Cap-and-Trade System for AI Regulation
- Cap-and-trade for AI is a regulatory framework that caps AI-relevant resources like compute and emissions and enables trading of allowances.
- The system incentivizes efficiency by allowing firms to sell surplus allowances, penalizing overuse, and addressing market imbalances.
- Effective governance relies on precise resource accounting and benchmarking, such as through CAP trade-offs and sparse computing metrics.
A cap-and-trade system for AI is a regulatory arrangement that places a quantitative limit on an AI-relevant resource, allocates permissions to use that resource, permits exchange of those permissions, and penalizes overuse. In the current literature, the capped quantity is not uniform: some proposals cap AI deployment/inference FLOPs, some seek to cap AI-related greenhouse-gas emissions through emissions trading, and some advocate a compute cap without any trading market. A related systems literature does not propose a legal market, but it treats deployment as operating under explicit resource caps and makes the trade-off among Cost, Accuracy, and Performance (CAP) visible and measurable for sparse AI systems (Bornstein et al., 27 Jan 2026, Hacker, 2023, Miotti et al., 2023, Jiang et al., 16 May 2025).
1. Conceptual scope and policy motivation
The modern discussion of AI cap-and-trade is anchored in a shared diagnosis: AI development is often dominated by hyper-scaling—larger models, more data, more GPUs, and more FLOPs—because using more resources is a simpler path to improved AI performance. The literature identifies three recurring policy concerns. First, high compute costs create accessibility barriers for academics, startups, and smaller firms. Second, when performance gains can be bought through scale, firms have weak incentives to innovate on efficiency. Third, rising AI use increases electricity, cooling, infrastructure use, emissions, and water use, making sustainability a central governance issue (Bornstein et al., 27 Jan 2026, Hacker, 2023).
A second line of argument treats compute as the natural control variable. One proposal uses compute as a proxy for model capability and argues that dangerous AI is likely to emerge from very large-scale computational effort. Another proposal focuses instead on inference/deployment compute, on the ground that a blanket cap on total AI compute would be too harsh and could harm state-of-the-art research. The common premise is that AI governance can be organized around a measurable resource rather than around subjective capability descriptions alone (Miotti et al., 2023, Bornstein et al., 27 Jan 2026).
A concise typology of the main regimes in the literature is given below.
| Regime | Capped quantity | Mechanism |
|---|---|---|
| AI cap-and-trade | Annual deployment/inference FLOPs | AI Allowances, trading, banking, penalties |
| Sustainable AI via ETS | Energy or carbon footprint / GHG emissions | Inclusion in the EU ETS, allowances, sanctions |
| Global compute cap | Compute in FLOPs | Moratorium threshold, danger threshold, verification |
| MoE-CAP governance framing | Cost, Accuracy, Performance under hardware constraints | Benchmarking and resource-allocation trade-offs |
This typology matters because the phrase “cap-and-trade system for AI” can denote either a literal market in AI-related allowances or, more broadly, a cap-based governance architecture for scarce AI resources. The literature is explicit that these are not identical models.
2. Compute-denominated cap-and-trade for AI deployment
The clearest market-based proposal is a cap-and-trade system for AI inference/deployment compute. In that design, a regulator or governing body sets a cap on the amount of AI deployment compute that certain firms may use; firms receive AI Allowances corresponding to that cap; firms that use less than their allocation can sell surplus allowances; firms that need more compute can buy allowances; unused allowances can be banked; and firms that exceed their allowed compute are penalized or taxed (Bornstein et al., 27 Jan 2026).
The allocation rule is built around a benchmark and an assistance factor. The governing body sets a FLOP benchmark , interpreted as an efficiency benchmark such as Watts per FLOP, and each company receives an assistance factor , which can adjust allocations upward or downward to reduce leakage, reward clean energy, penalize fossil-fuel use, or punish prior violations or monopolistic behavior. The paper gives the benchmarked allocation formula as
$A_i = O_i \times B \times C_i \tag{2}$
where is the allowance allocation and is company output. It also discusses grandfathering,
$A_i = \gamma \times H_i \tag{1}$
where is a historical baseline and . The preferred method is benchmarking, because output is not capped but firms are rewarded based on efficiency relative to the benchmark (Bornstein et al., 27 Jan 2026).
The operational link between allowances and compute is expressed as
where 0 is the FLOPs that company 1 can use, 2 is the quantity of AI allowances, and 3 is FLOP-efficiency, defined as watts per FLOP. The more efficient the firm, the more compute it can effectively deploy with the same allowance. This is the mechanism through which the proposal creates a market price for efficiency (Bornstein et al., 27 Jan 2026).
The same paper formalizes the incentive effect in a stylized utility model. Without governance, firms choose FLOP usage 4 by solving
5
with equilibrium
6
With cap-and-trade, the firm can buy or sell FLOPs 7, at market price 8, subject to a maximum allowable FLOP level 9:
0
subject to
1
Using KKT conditions, the equilibrium becomes
2
The paper’s main formal claim, Remark 1, is that rational firms under cap-and-trade use fewer FLOPs than under no governance because
3
The market price of allowances raises the effective marginal cost of compute from 4 to 5, thereby reducing equilibrium FLOP usage (Bornstein et al., 27 Jan 2026).
Within that framework, the proposal is explicitly aimed at accessibility and competition as well as emissions. Efficient firms can sell surplus allowances and create a new revenue stream. The paper emphasizes that efficient smaller firms may stay afloat by selling surplus allowances and that the system could partially offset their lack of direct access to massive compute clusters. It also includes stylized simulations in which FLOP usage is lower under cap-and-trade than with no governance, and utility may be higher under cap-and-trade for some parameter settings, especially when the cap 6 is sufficiently large (Bornstein et al., 27 Jan 2026).
3. Emissions trading and sustainable AI regulation
A parallel proposal situates AI within climate law rather than compute governance. Its core claim is that the most straightforward way to incentivize limits on GHG emissions is to include AI, and potentially ICT more generally, in the EU Emissions Trading System (ETS). The ETS is described as the EU’s primary climate policy tool, built on a cap-and-trade regime in which a declining aggregate ceiling on emissions is combined with trade in allowances, thereby creating a financial incentive to reduce emissions where cheapest (Hacker, 2023).
In this framework, the capped quantity is not FLOPs but the energy or carbon footprint of AI systems. Covered entities would need allowances or permits corresponding to the emissions tied to training and running AI systems. If they exceed the cap, they would face financial sanctions or have to purchase additional allowances. The intended effect is to pressure firms to make AI systems more efficient, use more renewable energy, or reduce unnecessary training and deployment. The paper argues that this route is more robust than a carbon tax in the EU context because the ETS sets clear net emissions targets, creates workable price incentives, is already familiar in EU climate law, and is, in the author’s view, “the easiest and most feasible mechanism” to systematically price AI’s GHG externalities (Hacker, 2023).
The proposal is not limited to simple ETS inclusion. It also advances a differentiated model based on social usefulness classes. On that model, regulators define classes such as high-benefit, medium-range, and low-benefit, then assign soft or hard consumption caps to each class. The paper gives examples: high-benefit uses include medicine, transport, administration, and employment-related uses; low-benefit uses include entertainment, marketing, and advertising. The caps would designate “the amount of energy that can be used, or GHG that may be emitted, to train and run an ML system in that specific sector and use case.” This is a rationing model in which energy or carbon is treated as a scarce regulatory resource and allocated more generously to AI applications with greater public value (Hacker, 2023).
The paper also addresses a technical allocation problem for foundation-model ecosystems. Because a foundation model may be fine-tuned and deployed across many different applications, emissions are not always easy to assign to one downstream use case. Two methods are proposed: allocate a fraction of the foundation model’s costs to each use case, for example “a specific fraction (e.g., 1/1000) of the foundation models costs for each use case,” or measure only the incremental emissions of fine-tuning, deployment, and running the model for a specific use case. This emphasis on allocation methodology shows that the proposal is intended as a technically adaptable cap-and-trade regime rather than a purely symbolic one (Hacker, 2023).
The sustainability literature treats this emissions-based mechanism as part of a broader regulatory toolkit that includes co-regulation, sustainability-by-design, restrictions on training AI models, transparency obligations, sustainability impact assessments, and GDPR-based balancing of climate costs against individual rights. The cap-and-trade component is the strongest ex ante control and is presented as the most intrusive but potentially most effective way to limit AI’s climate impact (Hacker, 2023).
4. Cap-based AI governance without trading
Not all cap-based AI governance creates a market in tradable allowances. A prominent proposal is an international treaty centered on a global compute cap for advanced AI. Its main provision is a ban on the development of AI systems above an agreed-upon computational resource threshold. The treaty is structured as a cap-based regulatory regime that uses computational resources as the limiting quantity and relies on monitoring, verification, reporting, and enforcement. It is therefore structurally similar to a cap-and-trade system in that it sets a ceiling on a measurable resource, but it is not a market-based cap-and-trade system in the classic emissions sense (Miotti et al., 2023).
The treaty operationalizes the regime through two thresholds. The Moratorium Threshold is a hard ceiling above which development is prohibited, with initial value
7
The Danger Threshold is a lower threshold that triggers heightened regulation, with initial value
8
Below 9 FLOP, standard AI development is allowed. Between $A_i = O_i \times B \times C_i \tag{2}$0 and $A_i = O_i \times B \times C_i \tag{2}$1 FLOP, development is allowed only under stronger oversight and safeguards. Above $A_i = O_i \times B \times C_i \tag{2}$2 FLOP, development is prohibited. The treaty defines compute as “the processing power and other electronic resources used to train, validate, deploy, and run artificial intelligence algorithms and models,” and defines FLOP as “single-precision (32-bit) floating point operations” (Miotti et al., 2023).
The compliance architecture is extensive. State Parties must self-report the amount and locations of large concentrations of advanced hardware; allow comprehensive verification of advanced hardware in declared facilities; support detection of undeclared or secret facilities with large concentrations of advanced hardware; and recognize the need for a protocol allowing investigation by independent evaluators within their borders. Article VI proposes negotiations to create an international agency responsible for verifying treaty obligations, ensuring enforcement, researching powerful AI systems, and adjusting thresholds (Miotti et al., 2023).
The treaty goes beyond static prohibition. It requires regulation above the Danger Threshold so that developers and users demonstrate safeguards such as information security requirements, probabilistic risk assessments, predictions of dangerous capabilities, third-party auditing, and other safety and rights protections. It also includes reporting suspected violations to the United Nations Security Council, establishing an international hotline, creating communication channels for civilian AI developers, protecting whistleblowers, and developing and testing emergency response plans capable of stopping a training run before or immediately after it crosses the cap or stopping proliferation of a dangerous model, for example by withdrawing API access (Miotti et al., 2023).
This regime differs from conventional cap-and-trade in five stated respects: no tradable allowances, no market mechanism, binary compliance above the cap, a security-oriented rather than environmental objective, and hardware-centric enforcement. For that reason, it is best described as a cap-based international control regime, not a literal cap-and-trade market (Miotti et al., 2023).
5. Measurement, accounting, and the CAP perspective in sparse AI systems
A cap-and-trade system for AI depends on what is being counted, and a systems paper on sparse Mixture-of-Experts introduces a benchmarking framework that makes this accounting problem unusually explicit. MoE-CAP is presented as a benchmarking and decision framework for sparse Mixture-of-Experts systems that foregrounds the structured tension among Cost (C), Accuracy (A), and Performance (P). The paper’s core claim is that, with current hardware, one cannot generally maximize Cost, Accuracy, and Performance all at once. Sparse MoE systems tend to optimize two of the three dimensions at the expense of the third, a dynamic termed the MoE-CAP trade-off (Jiang et al., 16 May 2025).
The framework is motivated by the architecture of sparse MoE models. Sparsity allows very large parameter counts while activating only a small subset of experts per token, but it also creates highly uneven demands on compute, memory, and communication. Modern MoE serving stacks increasingly rely on heterogeneous resources—GPUs plus CPUs, DRAM plus HBM plus SSDs, and links such as PCIe and NVLink—so the paper argues that realistic benchmarking must account for the full system rather than only GPU cost or dense-model assumptions. The CAP trade-off is then defined over all relevant hardware components, including compute, communication, and memory across heterogeneous tiers; downstream task quality such as exact match, F1, or win rate; and deployment-time efficiency such as TPOT, throughput, latency, bandwidth, or FLOPS (Jiang et al., 16 May 2025).
To visualize the trade-off, the paper introduces the CAP Radar Diagram, used as a comparative benchmark rather than a single score. It categorizes systems as PA-optimized when performance and accuracy are strong but cost is high, PC-optimized when performance and cost are good but accuracy is reduced, and CA-optimized when cost and accuracy are balanced but performance suffers. In one comparison on Qwen3-30B-A3B, SGLang is fastest with strong accuracy but highest cost; K-Transformers reduces cost and improves speed relative to MoE-Infinity but with lower accuracy; and MoE-Infinity maintains the same accuracy as SGLang while reducing purchase cost by nearly 60%, but latency rises by $A_i = O_i \times B \times C_i \tag{2}$3. In another comparison on Qwen3-235B-A22B, MoE-Infinity is most energy-efficient and accurate but slowest; SGLang-FP8 is more than $A_i = O_i \times B \times C_i \tag{2}$4 faster but costs $A_i = O_i \times B \times C_i \tag{2}$5 more power and loses some accuracy; and AWQ offers a middle ground (Jiang et al., 16 May 2025).
The paper’s major technical contribution is the introduction of sparsity-aware metrics. It argues that vanilla MBU and MFU assume dense activation and therefore systematically overestimate memory and FLOP demand when only a subset of experts is used. To correct this, it defines Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU). Conceptually, S-MBU replaces full-model size with the activated size, using the total activated parameter size per token $A_i = O_i \times B \times C_i \tag{2}$6, KV-cache size $A_i = O_i \times B \times C_i \tag{2}$7, and time per output token $A_i = O_i \times B \times C_i \tag{2}$8. S-MFU estimates compute utilization under actual sparse routing rather than pretending that every expert is evaluated, using token throughput $A_i = O_i \times B \times C_i \tag{2}$9, attention FLOPs 0, router FLOPs 1, active expert count 2, and expert-module term 3. The paper explicitly notes that dense models are a special case with 4 and all indicators equal to 1 (Jiang et al., 16 May 2025).
These metrics are then used to derive practical hardware requirements:
5
and
6
The interpretation is that if hardware utilization is only partial, raw theoretical requirement must be scaled upward to obtain the real hardware needed to hit a target latency or throughput. This is directly relevant to cap-based governance because a cap or allowance system can only work if the underlying resource accounting is not systematically inflated by dense-model assumptions (Jiang et al., 16 May 2025).
The empirical implications are substantial. The paper states that full activation of DeepSeek-R1 requires 18,901 GB/s of bandwidth, achievable only on high-end hardware like DGX-H100 with expert parallelism, but at batch size 1 the requirement falls to 1,040 GB/s, making consumer GPUs such as an RTX 4090 feasible with offloading. Under a TPOT SLO of 0.25 s/token, Apple M3 Max can support DeepSeek-V2-Lite at batch size 32, NVIDIA Orin AGX can support batch size up to 16, and Orin NX up to 4. In appendix experiments on Mixtral-8x7B and Qwen, vanilla MBU overestimates memory use by more than 260% in the Mixtral case, while S-MBU matches profiled memory usage with less than 1% error versus HuggingFace Transformers in those tests (Jiang et al., 16 May 2025).
The paper is explicit that MoE-CAP is not a policy mechanism itself. However, it states that, from a governance or “cap-and-trade” perspective, a deployment should be understood as operating under explicit resource caps, and systems can “trade” among cost, accuracy, and performance by shifting burden across memory tiers, compute tiers, and model sparsity. This suggests an analytical bridge between market-based governance and systems benchmarking: the policy question of what to cap is inseparable from the systems question of how to measure the relevant resource (Jiang et al., 16 May 2025).
6. Adaptive mechanism design, implementation concerns, and unresolved questions
A further strand of work does not regulate AI directly but supplies a methodology for designing cap-and-trade mechanisms under strategic adaptation. A carbon market simulator models a government agent that allocates carbon credits and many enterprise agents that produce, invest in emissions reduction, and trade allowances. The market is cast as a hierarchical, model-free multi-agent reinforcement learning (MARL) problem: the government’s allocation problem is an MDP, enterprise behavior is a partially observable stochastic game, and the mechanism itself adapts through learning rather than remaining fixed. The paper explicitly notes that this general pattern could be adapted to AI compute quotas, bandwidth or spectrum allocation, energy usage limits, access to shared infrastructure, or other AI-related governance/resource control problems (Wang et al., 2024).
In that simulator, the government reward is defined as
7
where productivity is measured from enterprise income, equality is derived from a Gini-style measure, and 8 measures excess emissions. The government action specifies the proportion of the total yearly carbon credit allocated in the current period, the allocation weights for each enterprise, and punishment severity; 10% of total credit is reserved for a government-certified project, and the remaining 90% is distributed to enterprises according to weights. The comparison set includes Emission / grandfathering, Emission intensity / benchmarking, and Enterprise size baselines, as well as temporal trajectories labeled Flat, Decreasing, and Convex. The reported qualitative finding is that the MARL policy yields more diversified allocation, more government-certified project construction, more production than some baselines, and fewer excess emissions, thereby showing a better balance of productivity, equality, and climate performance (Wang et al., 2024).
For AI governance, this methodology is significant because many implementation questions remain unresolved in the legal proposals. The compute cap-and-trade paper states that it does not fully specify a real-world monitoring and auditing system, and it leaves open how regulators would measure FLOPs at scale, how to prevent gaming or leakage, whether inference-only caps are sufficient if training remains unconstrained, how to choose benchmarks and assistance factors fairly, whether regulatory burden might still favor incumbents, and how to reconcile efficiency regulation with national security and competitiveness goals. It also warns about AI leakage, the growth vs. regulation tradeoff, and the argument that unrestricted AI development may be strategically necessary (Bornstein et al., 27 Jan 2026).
The sustainable AI literature raises a different but related set of implementation issues: sectoral coverage in the current ETS is incomplete; the scope of an AI-specific ETS would require further work on what counts as AI or ICT, whether the scheme should cover only training or also deployment and inference, whether data centres and supporting infrastructure should be included, and how to deal with cross-sector cloud services and distributed compute. It also emphasizes that transparency is necessary but insufficient: disclosure can support regulators, NGOs, and intermediaries, but it does not itself force reductions (Hacker, 2023).
The treaty literature adds still another layer of difficulty. Compute thresholds are acknowledged to be imperfect proxies because the same capability can be achieved with fewer FLOPs over time due to algorithmic progress, better architectures, better prompting methods, and other efficiency improvements. Verification is hard because it depends on hardware reporting, inspections, detection of secret facilities, and independent evaluators across national borders. Thresholds may therefore need to be revised downward frequently, and the system depends on good-faith international cooperation despite strategic competition (Miotti et al., 2023).
Taken together, these works define a field rather than a single instrument. A cap-and-trade system for AI may denote a market in deployment-compute allowances, an emissions-trading regime for AI’s climate externalities, a cap-based compute prohibition without trading, or a benchmarking framework that exposes the resource trade-offs on which any cap would have to rely. Across these variants, the recurring technical problem is resource accounting; the recurring economic problem is how to price or ration scarcity; and the recurring governance problem is how to preserve innovation while constraining concentrated, inefficient, or high-risk uses of compute and energy (Bornstein et al., 27 Jan 2026, Hacker, 2023, Miotti et al., 2023, Jiang et al., 16 May 2025)