Compute Concentration Is a First-Order Socio-Economic Risk; ML Must Measure It

Track: Track 2: Socio-Economical and Future Visions

Keywords: Compute Governance, Compute Concentration, Market Structure, Measurement, Resource Allocation, Sustainability

TL;DR: We frame compute concentration as a vital post-AGI risk, proposing 'Compute Cards' and indices (Gini, Top-k) to measure it. This transforms abstract power into a monitored variable, enabling accountable governance and independent oversight.

Abstract:

Post-AGI governance proposals often emphasize rapid adaptation, stabilization, and preventing gradual disempowerment of human institutions. Yet many policy levers—from liability regimes to treaties to benefit-sharing—implicitly assume that researchers, regulators, and civil society can observe and contest the allocation of the key bottleneck input: compute. We argue that compute concentration is a first-order socio-economic and safety risk in post-AGI scenarios because it shapes (i) who can build, (ii) who can audit, and (iii) who can deploy and gate high-impact systems. We propose a practical measurement layer: compute provenance reporting via standardized Compute Cards (public, banded disclosures plus a confidential annex for trusted auditors), paired with simple, robust concentration
indices (Top-k, HHI, Gini) computed across training and inference. Using a public dataset of estimated training costs for notable frontier models, we illustrate extreme inequality (Gini ≈ 0.89; Top-1 share ≈ 54%). We then outline an MVP “dashboard” pipeline, a disclosure threat model, and a research agenda that links measurement to policy evaluation (e.g., public compute, procurement, and competition interventions) while respecting safety and security constraints.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Zijin_Wu3

Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 29