, including licensing terms, carbon emissions from training, and organizational ownership. This granularity provides stakeholders with a detailed view of the ecosystem’s technical, legal, and environmental underpinnings.

The project is led by Rishi Bommasani, Society Lead at CRFM, and Percy Liang, CRFM Director, with contributions from a global community of researchers and practitioners. Its academic rigor is further validated by the publication of an associated paper in the AAAI/ACM Conference on AI, Ethics, and Society (AIES).

Notably, Ecosystem Graphs has already gained traction among influential stakeholders. The framework has been adopted by the 2024 AI Index Report to analyze trends in model development and deployment, and by the UK’s Competition and Markets Authority (CMA) to assess market concentration and potential antitrust concerns in the AI sector.

The Urgency of Transparency in an Opaque Era

The release of Ecosystem Graphs coincides with a broader industry shift toward secrecy. Foundation models emerged as a dominant AI paradigm around 2020, accelerated by the launch of OpenAI’s GPT-3. By late 2022, ChatGPT had become the fastest-growing consumer application in history, signaling the transformative potential of these technologies. Yet, as of 2024–2026, major developers have progressively restricted disclosures about training data, compute, and architecture, citing competitive advantage and safety concerns. OpenAI’s GPT-4 report, which withheld details on its architecture and dataset construction, marked a turning point in this trend.

This lack of transparency poses significant challenges across sectors. For executives and investors, it obscures supply-chain vulnerabilities, market consolidation risks, and potential acquisition targets. For developers and founders, it complicates the discovery of permissively licensed models (e.g., Mochi 1) or emerging modalities like video generation, where models such as Google’s Veo 2 and RunwayML’s offerings are rapidly evolving. For policymakers and auditors, it hinders efforts to conduct safety audits, antitrust analyses, or regulatory oversight, as the technical and organizational relationships between models remain unclear.

Ecosystem Graphs counters this opacity by providing a structured, publicly accessible resource that documents the AI ecosystem’s complexities. Its community-driven approach ensures that the graph remains dynamic and responsive to the rapidly evolving landscape of foundation models.

Hubs of Power and the Risks of Concentration

One of the framework’s most critical contributions is its identification of “hub” assets—high-connectivity nodes that concentrate disproportionate influence over the AI ecosystem. These hubs include:

• The Pile: A massive open dataset used to train models by Meta, Microsoft, Stanford, Tsinghua University, and Yandex, among others. Its widespread adoption makes it a linchpin in the development of numerous foundation models.
• P3/xP3: Instruction-tuning datasets that have become foundational for fine-tuning large language models, enabling downstream applications across sectors.
• PaLM: Google’s internal family of foundation models, which underpins a vast array of downstream applications and services, from search to enterprise tools.
• ChatGPT API: A deployment hub that enables thousands of third-party applications, creating widespread dependency on OpenAI’s infrastructure across industries.

These hubs highlight the ecosystem’s vulnerability to cascading failures. For instance, LAION-5B—currently the subject of ongoing litigation over copyright infringement in its training data—serves as the backbone for Stable Diffusion and its derivatives. A legal ruling against LAION-5B could disrupt not only Stable Diffusion but also the myriad applications built atop it, from creative tools to enterprise solutions. Similarly, reliance on proprietary APIs like OpenAI’s or Google’s introduces single points of failure for businesses and developers, where an outage or policy change could have far-reaching consequences.

Beyond technical dependencies, the framework also sheds light on the concentration of institutional power. The dominance of a few key players—such as OpenAI, Google, Meta, and Microsoft—in both model development and downstream deployment raises questions about market consolidation and competitive fairness. Regulators, including the UK CMA, are increasingly scrutinizing these dynamics, using tools like Ecosystem Graphs to assess potential antitrust risks.

However, the framework is not without limitations. As a community-driven project, it faces challenges in long-term maintenance, data verification, and coverage bias. The current graph overrepresents Western, English-language assets, and its reliance on voluntary contributions may lead to gaps in documenting non-Western or proprietary systems. Additionally, while the graph maps technical dependencies, it does not fully capture unresolved legal risks, such as copyright disputes over training data or licensing ambiguities. For example, ongoing litigation involving LAION-5B and Stable Diffusion highlights the legal uncertainties that technical dependency mapping alone cannot resolve.

Strategic and Regulatory Implications

Ecosystem Graphs has already demonstrated its value to policymakers and industry leaders alike. The UK CMA has leveraged the framework to analyze market concentration and competitive dynamics in the AI sector, particularly in its scrutiny of partnerships between major tech firms and AI startups. Similarly, the 2024 AI Index Report used the graph to assess trends in model development, deployment, and transparency, providing a data-driven foundation for discussions about the future of AI governance.

For industry leaders, the tool offers actionable strategic insights. By identifying hubs like The Pile or PaLM, executives can better understand market consolidation, potential acquisition targets, and supply-chain dependencies. Investors, meanwhile, can use the graph to evaluate the resilience of AI startups against cascading failures or legal risks tied to upstream assets. For example, a startup heavily reliant on a proprietary API or a legally contested dataset may face higher operational and regulatory risks.

Developers and founders stand to benefit from the framework’s ability to surface permissively licensed models and emerging modalities. The graph’s documentation of Mochi 1—an Apache 2.0-licensed model—highlights alternatives to proprietary systems, while its tracking of video generation models like Google’s Veo 2 or RunwayML’s offerings provides a roadmap for innovation in multimodal AI. By understanding the technical and legal lineage of these assets, developers can make more informed decisions about which models to adopt, modify, or build upon.

For policymakers and auditors, Ecosystem Graphs provides a structured basis for regulatory oversight, safety audits, and antitrust analysis. By mapping institutional power and downstream impact, the framework enables more effective assessment of market dominance, competitive practices, and potential risks to consumers and businesses. This is particularly critical as foundation models continue to reshape industries, from healthcare to finance, and as governments worldwide grapple with the need for balanced, effective AI regulation.

As the AI ecosystem evolves, the need for transparency and accountability will only grow. Ecosystem Graphs represents a critical step toward addressing this need, offering a rare window into the complex, interdependent world of AI development. However, its long-term success will depend on sustained community engagement, rigorous data validation, and expansion beyond its current Western-centric scope. For now, it provides professionals across sectors with an indispensable tool for navigating the opportunities and risks of the foundation model era.