The Inference Dependence Score: Open Index of Structural AI Dependence

The vast majority of nations will be using their frontier AI technology through foreign API services (inference), but they will not have any domestic frontier AI model or compute infrastructure. This study claims that just spreading out AI hardware does not spread out AI influence. In contrast, the control is increasingly concentrated at the inference layer, where fewer than five companies decide who gets to use high-performance AI, what terms and conditions they must meet to be able to do so, and how much it will cost.

This change poses an overlooked governance problem that has major consequences for catastrophic AI risk. In case governments, enterprises, and many of the most important public institutions depend on the same few suppliers for their ability to use increasingly powerful AI technology, then the inference layer becomes a potential source of systemic vulnerability and coercion in times of geopolitical crisis. While governance of AI has always focused on the compute infrastructure, on frontier AI models, and on laboratories, the issue of dependence on foreign-owned inference infrastructure was largely ignored.

To bridge this gap, I will build a replicable and open framework called the Inference Dependence Index (IDI) to evaluate a nation’s dependency on foreign-owned AI inference providers. I will also build a framework for National AI Gateways, which takes inspiration from interoperable public digital infrastructure like NAPAS in Vietnam and UPI in India. In contrast to promoting technological autarky, this solution considers ways in which compatible National AI Gateways operating within global AI ecosystems can cut down unwanted dependencies.

Project Timeline

This project will be completed within a six-month timeline.

Months 1-2: The working paper from the AI Safety Hackathon will be expanded into a 20-page research paper. This paper will use the relevant literature on weaponized interdependence (Farrell & Newman), compute governance (Heim et al.), structured access (Shevlane), and cloud governance/jurisdiction (Lehdonvirta et al.) to come up with a cohesive theoretical framework. The IDI method will be further refined through structured feedback from AI governance researchers.

Months 2-4: The IDI will be scaled from the initial pilot countries (Vietnam, Indonesia, and Singapore) into a comparative analysis for eight to ten Asia-Pacific countries. The research methods, weightings, and evidence base used will be openly shared to allow other researchers to validate, critique, or improve the research findings.

Months 4-5: The next iteration of IDI calculator v2 will be publicly released as an open source web app built off HTML and Streamlit prototypes created earlier.

Months 5-6: A policy brief will be published outlining the National AI Gateway model along with the possible approaches to reduce dependency on foreign inference layers.

The findings from the project will be distributed via AI governance research communities, regional policy networks, and public writing.

About me

I am an independent researcher in AI governance. In my role as political assistant at the US Consulate General, I evaluated issues pertaining to technology governance, human rights, and emerging policies. Thereafter, I worked as manager of AI trust & safety/model alignment projects at Pareto.ai, managing data quality/governance for frontier AI systems. I also write on topics involving technology governance/digital rights and am currently studying for an MSc in Applied AI. I have just completed BlueDot Impact's AGI Safety Fundamentals Program. This project is a product of my exposure to the non-frontier AI ecosystem. The draft paper, methodology, and software designs have been created; this grant will allow me to verify, grow, and release them as public infrastructure for AI governance.

Project Outcomes

Upon completion of this project, I expect to achieve the following outcomes:

A publicly available working paper covering a survey of the conceptual frameworks for inference dependency/AI power concentration/governance of the inference layer.

Methodology of Inference Dependency Index (IDI) and an open comparative dataset of 8-10 Asian/Pacific nations.

Open-source version of the IDI calculator v2 that enables independent researchers/policymakers to calculate the inference dependency of nations.

Policy brief outlining National AI Gateways translating the research into actionable governance options for states importing AI.

These projections will be used as the infrastructure for measuring the power of AI governance. With quantifying inference dependence, this project provides a basis for the researchers and governments of different nations to understand the phenomenon of AI power concentration and to examine governance measures before the dependency is set.