Market-Based Compute Permits for Frontier AI Safety

Compute is one of the most tractable levers for governing transformative AI, yet we still lack concrete, well specified designs for market based compute governance at national and international levels. Building on recent work on compute caps, licensing, and hardware enabled oversight, this project explores how tradable compute permits could be used to constrain frontier training runs, align incentives for labs and states, and support auditability and monitoring.

Core questions include: How should compute permits be allocated, auctioned, or grandfathered, and at what level of granularity (for example per training run, per lab, or per data center)? Under what conditions would international trading of permits increase or reduce risk? How can enforcement, monitoring, and penalties be structured to discourage evasion without creating strong incentives for regulatory arbitrage or secrecy? How should such schemes interact with export controls on chips, data center regulation, and existing or proposed AI safety standards?

Mentees will work with me to 1. Survey and synthesize existing proposals on compute governance, AI hardware policy, and emissions style permit systems. 2. Develop simple economic or game theoretic models that capture key tradeoffs in compute permit design, including distributional and strategic effects across states and labs. 3. Implement and analyze a small number of illustrative scenarios (for example a stylized model with two or three countries with different risk preferences and industrial bases) using Python or similar tools. 4. Coauthor an accessible working paper or policy brief, plus slides for SPAR Demo Day, aimed at decision makers in governments and multilateral institutions.

The project is scoped so that mentees can contribute meaningfully within three months at about 5 to 10 hours per week, with stretch goals such as a small numerical simulation or a deeper case study on a particular jurisdiction if time and skills allow. I am happy to tailor the technical depth to the strengths of the selected mentees, for example emphasizing modeling and code for more technical teams, or qualitative institutional design and comparative policy analysis for more policy oriented teams.

Governing access to high-end compute is one of the most promising ways to constrain the development and deployment of highly capable AI systems. Well-designed market-based compute permits could make it significantly harder to train dangerous models in secret, create clearer incentives for labs to comply with safety standards, and provide regulators with quantitative levers to manage aggregate risk.

This project advances AI safety by clarifying which permit designs actually reduce risk and which might create new vulnerabilities. By grounding proposals in simple economic and game theoretic models, and connecting them to real world institutions such as export controls, standards, and licensing regimes, we aim to produce concrete options that policymakers can adopt or adapt. The work builds on my prior research on compute permits, semiconductor supply chains, and global AI governance, and is intended to feed into ongoing conversations in governments, think tanks, and multilateral forums about how to implement effective compute based oversight.

Mentees will act as junior coauthors on the project. They will help scope the research questions, conduct literature reviews on compute governance and permit systems, build and analyze simple models or case studies, and contribute directly to writing and visualization. I will set the overall direction and ensure that the project stays tractable and relevant for AI safety, but mentees will have real ownership over subquestions, modeling choices, and sections of the final write up. By the end of the program, the goal is for each mentee to have a clear, citable contribution to a working paper or policy brief and to have deepened their understanding of compute based AI governance.

Applicants should meet at least one of the following profiles: 1. Economics or quantitative social science background, ideally having completed intermediate microeconomics or game theory and at least one course in econometrics or statistics. 2. Computer science, mathematics, or related technical background, with strong interest in AI governance and some familiarity with public policy or international relations.

In all cases, mentees should have: • High proficiency in either Python or R, including ability to implement simple models, run simulations, and produce clear figures. • Ability to read and synthesize research papers and policy reports, and to write clearly for a mixed technical and policy audience. • Genuine interest in AI safety and global governance, and willingness to think carefully about incentive design and institutional feasibility.

Prior experience with any of the following is a plus but not strictly required: mechanism design, industrial organization, macroeconomics of technology, compute governance, or work on AI policy and regulation.

Minimum 5 hours per week, with 7 to 10 hours per week preferred for mentees who want to take on more ambitious modeling or writing tasks.

No strict geographical preference. I expect mentees to be able to attend a weekly group meeting scheduled in a time window that overlaps with Central European and Jakarta working hours, for example between 08:00 and 18:00 in UTC+1 to UTC+7. Occasional asynchronous collaboration is fine, but applicants should be able to join at least one regular live call each week.