> ## Content Index
> Fetch the complete content index at: https://www.aigl.blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# Computing Power and the Governance of Artificial Intelligence
- URL: https://www.aigl.blog/computing-power-and-the-governance-of-artificial-intelligence/
- Published: 2026-09-18T18:59:57.000Z
- Updated: 2026-09-18T18:59:57.000Z
- Description: This paper examines compute as a lever for governing AI development and deployment. It outlines compute’s distinctive properties, policy mechanisms for visibility, allocation and enforcement, and guardrails against privacy, economic and centralisation risks.
- Author: Jakub Szarmach
- Tags: Report, AI Governance, Risk Management, AI Security, Technical Controls, Supply Chain & Third-Party Risk, #aigl-library

## ⚡ Quick Summary

This report, dated 14 February 2024, examines how computing power, or “compute”, can be used to govern artificial intelligence. It argues that AI-relevant compute is an especially useful intervention point because it is detectable, excludable and quantifiable, and because its supply chain is highly concentrated. The paper focuses particularly on compute-intensive frontier models while recognising that data, algorithms and human capital also shape AI capabilities.

Its central framework identifies three governance capacities that compute can enhance: regulatory visibility into AI capabilities and use; allocation of resources towards safe and beneficial development; and enforcement of restrictions on reckless or malicious development and use. It maps illustrative mechanisms for each capacity, including reporting of training-compute use, an international AI chip registry, privacy-preserving workload monitoring, compute subsidies, network-based “compute caps”, remote enforcement and multiparty control. The report concludes that these mechanisms vary substantially in readiness and should be bounded by safeguards, including a focus on industrial-scale AI compute, privacy-preserving practices, periodic review and substantive and procedural protections.

## 🧩 What’s Covered

The report progresses from technical context to policy mechanisms, risks and guardrails.

- **AI inputs and lifecycle:** It defines the “AI triad” of data, algorithms and compute, with human capital required to orchestrate them. It distinguishes development—design, training and enhancement—from deployment, and explains that training is generally the most compute-intensive development stage while inference can account for most total AI compute at scale.
- **Existing compute governance:** It describes current government interventions: domestic compute investment, subsidised compute access, semiconductor export controls and compute-based reporting. The discussion includes U.S. Executive Order 14110 thresholds of more than 10^26 operations, or 10^23 operations for models trained primarily on biological sequence data.
- **Why compute is governable:** It develops four claimed advantages over data, algorithms and talent: large clusters and data centres can be detected; hardware access can be withheld; chip performance and compute use can be measured; and key stages of the advanced-chip supply chain are concentrated. It also considers the difference between governing development and governing deployment.
- **Visibility mechanisms:** It considers using public compute information and geospatial intelligence to estimate capabilities; required reporting by cloud providers and developers; an international registry recording transfers of advanced AI chips; and workload monitoring using techniques such as cryptographic logs and proof-of-learning.
- **Allocation mechanisms:** It discusses directing compute towards beneficial and defensive applications, addressing unequal access between and within countries, changing the overall pace of AI progress through quotas or a compute reserve, and possible international “CERN for AI” models.
- **Enforcement mechanisms:** It sketches limits on chip-to-chip networking, hardware-based remote authorisation, multisignature control of risky training runs, and digitally enforced cloud-compute conditions such as requiring audit results before large-scale access.
- **Risks and guardrails:** It addresses privacy, leakage of strategic and commercial information, economic costs, centralisation, efficiency improvements, dangerous low-compute models and evasion. Its five guardrails include excluding small-scale and non-AI compute, using privacy-preserving technology, limiting ex ante controls to appropriate risks, regularly revisiting thresholds, and providing safeguards.

## 💡 Why it matters?

The report gives policymakers, cloud providers, AI developers and standard-setting bodies a structured way to assess an upstream governance lever rather than relying only on controls at deployment. Its distinction between visibility, allocation and enforcement helps clarify what a proposed compute intervention is meant to achieve and what information or technical capability it would require.

It also stresses that compute thresholds are only proxies for capability and risk. This is practically important for governance designs that use reporting, access conditions or hardware controls: the paper connects such measures to privacy, information-security and anti-centralisation safeguards rather than treating greater monitoring or control as inherently beneficial.

## ❓ What’s Missing

The report deliberately does not prescribe a single policy package. Many of its Section 4 mechanisms are described as exploratory sketches rather than “shovel-ready interventions”, and it says their feasibility, robustness and desirability need further analysis. It does not supply operational specifications for an AI chip registry, privacy-preserving workload monitoring, remote enforcement or multiparty control, nor does it resolve their institutional governance. The analysis also acknowledges that compute is an imperfect proxy for capability: specialised low-compute systems, post-training changes, model-weight release and algorithmic or hardware advances can reduce the reach and durability of compute-based measures. Jurisdiction-specific legal implementation is not developed as a primary focus.

## 👥 Best For

Useful for AI governance and national-security policy teams assessing compute-related interventions; cloud and data-centre operators considering reporting or access controls; and researchers examining verification, privacy and supply-chain questions. It is particularly relevant when comparing measures aimed at frontier-model development with measures focused on deployment or capability evaluation.

## 📄 Source Details

*Computing Power and the Governance of Artificial Intelligence* is an English-language paper dated 14 February 2024, with 19 named authors led by Girish Sastry, Lennart Heim and Haydn Belfield. No publisher, edition or document URL is identified in the supplied pages. The available text extraction covers pages 1–80 of a 104-page PDF, including the report’s main sections, appendices and the start of its references. Figures are available at [https://github.com/lheim/CPGAI-Figures](https://github.com/lheim/CPGAI-Figures?ref=aigl.blog).