If nations wanted to verify compliance with international AI agreements, they'd need to know where all the AI chips are. How well can we currently answer that question, and what would it take to do better?
About the project
International agreements on AI development would require verifying that all major AI developers are following the rules. This verification problem has two parts: tracking new chips from the fab, and accounting for chips already deployed. The first is relatively straightforward—you can monitor production and shipments from the source. The second is harder: millions of AI accelerators are already distributed across data centers worldwide, and we don't have a clear picture of how many exist, where they are, or who controls them.
This project aims to assess the current state of knowledge about the global AI hardware supply. We'll identify the key quantities that matter (chips produced per year by type, chips deployed by region and owner, chips in transit, chips decommissioned), catalog the available data sources (fab production data, company disclosures, revenue figures, satellite imagery of data centers, import/export records), and evaluate how confident we can be in each. We'll then identify gaps and propose methods that could improve coverage—potentially including recommendations for what the intelligence community or international bodies should be tracking.
Theory of change
You can't verify an international AI agreement if you don't know where the chips are. This project describes the current state of the art in locating the chips and identifies ways that we can improve on that state of the art.
Your role
Mentees will be research assistants. I'll set the research direction and outline the structure of the final report. Mentees will dig into data sources, read industry reports, and talk to experts. We'll meet weekly to discuss findings and figure out what to look at next. The final output will be a published report.
Prerequisites
- You regularly read research from Epoch, SemiAnalysis, and similar sources
- You have published research or technical writing in related areas
- You can explain the following terms without looking them up: dies, yield, FLOP vs FLOP/s, memory bandwidth vs interconnect bandwidth, H20 vs H100 vs B100 vs B200.
Application question(s)
- Please link 1 piece of relevant writing
- Please describe in 1-3 sentences your relevant experience
- Please describe your career goals over the next few years in 1-3 sentences
About the mentor
Hi! I'm a DPhil student at Oxford focusing on compute governance. I'm interested in on-chip mechanisms and other means to enforce domestic and international AI policies. I spend part of my time investigating grants on AI at Longview Philanthropy.