Computing Power and the Governance of Artificial Intelligence

TL;DR

The paper analyzes how regulating AI-relevant compute can complement AI governance by making development more observable, steering progress toward safer and beneficial uses, and strengthening enforcement of norms. It argues that compute is uniquely regulable due to its detectability, excludability, quantifiability, and concentrated supply chain, and that frontier AI models are especially compute-intensive, following scaling laws that tie performance to compute. The authors outline three governance capacities—visibility, allocation, and enforcement—and present illustrative mechanisms (e.g., reporting, chip registries, and digitally enforceable policies) alongside guardrails to mitigate privacy, economic, and centralization risks. While highlighting potential benefits for public safety and equity, the paper cautions about unintended consequences and underscores the need for careful design, piloting, and ongoing research to adapt policies as technology evolves.

Abstract

Computing power, or "compute," is crucial for the development and deployment of artificial intelligence (AI) capabilities. As a result, governments and companies have started to leverage compute as a means to govern AI. For example, governments are investing in domestic compute capacity, controlling the flow of compute to competing countries, and subsidizing compute access to certain sectors. However, these efforts only scratch the surface of how compute can be used to govern AI development and deployment. Relative to other key inputs to AI (data and algorithms), AI-relevant compute is a particularly effective point of intervention: it is detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain. These characteristics, alongside the singular importance of compute for cutting-edge AI models, suggest that governing compute can contribute to achieving common policy objectives, such as ensuring the safety and beneficial use of AI. More precisely, policymakers could use compute to facilitate regulatory visibility of AI, allocate resources to promote beneficial outcomes, and enforce restrictions against irresponsible or malicious AI development and usage. However, while compute-based policies and technologies have the potential to assist in these areas, there is significant variation in their readiness for implementation. Some ideas are currently being piloted, while others are hindered by the need for fundamental research. Furthermore, naive or poorly scoped approaches to compute governance carry significant risks in areas like privacy, economic impacts, and centralization of power. We end by suggesting guardrails to minimize these risks from compute governance.

Figures (16)

• Figure 1: Summary of the core concepts in the report. Compute is attractive for policymaking because of four properties. These properties can be leveraged to design and implement policies that enable three critical capacities for the governance of AI.
• Figure 2: The AI Triad. The three key technical inputs to AI are data, algorithms, and compute. Human capital is required for all inputs.
• Figure 3: A Simplified AI lifecycle. In the compute-intensive Development stage, the model is designed, trained, and enhanced. The model is then put to use in the Deployment Stage. Many copies of the model can be run during Deployment.
• Figure 4: Training compute used for notable ML models has been doubling every six months since the emergence of the Deep Learning Era. Executive Order 14110 introduced a notification requirement for models trained with more than $10^{26}$ operations (and $10^{23}$ operations if trained on using primarily biological sequence data).
• Figure 5: The importance of compute AI in a historical context. (Data from epochAITrends2023sevillaComputeTrendsThree2022.)