Mapping Domestic AI-enabled Power Concentration Chokepoints

Current research and literature on risk from AI-enabled extreme power concentration is at an early stage. What exists right now is foundational work which scopes the risk, attempts to decompose the constitutive elements of potential EPC threat models, and offers suggestions often focused on increasing transparency and oversight at frontier labs to flag dangerous capabilities that power-seeking actors could exploit. This is essential and formative work.

I seek to contribute to this nascent canon with a report grounded in a domestic (US) analysis of the current institutional configuration (financial infrastructure, current regulatory frameworks, electoral and administrative machinery of government, law enforcement and surveillance apparatuses, information and communication layer, control over critical infrastructure, relevant legal-judicial backstops -- and access to frontier AI capability and compute as it intersects across these domains), that offers conditional threat models based on both a current capability (accounting for incremental advances within the current research paradigm), and a capability leap (a substantial jump in capability from the next generation of R&D) assumption. In this analysis and modelling I want to consider whether threat model trajectories are invariant between these two conditions, or sensitive to them, and in what particular ways.

The result of this analysis and modelling is a report that proposes two threat models based on current institutional arrangements and precedents, within a ~5 year window, and identifies the various chokepoints in these models that are accessible to intervention which go beyond the current increasing-transparency and better-auditing mitigations. I do not expect the identified chokepoints to be equally accessible to intervention or equally impactful if closed, and so they will be graded along two axes: tractability and marginal impact.

The yield of this report is to provide:

A) a piece of foundational literature in the EPC canon to increase subject-matter knowledge for those interested or working in AI safety and invite critique in order to sharpen the discussion and strategy on AI-enabled EPC
B) a clear roadmap of intervention strategies based on both a current-capability, and a capability-leap assumption, which researchers, policy-makers, and grantmakers can reference when making strategic funding and intervention decisions

A third personal yield is to provide me with the deep knowledge of the subject matter, and a research deliverable in order to pursue research fellowships and eventually full time work in AI safety as a researcher or grantmaker with a focus on strategy and governance. It is in this respect that this is also a grant for career transition funding as I intend to devote myself full time to the drafting of this report over the following 3-5 months (depending on the funding secured).

The grant covers a living stipend plus one-time conference travel.

Minimum (~$21k) — 3 Months. Stipend of ~$6k/month × 3 months (~$18k), plus ~$3k one-time travel to the Cambridge Conference on Catastrophic Risk. This funds the core deliverable: the scoping report with its two-condition threat models and graded chokepoints. The stipend is a tax-inclusive gross figure tracking documented Toronto living costs (~CAD $6,400/month); it is not full income replacement.

Ideal (~$36k) — 5 Months. The same ~$6k/month stipend extended to 5 months (~$30k), the same ~$3k one-time travel, plus a ~$3k contingency buffer for travel-cost and tax-gross-up variance. The two additional months are spent on depth, developing the intervention analysis and stress-testing the chokepoints across both capability conditions to establish whether they hold trajectory-invariant.

Full line-item breakdown: [spreadsheet link].