Big AI’s energy problem triggers call for market controls

A new research paper published as an academic preprint on arXiv argues that without a structural shift in incentives, the artificial intelligence (AI) sector risks becoming both environmentally unsustainable and economically exclusionary.

Titled AI Cap-and-Trade: Efficiency Incentives for Accessibility and Sustainability, the study proposes a market-based governance framework that treats AI computation as a scarce resource, using a cap-and-trade system to reward efficiency, curb excessive energy use, and rebalance competition within the AI ecosystem.

The rising cost of scale in artificial intelligence

Over the past decade, progress in AI has largely followed a single trajectory: scale. Larger models, trained on more data and run on massive clusters of GPUs, have consistently delivered improvements in performance across language, vision, and multimodal tasks. This scaling paradigm has driven many of the breakthroughs that underpin today’s generative AI systems, but it has also introduced structural distortions.

The study documents how the computational requirements of modern AI have grown exponentially. Training frontier models now requires vast quantities of floating-point operations, while deployment at scale demands continuous inference across millions or billions of user interactions. Unlike training, which is episodic, inference represents a persistent and expanding load on data centers, electricity grids, and hardware supply chains.

These demands carry economic consequences. Only a small number of firms can afford the capital investment required to build and operate large-scale AI infrastructure. As a result, advanced AI development has become increasingly concentrated among a handful of technology companies with access to capital, cloud infrastructure, and proprietary data. Smaller firms, startups, and academic researchers face rising barriers to entry, limiting diversity in innovation and slowing independent research.

The environmental impact is equally significant. Energy consumption associated with large AI models contributes directly to carbon emissions, particularly in regions where electricity generation relies on fossil fuels. As AI adoption spreads across sectors, from enterprise software to consumer applications, the cumulative footprint of inference threatens to rival or exceed that of other major digital industries.

The authors argue that current policy debates have failed to address these issues in a coherent way. Discussions often focus on regulating AI outputs, content risks, or model capabilities, while largely ignoring the physical and economic infrastructure that makes large-scale AI possible. In this context, the paper positions computational usage itself as a policy-relevant variable that deserves direct governance.

A cap-and-trade model for AI computation

To address these challenges, the study proposes an AI-specific cap-and-trade system modeled on mechanisms used in environmental regulation. Instead of limiting emissions directly, cap-and-trade systems place a ceiling on total resource use and allow market participants to trade allowances, creating financial incentives for efficiency.

Applied to AI, the proposed framework introduces a cap on total computational usage, measured in floating-point operations, allocated across AI firms through a system of tradable allowances. Companies would receive a defined number of AI computation permits per year, calibrated to account for output levels and baseline efficiency. Firms that use fewer computational resources than their allocation would retain surplus permits that could be sold to others, while firms that exceed their allocation would need to purchase additional allowances.

Crucially, the study focuses the cap-and-trade mechanism on inference rather than training. The authors argue that restricting training could stifle research and experimentation, particularly in academic and early-stage settings. Inference, by contrast, represents the largest and most continuous source of computational demand and emissions, making it a more effective and less disruptive target for regulation.

By pricing computation, the system reshapes incentives across the industry. Rather than competing solely on scale, firms would be rewarded for developing more efficient architectures, optimizing inference pipelines, reducing redundancy, and investing in techniques such as sparsity, quantization, and model compression. Efficiency would become a competitive advantage rather than a secondary consideration.

The paper’s economic analysis shows that under reasonable assumptions, firms operating within a cap-and-trade system would rationally choose lower computational usage than they would in an unregulated environment. At the same time, the system preserves flexibility by allowing firms to decide how to achieve efficiency gains, rather than mandating specific technical approaches.

Importantly, the authors emphasize that the system is designed to improve accessibility as well as sustainability. Smaller and more efficient firms, which often struggle to compete with hyperscalers, could generate revenue by selling unused computation allowances. This creates a new pathway for participation in the AI economy that does not depend on building massive infrastructure.

Implications for competition, sustainability, and AI governance

The proposed AI cap-and-trade model carries implications that extend beyond environmental impact. By introducing scarcity and pricing into computational usage, the framework challenges the assumption that unlimited scaling is the natural or optimal path for AI progress.

From a competition perspective, the study suggests that efficiency-based incentives could counteract the concentration of power in the AI sector. Firms that rely on brute-force scaling would face higher costs, while those that innovate on efficiency could gain relative advantage. This shift could diversify the field of AI developers, supporting startups, research institutions, and regional players that are currently priced out of large-scale deployment.

The framework also reframes sustainability as an economic design problem rather than a moral or reputational issue. Instead of relying on voluntary commitments or corporate pledges, the system embeds environmental considerations directly into market behavior. Firms seeking to maximize profit would have reason to reduce computational waste, aligning private incentives with public goals.

The authors acknowledge potential concerns about implementation. Measuring computational usage accurately, preventing gaming of the system, and coordinating policy across jurisdictions would require careful design. There is also the question of governance authority: determining who sets the cap, how allowances are distributed, and how compliance is enforced.

However, the study states that these challenges are manageable and familiar from other regulated markets. Energy usage, emissions, and financial transactions are already monitored at scale in many industries. Advances in cloud accounting and hardware telemetry make it increasingly feasible to track AI computation with precision.

The paper also addresses fears that regulation could slow innovation or undermine national competitiveness. By focusing on inference and preserving flexibility in how efficiency is achieved, the proposed system avoids prescribing technical solutions or limiting research directions. Instead, it channels innovation toward making AI more efficient, robust, and accessible.

In the broader context of AI governance, the proposal represents a shift from content-centric regulation toward infrastructure-aware policy. As governments debate how to manage the risks and benefits of AI, the study argues that ignoring the physical and economic foundations of AI systems leaves a critical gap. Governance that overlooks energy use and market concentration risks addressing symptoms rather than causes.