Compute Threshold (AI Governance)
compute-threshold · Compute governance
A regulatory trigger expressed as floating-point operations (FLOPs) consumed during model training, above which specific reporting, evaluation, or governance obligations attach.
Definition & scope
Compute thresholds operationalize the intuition that capability scales (imperfectly) with training compute. Jurisdictions have adopted different thresholds: US EO 14110 used 10²⁶ FLOPs for foundation-model reporting; EU AI Act Art. 51 uses 10²⁵ FLOPs as the systemic-risk presumption; China's GenAI Measures use no compute threshold (registration triggered by public-facing deployment instead); UK AISI commitments are voluntary and capability-based rather than compute-thresholded. Critics note that thresholds become outdated as algorithmic efficiency improves and that compute alone is an imperfect capability proxy.
Locus of dispute: Is compute-thresholding a defensible proxy for governance-relevant capability? Algorithmic-efficiency improvements (DeepSeek R1 reportedly demonstrating frontier-tier reasoning with substantially less training compute than 10²⁵-FLOP-class models, though the exact training-FLOP count is not publicly disclosed by the provider) destabilize the threshold; field is split on whether compute thresholds should be indexed to efficiency, replaced by behavioural evaluation, or kept fixed for predictability.
Mechanism: how training compute is measured and why it is governable
A compute threshold operationalises capability indirectly: rather than testing behaviour, it counts the floating-point operations (FLOP) spent training a model. For dense transformers, training compute is conventionally C ≈ 6ND, with N parameters and D tokens 12. Because N and D are known ex ante, the EU AI Act presumes systemic risk above 10^25 FLOP (Regulation (EU) 2024/1689, Art. 51(2)). Compute is also uniquely governable: Sastry et al. 2024 call it "detectable, excludable, and quantifiable, and ... produced via an extremely concentrated supply chain" 3, and Heim et al. 2024 argue providers "should have legal obligations" to keep records and report frontier training 4. The presumption is rebuttable — Annex XIII weighs parameters, data, and benchmarks — so it screens, not verdicts 5.
Open critiques and debates
Critics identify specific failure modes. Hooker 2024 argues "the relationship between compute and risk is highly uncertain and rapidly changing" and that fixed thresholds "overestimate our ability to predict what abilities emerge at different scales," so a FLOP line over-includes benign large models and misses capable small ones 6. Because C ≈ 6ND captures only training FLOP, fine-tuning and test-time gains escape it — a loophole where "enhancement techniques ... capable of decreasing training compute usage while preserving ... model capabilities" stay below threshold 7. Ho et al. 2024 measure compute-to-fixed-performance halving about every eight months 8. Casper, Krueger & Hadfield-Menell 2025 warn demanding evidence first can delay regulation 9; Reuel et al. 2024 list measurement and verification as open problems 10.
Distributional and geopolitical stakes of compute-based governance
A compute threshold is not jurisdiction-neutral: enforceability depends on where training hardware physically sits. Lehdonvirta, Wü & Hawkins 2024 census hyperscale cloud regions and find a divide between a "Compute North" hosting frontier-capable infrastructure and a "Compute South" without it, shaping who can wield compute-based governance 11. The asymmetry is deliberate: Kollar & Stokols 2026 show US and Chinese sovereign-compute drives reorganise land, energy, and regulation 12, while Weymouth 2025 finds states asserting "strategic digital sovereignty" through selective alliances, fragmenting infrastructure into techno-blocs 13. Because thresholds run through the concentrated supply chain Sastry et al. 2024 identify 3, cross-border verification 14 becomes a precondition for any cross-bloc regime.
Use in governance
How instruments operationalise this concept
| Instrument | Jurisdiction | Status |
|---|---|---|
| EU AI Act | EU | in force |
| Executive Order 14110 on Safe, Secure, Trustworthy AI | US | partial |
Appears in topic articles
Editorial note
When citing a specific FLOP threshold, always pair it with the jurisdiction and instrument. '10²⁵ FLOPs' is meaningful only under EU AIA; the same number has different implications in other regimes.
See also
Further reading
Sources on the broader topics this concept relates to — complementing, not standing in for, the primary sources cited inline above. 65 academic & grey-literature sources; catalogued metadata with a primary link; one-line findings are ✦ AI-generated summaries, labeled as such (charter §7.9). Browse the full literature index.
- Geopolitical ecologies of cloud capitalism: Territorial restructuring and the making of national computing power in the U.S. and China Peer-reviewed✦ AIUS and Chinese drives for sovereign AI/cloud dominance depend on reorganizing land, energy and regulatory systems to sustain large-scale national computing power.
- An interdisciplinary account of the terminological choices by EU policymakers ahead of the final agreement on the AI Act: AI system, general purpose AI system, foundation model, and generative AI Peer-reviewed✦ AITraces how the AI Act's legal text shifted across versions among the terms 'AI system, general purpose AI system, foundation model, and generative AI', exposing definitional instability in the regime.
- The EU model of AI governance: regulating artificial intelligence through law and policy Peer-reviewed✦ AIAnalyses how the AI Act's risk-based model handles general-purpose and foundation models whose 'autonomous content generation challenges legal categories of authorship, accountability, and control'.
- Generative AI and data protection Peer-reviewed✦ AIExamines friction between foundation-model training and the GDPR, noting models that 'memorize and leak pieces of training data' cannot be treated as anonymous.
- Defending Compute Thresholds Against Legal Loopholes Preprint✦ AIIdentifies 'enhancement techniques that are capable of decreasing training compute usage while preserving... model capabilities', exposing loopholes in compute-reporting thresholds.
- Digital Disintegration: Techno-Blocs and Strategic Sovereignty in the AI Era Peer-reviewed✦ AIArgues states increasingly assert 'strategic digital sovereignty...through selective alliances with firms and other governments,' fragmenting global AI infrastructure into techno-blocs rather than multilateral order.
- GPTs are GPTs: Labor market impact potential of LLMs Peer-reviewed✦ AIFinds around 80% of the U.S. workforce "could have at least 10% of their work tasks affected" by LLMs, which exhibit "traits of general-purpose technologies".
- Computing Power and the Governance of Artificial Intelligence Preprint✦ AIArgues compute is a uniquely governable lever because it is "detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain".
- Training Compute Thresholds: Features and Functions in AI Regulation Preprint✦ AIFinds "training compute currently is the most suitable metric to identify GPAI models", but thresholds should only trigger further scrutiny, not determine risk measures alone.
- Compute North vs. Compute South: The Uneven Possibilities of Compute-based AI Governance Around the Globe Peer-reviewed✦ AICensus of hyperscale cloud regions shows a divide between "Compute North" states hosting training-relevant compute and a Compute South, shaping who can wield compute-based governance.
- Generative AI in EU law: Liability, privacy, intellectual property, and cybersecurity Peer-reviewed✦ AIExamines how the EU AI Act, liability regimes, GDPR, copyright and cybersecurity rules apply to generative AI, identifying gaps and proposing targeted regulatory refinements.
- Infrastructuring AI: The stabilization of 'artificial intelligence' in and beyond national AI strategies Peer-reviewed✦ AIShows the UK National AI Strategy 'stabilises: AI as an autonomous and inevitable force', revealing how national strategies fix actors, capital flows, and power relations.
+ 53 more across this concept's topics — see the literature index.
References
Sources cited inline in the analysis (linked from the superscript markers), then the primary instrument sources behind the classifications.
- Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, Dario Amodei (2020) Scaling Laws for Neural Language Models, arXiv (cs.LG). arXiv:2001.08361 — Establishes that model 'loss scales as a power-law with model size, dataset size, and the amount of compute', the empirical basis for compute-threshold regulation of foundation models. ↩
- Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, et al. (DeepMind) (2022) Training Compute-Optimal Large Language Models, arXiv (cs.CL); NeurIPS 2022. arXiv:2203.15556 — The 'Chinchilla' study shows 'model size and the number of training tokens should be scaled equally', complicating compute-only regulatory thresholds. ↩
- Sastry, Heim, Belfield, Anderljung, Brundage, et al. (2024) Computing Power and the Governance of Artificial Intelligence, arXiv. arXiv:2402.08797 — Argues compute is a uniquely governable lever because it is "detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain". ↩
- Lennart Heim, Tim Fist, Janet Egan, Sihao Huang, Stephen Zekany, Robert Trager, Michael A. Osborne, Noa Zilberman (2024) Governing Through the Cloud: The Intermediary Role of Compute Providers in AI Regulation, arXiv (cs.CY). arXiv:2403.08501 — Argues 'compute providers should have legal obligations' to secure infrastructure, keep records, verify activity and report frontier training as regulatory intermediaries. ↩
- Heim & Koessler (2024) Training Compute Thresholds: Features and Functions in AI Regulation, arXiv. arXiv:2405.10799 — Finds "training compute currently is the most suitable metric to identify GPAI models", but thresholds should only trigger further scrutiny, not determine risk measures alone. ↩
- arXiv:2407.05694 ↩
- Matteo Pistillo, Pablo Villalobos (2025) Defending Compute Thresholds Against Legal Loopholes, arXiv (cs.CY). arXiv:2502.00003 — Identifies 'enhancement techniques that are capable of decreasing training compute usage while preserving... model capabilities', exposing loopholes in compute-reporting thresholds. ↩
- arXiv:2403.05812 ↩
- arXiv:2502.09618 ↩
- Anka Reuel, Ben Bucknall, Stephen Casper, Tim Fist, Lennart Heim, et al. (34 authors) (2024) Open Problems in Technical AI Governance, arXiv (cs.CY). arXiv:2407.14981 — Catalogs open problems in 'technical analysis and tools for supporting the effective governance of AI', including compute measurement, verification and reporting gaps. ↩
- Lehdonvirta, Wú & Hawkins (2024) Compute North vs. Compute South: The Uneven Possibilities of Compute-based AI Governance Around the Globe, AIES Proceedings. 10.1609/aies.v7i1.31683 — Census of hyperscale cloud regions shows a divide between "Compute North" states hosting training-relevant compute and a Compute South, shaping who can wield compute-based governance. ↩
- Justin Kollar, Andrew Stokols (2026) Geopolitical ecologies of cloud capitalism: Territorial restructuring and the making of national computing power in the U.S. and China, Environment and Planning A: Economy and Space. 10.1177/0308518X251369704 — US and Chinese drives for sovereign AI/cloud dominance depend on reorganizing land, energy and regulatory systems to sustain large-scale national computing power. ↩
- Stephen Weymouth (2025) Digital Disintegration: Techno-Blocs and Strategic Sovereignty in the AI Era, International Organization. 10.1017/S0020818325101070 — Argues states increasingly assert 'strategic digital sovereignty...through selective alliances with firms and other governments,' fragmenting global AI infrastructure into techno-blocs rather than multilateral order. ↩
- Akash R. Wasil, Tom Reed, Jack William Miller, Peter Barnett (2024) Verification methods for international AI agreements, arXiv (cs.CY). arXiv:2408.16074 — Surveys '10 verification methods that could detect... unauthorized AI training... and unauthorized data centers', mapping the technical basis for compute-disclosure regimes. ↩
- Regulation (EU) 2024/1689, Art. 51(2) + Annex XIII pt. (a)
Cite this article 8 formats · BibTeX, RIS, APA, Chicago, … · 1-click copy
Persistent identifier: https://policywindow.org/wiki/compute-threshold — committed-stable URL with content-versioning via ?asOf= (rollout pending per methodology §7). DOIs via Zenodo are on the roadmap.
Article tools — track changes, suggest an edit
View history — every captured revision of this article · What links here
Social-science evidence — the “so-what”
What the peer-reviewed social science shows: whether the harm this concept addresses is empirically real, and whether governance of it works. The badge is the epistemic status of the evidence(not the policy debate) — “thin” or “absent” efficacy evidence is itself a finding (the “second silence”). Each epistemic-status label is Policy Window's editorial assessment of the cited evidence base (a structured classification), not a verdict any single source issues.
The category is real and coherently defined — training compute (FLOP) is quantifiable, measurable early in the lifecycle, externally verifiable, and correlates with capability, which is why Heim & Koessler 2024 argue it is the most suitable available trigger metric — but its defensibility as a *capability* proxy is genuinely contested: Ho et al. 2024 measure that the compute needed to reach a fixed language-model performance level has halved roughly every 8 months (95% CI ~5–14 months), so a fixed FLOP line drifts relative to capability and the same capability becomes reachable below threshold over time. Caveat: compute is an imperfect-but-real proxy, not a fiction; the dispute is over the threshold's durability, not whether the underlying correlation exists.
Sources: Heim & Koessler 2024 (Training Compute Thresholds: Features and Functions in AI Regulation, arXiv:2405.10799); Ho et al. 2024 (Algorithmic Progress in Language Models, arXiv:2403.05812)
There is no impact evaluation showing that any compute-threshold regime reduces downstream harm: the two flagship thresholds — the US EO 14110 10^26-FLOP reporting requirement (the order was rescinded January 20, 2025, before any efficacy assessment) and the EU AI Act's 10^25-FLOP systemic-risk trigger (Article 51, applicable from August 2025) — are recent and untested, and proponents themselves caution that thresholds should serve only as an initial filter, never in isolation to determine mitigations (Heim & Koessler 2024). Critics further argue, by analogy to tobacco and fossil-fuel precedents, that the demand for proof of efficacy can itself be weaponized to delay regulation (Casper, Krueger & Hadfield-Menell 2025). The evidence that compute-thresholding governance reduces harm is absent.
Sources: Heim & Koessler 2024 (Training Compute Thresholds: Features and Functions in AI Regulation, arXiv:2405.10799); Casper, Krueger & Hadfield-Menell 2025 (Pitfalls of Evidence-Based AI Policy, arXiv:2502.09618)