Compute Thresholds in Regulation
An assessment of training-compute triggers as a regulatory instrument.
Abstract
Several jurisdictions have adopted training-compute thresholds as the trigger for regulatory obligations on advanced AI systems. This brief assesses the instrument on its own terms: is training compute measurable with sufficient precision, is it a good proxy for the risks the regulation targets, and how does the instrument behave as the technology changes. We find that compute is unusually measurable for a regulatory trigger and unusually administrable, that its correlation with capability has weakened as post-training has grown as a share of expenditure, and that thresholds set as fixed absolute values have a predictable obsolescence problem. We set out the design variants that address each weakness and the trade-offs each variant carries.
Key findings
- 01Training compute is measurable to within a factor of roughly two from external evidence, which is high precision relative to most regulatory triggers.
- 02The correlation between training compute and capability on held-out instruments has weakened as post-training expenditure has grown.
- 03Fixed absolute thresholds capture a growing number of systems each year with no change in policy, an effect that is rarely intentional.
- 04Relative thresholds and capability-based triggers each solve part of the problem and introduce different administrability costs.
A regulatory trigger has to be measurable, administrable and connected to the harm it targets. Compute thresholds do well on the first two. The third is where the argument lives.
This brief does not argue for or against compute thresholds. It assesses them as an instrument. Our view is that the debate has been conducted mostly at the level of principle, and that the practical questions — can it be measured, can it be enforced, does it track the thing you care about, what happens to it over time — have clearer answers than the principled questions and deserve more attention.
Measurability
Training compute is unusually measurable for a regulatory trigger. It can be estimated externally from hardware deployment, elapsed time and architectural disclosure to within roughly a factor of two, and verified internally from records that firms keep for their own operational reasons. Compare this with triggers based on revenue, market share, risk level or user harm, all of which require far more contested measurement.
The measurement caveats are real but bounded. Definitional choices — whether post-training counts, whether failed runs count, whether utilisation is measured or assumed — move the number by factors of two or more, which means a threshold regime must specify its definition precisely. Most existing instruments do so incompletely.
Proxy quality, and its decay
The substantive weakness is the link between compute and capability. In 2022, training compute was an excellent predictor of capability on essentially any instrument. It is a weaker predictor in 2026. The reason is compositional: an increasing share of capability derives from post-training investment, from data curation and from inference-time computation, none of which a pre-training compute threshold captures.
Frontier Labs evaluation programme. Systems within one order of magnitude of frontier compute at each date.
A correlation of 0.61 is not nothing. It remains among the stronger single-variable predictors available. But a trigger whose proxy quality is declining year on year has a built-in expiry, and instruments written today should account for that.
Obsolescence
A fixed absolute threshold captures more systems each year as the cost of a given compute level falls. This is sometimes the intent. Frequently it is not, and the drafting record suggests the effect was not always considered. A threshold set to capture two or three systems in the year of drafting may capture dozens within four years.
| Variant | Addresses | Introduces |
|---|---|---|
| Fixed absolute threshold | Simplicity, legal certainty | Predictable obsolescence |
| Indexed threshold | Obsolescence | Index design contested; less certainty |
| Relative threshold (top-N or top-percentile) | Obsolescence, scope stability | Requires continuous market observation |
| Capability-based trigger | Proxy quality | Measurement contestability, gaming surface |
| Compound trigger (compute and capability) | Both weaknesses partially | Complexity, higher compliance cost |
Sources and references
- [1]Statutory and regulatory instruments, nine jurisdictions.
- [2]Frontier Labs compute reconstruction methodology.
- [3]Held-out evaluation results, 2023–2026.
Source assessments, provenance records and the full data appendix are described in the methodology. Corrections to this publication are recorded on the corrections page.
Keywords
- policy
- regulation
- compute thresholds
- governance
Full edition
Read the complete report.
The full edition is released to named readers. Tell us who you are and it opens straight away.
- Extent
- 34 pages
- Price
- Free
Cite this publication
Policy Desk (2026). Compute Thresholds in Regulation. Frontier Labs. https://doi.org/10.5281/frontierlabs.2026.0012
Research infrastructure powered by AGI-S.