P100 Compute-threshold definitions and loopholes

Which compute counts toward a governance threshold, how are thresholds set and renegotiated, and how can runs be structured to evade them?

What compute counts toward thresholds, how thresholds are set, moved and renegotiated (including regional caps and coverage exemptions), and how runs evade them.

Status

Analyzed Evasion clause only: a simulated cost of structuring a 10^24, 10^25 or 10^26 FLOP run as clusters below every proposed registration threshold (DiLoCo-family methods over 100 Mbps links; extrapolated 10-300x beyond tested model size). The same paper proposes a 1,280 GB HBM memory threshold as a fix to the threshold definition, also paper only. Which compute counts (fine-tuning, distillation, mixture-of-experts, inference scaling) and how thresholds are renegotiated: no work beyond framing and stated difficulties.

these proposed or existing regulations could be evaded using $1.6M, $31M, and $3.8B, respectively, of hardware arranged in clusters smaller than any proposed registration requirement
Rahman (2026), Does Distributed Training Undermine Compute Governance? (MIRI), arXiv 2605.29359 (2026-05-28), L466–468

Posed by

Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1).

Sub-questions

IDSub-questionWhat the sources report
P100.1At thresholds of 10^24, 10^23 and 10^22 FLOP (the lower two set by this survey), how many facilities worldwide would require monitoring?open
P100.2Can metadata or authenticated end-user connections detect several concurrent runs that compose one modular/MoE system, and what is a 'distinct model'?open
P100.3Can fine-tuning and merging below threshold (often <1% of pre-training compute) be tracked, and how much capability does it add in compute-equivalent terms?a source suggests an answer (not checked)
P100.4What fraction of frontier compute goes to non-pre-training workloads (RL, synthetic data incl. discarded samples, inference)?open

Sub-questions not yet stated precisely enough to check

IDSub-question
P100.5How should compute thresholds for exclusion/registration track algorithmic progress and distributed training, and how are they renegotiated?
P100.6How should regions and region-wide worst-case compute caps be defined?
P100.7Must chips be exempted from logging until large-scale training can be distinguished, and what share of guarantees is lost?
P100.8Can 'logical location' (a state taking responsibility for chips in its borders) substitute for physical location tracking?
P100.9What technical requirement tiers / feature sets follow from attacker threat models in each export context?
P100.10Which compute counts toward a threshold (fine-tuning, distillation, MoE experts, inference scaling), and how are threshold quantities set?

Possibly relevant work found by automated search

Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.

For P100.3: Can fine-tuning and merging below threshold (often <1% of pre-training compute) be tracked, and how much capability does it add in compute-equivalent terms?

Data