Within Thresholds
Will Today's Compute Limits Still Matter Tomorrow?
Fixed thresholds may become outdated as frontier training runs become more common and capabilities advance.
On this page
- Why thresholds drift over time
- Forecasts for larger training runs
- Options for updating thresholds
Page outline Jump by section
Introduction
A recurring question in debates about AI governance and existential risk is whether fixed compute thresholds can keep pace with rapid AI progress. Compute‑based reporting thresholds — rules requiring extra oversight once a model’s training uses more than a certain number of computational operations — are attractive because they are measurable and seem technically grounded. Yet the AI landscape is evolving fast: hardware improves, training techniques become more efficient, and the number and variety of high‑compute projects mushroom. That raises a real concern within the AI doom and safety conversation: can a static number really remain a reliable signal of future danger, or will thresholds become outdated as capabilities outgrow them? Current evidence suggests that without adaptive updating or contextualisation, fixed thresholds are likely to drift behind the frontier they’re meant to capture and may weaken as an early warning trigger. [Regulations.ai]regulations.aiCompute ThresholdAI Regulation Glossary | Regulations.ai…
Why Fixed Compute Thresholds Drift Over Time
Training compute thresholds are grounded in the idea that larger amounts of compute generally produce more capable models. Regulators in the United States and the European Union have set milestone values — for example, about 10²⁵–10²⁶ floating‑point operations — above which models trigger reporting or scrutiny requirements. These values aim to capture models at the frontier of capability, which are most likely to pose novel or high‑impact risks. [Regulations.ai]regulations.aiCompute GovernanceDEFINITION Compute Governance refers to regulatory and policy approaches that lev…
However, several trends challenge the long‑term stability of a fixed threshold:
- Algorithmic and efficiency improvements: As AI research proceeds, better architectures, optimisation techniques, and training shortcuts can extract greater capability from the same or even smaller compute budgets. This means two models trained with the same amount of compute may differ widely in ability; a threshold that once reliably captured the most capable models may over time miss increasingly capable models trained more efficiently. [AI Security & Safety Directory]aisecurityandsafety.orgSource details in endnotes.
- Rapid proliferation of high‑compute projects: Forecasting work suggests the number of models exceeding current threshold levels could explode within a few years. One analysis projects that models above a 10²⁶ FLOP threshold — once rare — might grow from a handful to hundreds within a few years, making static thresholds less selective and more administratively burdensome. [Epoch AI]epoch.aiAIHow many AI models will exceed compute thresholds? | Epoch AIHow many AI models will exceed compute thresholds? | Epoch AI…
- Shifts in how AI scales: Some emerging research indicates that new modes of improving performance — such as inference‑time scaling rather than traditional training compute — could reshape how capabilities grow. If capability starts to decouple from the specific measure of training FLOPs, thresholds tied only to training compute risk becoming less predictive of real‑world power. [arXiv]arxiv.orgarXiv Inference Scaling Reshapes AI GovernancearXivInference Scaling Reshapes AI GovernanceFebruary 12, 2025…
These dynamics illustrate that thresholds based on a fixed compute value tend to drift relative to the evolving frontier. As models become more efficient or as the ecosystem grows more diverse, a threshold that once signalled cutting‑edge capability and risk may become either too weak or too blunt to serve its regulatory purpose.
Forecasts for Larger Training Runs
Empirical projections help make this drift concrete. Research that models future training compute trends estimates that the number of AI models exceeding absolute compute thresholds — those defined in fixed FLOP terms — will scale up sharply over the next few years. Under plausible development scenarios, the count of models over a 10²⁵ FLOP level is expected to grow dramatically by 2030, and even higher thresholds see similar growth. [Epoch AI]epoch.aiAIHow many AI models will exceed compute thresholds? | Epoch AIHow many AI models will exceed compute thresholds? | Epoch AI…
This matters for both governance and safety signalling:
- Regulatory burden: A threshold intended to catch only a handful of frontier projects could become so widely crossed that it no longer effectively distinguishes between high‑risk systems and more routine advanced models.
- Temporal lag: Thresholds set based on the state of the art at one point may quickly fall behind the frontier, providing an outdated snapshot of what counts as “high compute” without active revision.
These forecasts underscore how AI progress is not linear or static, and any governance tool tied closely to the scale of training runs must factor in both quantity and pace of change.
Options for Updating Thresholds
Acknowledging that compute thresholds drift over time, researchers and policymakers are exploring mechanisms to keep them relevant:
- Periodic revision authority: Some legislative frameworks explicitly empower regulators to raise or adjust threshold values regularly as technology evolves. This makes thresholds “living” rather than fixed administrative numbers. [Institute for Law & AI]law-ai.orgInstitute for Law & AIThe Role of Compute Thresholds for AI GovernanceInstitute for Law & AIFebruary 1, 2025…
- Effective or relative compute metrics: Instead of absolute FLOP counts, proposals have been floated to define thresholds relative to the current frontier — for example, relative to the largest training run worldwide at a given time — so that thresholds scale with the frontier itself. While challenging to implement, this could align governance more directly with capability benchmarks. [Institute for Law & AI]law-ai.orgInstitute for Law & AIThe Role of Compute Thresholds for AI GovernanceInstitute for Law & AIFebruary 1, 2025…
- Complementary metrics: Many analysts argue that thresholds should not stand alone but be combined with other indicators — such as performance evaluations, misuse potential assessments, or fine‑tuning trajectories — to capture risk more comprehensively. [governance.ai]governance.aitraining compute thresholds features and functions in ai regulationTraining Compute Thresholds: Features and Functions in AI Regulation | GovAIAugust 7, 2024…
- Algorithmic efficiency adjustments: In theory, thresholds could be calibrated not just on raw compute but on effective compute that accounts for algorithmic improvements. However, there is no agreed method for such normalisation, and measuring efficiency itself is complex. [Institute for Law & AI]law-ai.orgInstitute for Law & AIThe Role of Compute Thresholds for AI GovernanceInstitute for Law & AIFebruary 1, 2025…
Taken together, these options illustrate a common theme: thresholds must be dynamic and contextualised if they are to remain a meaningful signal of when heightened scrutiny is appropriate.
Implications for AI Safety and Risk Monitoring
Within discussions about AI doom and existential risk, the role of compute thresholds as an early warning trigger hinges on their continuing relevance against a shifting landscape. History shows that AI progress rarely honours static assumptions: hardware and software innovations continually reshape what is possible with a given resource budget.
If thresholds lag too far behind model capabilities, they risk becoming symbolic rather than practical triggers, failing to capture systems that might exhibit dangerous autonomy or loss of human control. Adaptive mechanisms — regular updates, relative definitions, and hybrid metrics — sharpen thresholds’ ability to highlight projects warranting deeper scrutiny before deployment.
But importantly, these mechanisms also reveal that thresholds are a screening tool, not a final arbiter. They flag candidates for oversight and evaluation, and must be paired with deeper assessments of capability, alignment, and misuse potential to meaningfully contribute to managing existential risk from advanced AI. [governance.ai]governance.aitraining compute thresholds features and functions in ai regulationTraining Compute Thresholds: Features and Functions in AI Regulation | GovAIAugust 7, 2024…
In this sense, compute thresholds that evolve along with technical progress are less about “keeping pace” as a static guardrail and more about feeding continuous safety evaluation pipelines that evolve as rapidly as the frontier they monitor.
Amazon book picks
Further Reading
Books and field guides related to Will Today's Compute Limits Still Matter Tomorrow?. Use these as the next step if you want deeper reading beyond the article.
The Coming Wave
Directly addresses how governance mechanisms must adapt as capabilities advance.
Human Compatible
Frames long-term control problems that static thresholds may fail to capture.
Superintelligence
Highlights why governance thresholds may require continual revision.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: regulations.ai
Title: Compute Threshold
Link: https://regulations.ai/glossary/compute-thresholdSource snippet
AI Regulation Glossary | Regulations.ai...
-
Source: epoch.ai
Title: AIHow many AI models will exceed compute thresholds? | Epoch AI
Link: https://epoch.ai/publications/model-counts-compute-thresholdsSource snippet
How many AI models will exceed compute thresholds? | Epoch AI...
-
Source: arxiv.org
Title: arXiv Inference Scaling Reshapes AI Governance
Link: https://arxiv.org/abs/2503.05705Source snippet
arXivInference Scaling Reshapes AI GovernanceFebruary 12, 2025...
Published: February 12, 2025
-
Source: law-ai.org
Title: Institute for Law & AIThe Role of Compute Thresholds for AI Governance
Link: https://law-ai.org/the-role-of-compute-thresholds-for-ai-governance/Source snippet
Institute for Law & AIFebruary 1, 2025...
Published: February 1, 2025
-
Source: governance.ai
Title: training compute thresholds features and functions in ai regulation
Link: https://www.governance.ai/research-paper/training-compute-thresholds-features-and-functions-in-ai-regulationSource snippet
Training Compute Thresholds: Features and Functions in AI Regulation | GovAIAugust 7, 2024...
Published: August 7, 2024
-
Source: blog.heim.xyz
Title: training compute thresholds
Link: https://blog.heim.xyz/training-compute-thresholds/Source snippet
Compute Thresholds — Features and Functions in AI RegulationApril 6, 2024 — TRAINING COMPUTE THRESHOLDS — FEATURES AND FUNCTIONS IN AI RE...
Published: April 6, 2024
-
Source: regulations.ai
Title: Compute Governance
Link: https://regulations.ai/glossary/compute-governanceSource snippet
DEFINITION Compute Governance refers to regulatory and policy approaches that lev...
-
Source: aisecurityandsafety.org
Link: https://aisecurityandsafety.org/en/glossary/compute-threshold/ -
Source: longtermwiki.com
Title: Compute Thresholds | Longterm Wiki
Link: https://www.longtermwiki.com/wiki/thresholdsSource snippet
January 30, 2026 — COMPUTE THRESHOLDS Concept COMPUTE THRESHOLDS Comprehensive analysis of compute thresholds (EU: 10^25 FLOP, US: 10^26...
Published: January 30, 2026
Additional References
-
Source: themoonlight.io
Link: https://www.themoonlight.io/en/review/training-compute-thresholds-features-and-functions-in-ai-regulationSource snippet
FEATURES AND FUNCTIONS IN AI REGULATION Here's a detailed breakdown of the paper, including the core methodology in more technical terms...
-
Source: aimodels.fyi
Link: https://www.aimodels.fyi/papers/arxiv/training-compute-thresholds-features-functions-ai-regulationSource snippet
TRAINING COMPUTE THRESHOLDS: FEATURES AND FUNCTIONS IN AI GOVERNANCE Published 8/7/2024 by Lennart Heim, Leonie Koessler OVERVIEW * The...
-
Source: cset.georgetown.edu
Title: If the paradigm of AI development of recent years continues, novel capabilities
Link: https://cset.georgetown.edu/article/regulating-the-ai-frontier-design-choices-and-constraints/Source snippet
the AI Frontier: Design Choices and Constraints | Center for Security and Emerging TechnologyOctober 26, 2023 — COMPUTE THRESHOLDS COULD...
Published: October 26, 2023
-
Source: fenwick.com
Title: Technological Challenges for Regulatory Thresholds of AI… | Fenwick
Link: https://www.fenwick.com/insights/publications/interesting-developments-for-regulatory-thresholds-of-ai-computeSource snippet
June 20, 2024 — JUNE 20, 2024 TECHNOLOGICAL CHALLENGES FOR REGULATORY THRESHOLDS OF [AI COMPUTE]({{ 'compute-kyc/' | relative_url }}) By: Zach Harned WHAT YOU NEED TO KNOW * Ge...
Published: June 20, 2024
-
Source: axi.lims.ac.uk
Link: https://axi.lims.ac.uk/paper/2405.10799Source snippet
Compute Thresholds: Features an...May 17, 2024 — ID: 2405.10799 ID: 2405.10799 Search TRAINING COMPUTE THRESHOLDS: FEATURES AND FUNCTIONS...
Published: May 17, 2024
-
Source: emergentmind.com
Title: Limitations of Compute Thresholds in AI Governance
Link: https://www.emergentmind.com/articles/2407.05694Source snippet
July 8, 2024 — ON THE LIMITATIONS OF COMPUTE THRESHOLDS AS A GOVERNANCE STRATEGY Published 8 Jul 2024 in cs.AI, cs.CL, cs.ET, and cs.LG |...
Published: July 8, 2024
-
Source: emergentmind.com
Title: Compute Thresholds in AI and Beyond
Link: https://www.emergentmind.com/topics/compute-thresholdsSource snippet
January 11, 2026 — COMPUTE THRESHOLDS IN AI AND BEYOND Updated 11 January 2026 * Compute thresholds are quantitative limits defined by op...
Published: January 11, 2026
-
Source: aisecurityandsafety.org
Title: compute governance
Link: https://aisecurityandsafety.org/en/guides/compute-governance/Source snippet
Controlling AI Through Hardware & Compute Access (2026) | AI Safety DirectoryApril 3, 2026 — COMPUTE GOVERNANCE: CONTROLLING AI THROUGH H...
Published: April 3, 2026
-
Source: youtube.com
Link: https://www.youtube.com/watch?v=FOHtGeUjgZYSource snippet
Lennart Heim: Governing the bottleneck of AI...
-
Source: ai-safety-atlas.com
Title: Compute Governance
Link: https://ai-safety-atlas.com/chapters/v1/governance/compute-governanceSource snippet
The most reliable is energy consumption - training runs that...
Topic Tree