Compute Governance and AI Policy

Regulating the New Oil of the Intelligence Age

Compute as the New Oil

The defining resource of the AI age is not data, though data matters. It is not algorithms, though architectures enable capability. It is compute: the computational processing power required to train and run large AI models. Compute has become the scarce, strategic resource that determines who can develop advanced AI and what those systems can do.

The comparison to oil is apt in several respects. Both are foundational inputs to transformative technologies. Both are concentrated in certain geographic regions and controlled by a limited number of entities. Both are subjects of strategic competition, export controls, and national security concerns. Both shape geopolitical relationships and economic power.

But compute differs from oil in crucial ways. It is not a natural resource extracted from the earth but a manufactured capability requiring sophisticated supply chains. It does not deplete with use but becomes more valuable as software improves. And its control is more complex—spread across chip design, fabrication equipment, semiconductor manufacturing, and cloud infrastructure.

Understanding compute governance is essential to understanding AI governance. Who controls the compute controls who can build AI.

The Compute Supply Chain

Modern AI compute depends on a complex global supply chain:

Chip design: The design of GPUs and specialized AI accelerators is dominated by a few companies—NVIDIA, AMD, Intel, and a growing number of startups. Design requires massive R&D investment and specialized expertise.

Fabrication equipment: The machines that manufacture chips—particularly extreme ultraviolet (EUV) lithography—are produced by even fewer companies, primarily ASML in the Netherlands. This is a critical chokepoint in the supply chain.

Semiconductor manufacturing: The actual production of advanced chips occurs at a limited number of facilities (fabs), concentrated in Taiwan (TSMC), South Korea (Samsung), and to a lesser extent the United States and China. Taiwan's dominance is particularly significant.

Advanced packaging: High-performance AI chips require sophisticated packaging techniques to connect multiple dies. This capability is also concentrated among specialized providers.

Cloud infrastructure: Most AI training occurs not on owned hardware but through cloud providers—AWS, Azure, Google Cloud, and others. These companies control access to compute at scale.

Each link in this chain represents a potential point of control or vulnerability for compute governance.

Export Controls and Techno-Nationalism

The strategic significance of compute has triggered an era of techno-nationalism, with export controls as the primary policy tool. The United States, recognizing the centrality of compute to AI capabilities, has implemented increasingly strict controls on advanced AI chip exports.

The October 2022 and October 2023 export control regimes restricted Chinese access to advanced GPUs and the manufacturing equipment needed to produce them. The goal was clear: slow Chinese AI development by denying access to the compute required for frontier model training.

These controls have had significant effects. Chinese AI labs report difficulties obtaining advanced chips. NVIDIA designed cut-down chips for the Chinese market that comply with export restrictions. Chinese investment in domestic semiconductor capabilities accelerated.

But the controls also have limitations. They address training compute but not inference. They can be circumvented through cloud access, smuggling, or domestic production of less advanced but still capable chips. And they create incentives for China to develop independent supply chains, potentially fragmenting the global semiconductor ecosystem.

The export control approach represents a bet that maintaining compute advantage will provide sustained strategic leverage. Whether this bet pays off depends on the pace of Chinese catch-up and the evolution of AI efficiency—whether capabilities can be achieved with less compute than currently required.

Domestic Investment and Industrial Policy

Recognition of compute strategic significance has triggered massive domestic investment in semiconductor capabilities:

The CHIPS Act: The United States allocated $52 billion to domestic semiconductor manufacturing, aiming to reshore advanced fabrication and reduce dependence on Taiwan. Similar initiatives exist in Europe, Japan, and South Korea.

National AI Research Resources: Proposals for government-funded compute infrastructure to enable academic and public interest AI research, addressing the concentration of large-scale training capability in private companies.

Sovereign AI initiatives: Multiple countries developing national AI compute infrastructure to ensure domestic capabilities and reduce dependence on foreign cloud providers.

These investments reflect a shift from market-led to state-directed semiconductor policy. The era of purely commercial decision-making in this sector is ending; strategic considerations increasingly dominate.

The effectiveness of these investments remains uncertain. Building semiconductor manufacturing capacity takes years. The expertise required is scarce and concentrated. And the economic efficiency of government-led industrial policy in this sector is historically mixed.

International Coordination Challenges

Effective compute governance requires international coordination, but achieving it is challenging:

Divergent interests: Countries differ in their relationships to the compute supply chain. Taiwan and South Korea benefit from their central positions. Europe has design capabilities but limited manufacturing. China is the target of controls. Coordinating policy across these divergent interests is difficult.

Asymmetric dependencies: The United States depends on Taiwan for advanced manufacturing; Taiwan depends on American design tools and military protection. China depends on foreign equipment; foreign companies depend on the Chinese market. These dependencies create complex negotiation dynamics.

Definition challenges: What counts as "advanced" AI chips requiring control? As capabilities evolve, controls must be updated. The definitions that make sense today may not make sense tomorrow.

Enforcement difficulties: Compute is intangible and can move across borders through cloud access. Physical chip controls can be circumvented. Verifying compliance is challenging.

Despite these difficulties, some coordination has emerged. The US has aligned with Japan and the Netherlands on equipment controls. The EU is developing its own export control frameworks. Multilateral discussions on AI governance increasingly include compute considerations.

Safety Standards and Responsible Development

Beyond national competition, compute governance is increasingly tied to safety considerations. The argument: advanced AI capabilities require significant compute, so controlling access to large-scale compute provides leverage for enforcing safety standards.

Several mechanisms have been proposed:

Training run registration: Requirements to report large-scale training runs to regulators, enabling oversight of frontier model development.

Safety evaluation mandates: Requirements for safety evaluation before deployment, potentially including third-party auditing.

Compute threshold triggers: Automatic regulatory requirements triggered by the scale of compute used in training, on the assumption that more compute enables more capable and potentially more dangerous models.

Compute caps: Limits on the total compute that can be used for single training runs, as a direct constraint on capability development.

Allocation mechanisms: Government involvement in allocating scarce compute to approved projects meeting safety and public benefit criteria.

These mechanisms raise governance questions. Who decides what safety standards apply? How are evaluations conducted and by whom? What rights of appeal exist for denied compute access? The concentration of compute oversight represents concentration of power over AI development.

Democratic Oversight and Public Interest

The concentration of compute in private corporations and its emergence as an instrument of state policy raises fundamental questions about democratic oversight:

Corporate concentration: Most frontier AI training occurs in a handful of large tech companies. Their decisions about what to build, how to train, and what to deploy shape AI capabilities without public accountability.

State power: Government control over compute access provides leverage over AI development that could be used for surveillance, censorship, or suppression of dissent as well as safety.

Public access: Meaningful public participation in AI development requires access to compute. Currently, this access is severely limited, creating an AI elite of those with resources to train models.

Global equity: Compute concentration reinforces global inequality. Wealthy countries and corporations monopolize advanced AI capabilities while the global majority is limited to API access or inferior systems.

Addressing these concerns requires institutional innovation: public compute resources, democratic governance mechanisms, transparency requirements, and international frameworks that distribute AI capability more equitably.

Unhinged View: The Folly of Compute Control

The emerging strategy of compute governance rests on a flawed assumption: that controlling the physical substrate of AI will control AI development. This assumption is increasingly incorrect and will become more so over time.

First, the compute required for training is declining in relative importance. Algorithmic improvements, training efficiency, and model compression enable more capability with less compute. The compute threshold for dangerous capabilities is falling, not rising.

Second, compute is becoming more distributed. Specialized chips, edge devices, and federated learning enable training without centralized data center scale. The chokepoint is temporary, not permanent.

Third, compute controls are creating the very threats they aim to prevent. By fragmenting the global semiconductor ecosystem, they increase the risk of conflict over Taiwan. By forcing China to develop independent capabilities, they create a bifurcated AI landscape with reduced communication and coordination.

Fourth, compute governance concentrates power in ways that threaten the values it claims to protect. The same infrastructure used for safety enforcement can be used for surveillance and control. The entities empowered by compute oversight are not necessarily aligned with public interest.

The proper approach is not compute control but capability resilience: ensuring that beneficial applications of AI outpace harmful ones, regardless of who has access to compute. This requires differential progress, not restriction. It requires investment in beneficial applications, not just control of dangerous ones.

Compute is not the new oil in the sense that controlling it provides lasting strategic advantage. It is the new oil in the sense that dependence on it creates vulnerabilities that smart policy would reduce rather than reinforce.

Alternative Governance Approaches

Beyond compute-centric governance, several alternative approaches merit consideration:

Model-based governance: Regulating systems based on demonstrated capabilities rather than training resources. Capability thresholds trigger requirements regardless of how the capability was achieved.

Application-based governance: Focusing on specific high-risk applications—bioweapons research, election manipulation, critical infrastructure control—rather than attempting to govern AI development generally.

Outcome-based governance: Regulating based on actual harms caused rather than potential risks from development. Liability frameworks that hold deployers accountable for damages.

Transparency governance: Requirements for disclosure of training data, model architecture, and evaluation results—enabling public and researcher oversight without direct control of development.

Insurance and bonding: Economic mechanisms that require developers to demonstrate financial capacity to cover potential harms, creating incentives for safety investment.

Each approach has advantages and limitations. A portfolio of mechanisms, adapted to different contexts, is likely more effective than reliance on compute control alone.

The Future of Compute Access

Several trends will shape future compute access:

Efficiency improvements: Continued algorithmic and hardware efficiency will reduce compute requirements for given capabilities. The democratization enabled by open source will accelerate.

Domestic capabilities: More countries will develop domestic semiconductor ecosystems, reducing dependence on the concentrated supply chain. The era of Taiwan-centric manufacturing may end.

Alternative substrates: Beyond silicon, new computing substrates—optical, neuromorphic, quantum—may emerge, changing the nature of AI compute and potentially disrupting current control strategies.

Distributed training: Techniques for training across distributed, potentially consumer-grade hardware will mature, enabling collective model training without centralized compute resources.

Cloud evolution: Cloud providers will remain central, but their relationship to users may evolve—from infrastructure providers to capability platforms with varying degrees of governance integration.

Key Takeaways

Compute has become the strategic resource governing AI capability, with a concentrated global supply chain creating chokepoints for both commercial and policy leverage.
Export controls and techno-nationalism have emerged as primary policy responses, attempting to maintain capability advantages through restricting access to advanced chips and manufacturing equipment.
Domestic industrial policy investments aim to reshore manufacturing and reduce supply chain dependencies, but face significant implementation challenges and time lags.
International coordination on compute governance is limited by divergent national interests, asymmetric dependencies, and enforcement difficulties across a globalized supply chain.
Compute-based safety governance raises concerns about concentration of power, with the same mechanisms usable for safety enforcement potentially enabling surveillance, censorship, and suppression of beneficial development.
Alternative governance approaches focusing on capabilities, applications, and outcomes may prove more effective and less prone to abuse than compute-centric control strategies.

References

Khan, S. (2022). "AI Chips: What They Are and Why They Matter." Center for Security and Emerging Technology. Comprehensive analysis of AI compute landscape.
Sastry, G., et al. (2024). "Computing Power and the Governance of AI." arXiv:2402.08797. Analysis of compute governance mechanisms and alternatives.
Ahmed, N., & Waheed, M. (2023). "The Grand Strategy of AI: Computing Power and the Governance of Frontier AI Systems." MIT Tech Review. Policy analysis of compute control strategies.
Heim, L. (2023). "Estimating Training Compute of Deep Learning Models." Epoch AI. Methodology for measuring and comparing AI training compute.
Villalobos, P., et al. (2022). "Will We Run Out of Data? An Analysis of the Limits of Scaling Datasets in Machine Learning." Epoch AI. Analysis of compute-data relationships in model training.
Sevastopulo, D., & Fildes, N. (2022). "US Curbs Exports of High-Performance AI Chips to China." Financial Times. Coverage of export control implementation.
Miller, C. (2022). Chip War: The Fight for the World's Most Critical Technology. Scribner. Historical analysis of semiconductor geopolitics.
Bureau of Industry and Security. (2023). "Export Controls on Advanced Computing and Semiconductor Manufacturing Items." Federal Register. Official export control regulations.
Shavit, Y. (2023). "Introduction to the U.S. AI Chip Ban and Its Effects on China." Center for Security and Emerging Technology. Analysis of export control impacts.
Erdil, E., & Besiroglu, T. (2023). "Algorithmic Progress in Language Models." arXiv:2310.12305. Analysis of efficiency improvements in AI training.

This essay represents a viewpoint within the UnhingedAI Collective. Compute governance is the new frontier of AI policy—but the attempt to govern through controlling physical resources may be fighting the last war while the battle moves to domains of efficiency and distribution.