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.
Key Insight: 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.
Key Data Points
Key Insights Summary
Compute Governance Is Already Happening
Governments are already targeting compute through investments in domestic capacity, export controls, compute subsidies, and compute-based reporting thresholds in regulations like the US Executive Order 14110.
Four Properties Make Compute Governance Attractive
Compute is detectable (large-scale AI requires visible infrastructure), excludable (physical hardware can be controlled), quantifiable (easily measured and reported), and has concentrated supply chains (few key producers).
Compute Enables Three Governance Capacities
Compute governance can enhance regulatory visibility into AI development, enable targeted allocation of resources, and improve enforcement of AI regulations and standards.
Supply Chain Concentration Enhances Governability
The AI chip supply chain is extremely concentrated, with TSMC dominating advanced fabrication, ASML having a monopoly on EUV lithography, and NVIDIA controlling most AI chip design.
Compute Thresholds Enable Proactive Governance
Using compute thresholds (like the 10²⁶ operations threshold in US Executive Order 14110) allows policymakers to identify and regulate potentially risky AI systems before deployment.
Risks Require Careful Implementation
Compute governance carries risks including threats to privacy, centralization of power, and potential for evasion. These require guardrails like focusing on industrial-scale compute and privacy-preserving technologies.
Content Overview
Document Contents
Introduction and Summary
Artificial intelligence (AI) has made tremendous strides over the past decade, fueled in large part by a sharp exponential increase of computing power applied to the training of deep neural networks. This increased computing power ("compute") has been a key enabler of the current wave of AI, including large language models and "generative AI."
The central thesis of this paper is that governing AI compute can play an important role in the governance of AI. Other inputs and outputs of AI development (data, algorithms, and trained models) are easily shareable, non-rivalrous intangible goods, making them inherently difficult to control; in contrast, AI computing hardware is tangible and produced using an extremely concentrated supply chain.
Overview of AI Capabilities, AI Governance, and Compute
The three key technical inputs to producing AI capabilities are data, algorithms, and compute, also referred to as the "AI triad." People provide the necessary technical and scientific expertise ("talent," or human capital) to orchestrate the AI triad in order to produce a trained model.
Compute has played a particularly prominent role in recent AI progress. The advent of the deep learning era around 2010-2012 can be attributed to the initial use of GPUs for training AI systems. This enabled AI systems to grow significantly in size, providing the "deep" in "deep learning."
AI governance refers to the study or practice of local and global governance systems that govern or should govern AI research, development, deployment, and use. Compute governance—the topic of this paper—is one tool for AI governance.
Why Compute Governance Is Attractive for Policymaking
Compute is an appealing lever for AI governance for two main reasons. First, compute plays a crucial role in developing and deploying cutting-edge AI systems. All else equal, the amount of compute used is one of the most reliable indicators of the potential impact of a system.
Second, governing compute is technologically feasible. This is a consequence of four features of compute that other inputs to AI progress don't share:
- Detectability: Large-scale AI training and deployment is highly resource intensive, requiring specialized chips in high-performance clusters hosted in large data centers.
- Excludability: The physical nature of hardware makes it possible to exclude users from accessing AI chips.
- Quantifiability: Computational power can be easily measured, reported, and verified.
- Supply chain concentration: AI chips are produced via a highly concentrated supply chain, with key steps dominated by a small number of actors.
Compute Can Enhance Three AI Governance Capacities
Compute can be used to improve society's capacity to govern AI in at least three key ways:
Visibility refers to the ability to understand how actors use, develop, and deploy AI. Compute governance can enhance visibility through mechanisms like required reporting of compute usage, international AI chip registries, and privacy-preserving workload monitoring.
Allocation refers to the ability to direct and influence the trajectory of AI development by changing the distribution of AI capabilities. This includes differentially advancing beneficial AI development, redistributing AI access across countries, changing the overall pace of AI progress, and collaborating on joint AI megaprojects.
Enforcement refers to the ability to respond to violations of norms or laws related to AI. Compute can enhance enforcement through mechanisms like compute caps, hardware-based remote enforcement, preventing risky training runs via multiparty control, and digital norm enforcement.
Risks of Compute Governance and Possible Mitigations
Compute governance carries several risks that require careful consideration:
- Unintended Consequences: Threats to personal privacy, leakage of sensitive information, and risks from centralization of power.
- Feasibility and Efficacy Issues: Algorithmic and hardware progress may reduce governance effectiveness, low-compute models might still develop dangerous capabilities, and there are incentives for evasion and circumvention.
To mitigate these risks, the paper suggests several guardrails:
- Exclude small-scale AI compute and non-AI compute from governance
- Research and implement privacy-preserving practices and technologies
- Only use compute-based controls for risks where ex ante controls are justified
- Periodically revisit controlled computing technologies
- Implement all controls with substantive and procedural safeguards
Conclusion
Compute is a crucial input to AI that possesses unique properties making it particularly amenable to governance interventions. These properties—detectability, excludability, quantifiability, and supply chain concentration—can be leveraged to enhance three key governance capacities: visibility, allocation, and enforcement.
While compute governance holds promise for addressing AI risks, it also carries significant risks of its own, particularly around privacy, centralization of power, and potential for evasion. Careful implementation with appropriate guardrails is essential to realize the benefits while minimizing the downsides of compute-based AI governance.
Appendix: The Compute-Uranium Analogy
The paper explores an analogy between compute and uranium in nuclear governance. Both are:
- Essential inputs to powerful technologies with dual-use potential
- Produced via concentrated supply chains with key bottlenecks
- Physical goods that can be tracked and controlled more easily than intangible knowledge
- Subject to export controls and international monitoring regimes
This analogy suggests that lessons from nuclear governance may inform approaches to compute governance, while also highlighting important differences between the two domains.
Note: The above is only a summary of the paper content. The complete document contains extensive analysis, policy mechanisms, and detailed discussion of risks and mitigations. We recommend downloading the full PDF for in-depth reading.