Skip to content

AI Cap-and-Trade: Efficiency Incentives for Accessibility and Sustainability

Conference: ICML 2026
arXiv: 2601.19886
Code: None (position + theoretical analysis paper)
Area: AI Governance / Economics / Sustainability
Keywords: Cap-and-trade, FLOP quota, AI accessibility, energy consumption incentives, KKT analysis

TL;DR

Drawing inspiration from carbon cap-and-trade, the authors propose a quota-trading market for AI inference FLOPs (AI Allowance). Using KKT conditions, they prove that under reasonable parameters, this mechanism strictly reduces FLOP usage across companies, thereby simultaneously mitigating both the energy consumption of large models and the market exclusion of smaller companies.

Background & Motivation

Background: The large model paradigm is trending toward ultra-large scale—bigger models, more data, more GPUs. OpenAI processes ~2.5 billion queries per day, with annual inference consuming about 1 ronnaFLOP (\(10^{27}\) FLOPs), requiring around 120,000 H100s. Each ChatGPT/Gemini query uses 0.24–0.34 Wh, with OpenAI consuming ~850,000 kWh and ~350 tons CO₂ daily, far exceeding the EPA "major polluter" threshold (100 tons/year).

Limitations of Prior Work: (1) Academia and small companies are squeezed out by GPU costs, with 70% of AI PhDs moving to industry; (2) Data center energy consumption is projected to double to 1000 TWh by 2030, with water usage reaching 120 billion liters; (3) Existing AI governance (EU AI Act, California SB-1047) focuses on compliance and safety, with almost no market-based "efficiency incentive" mechanisms.

Key Challenge: The AI industry is inherently biased toward "hyper-scaling > efficiency"—as long as compute is affordable, there is no external cost compelling energy savings. The negative externality of energy consumption is not priced in.

Goal: To design a market-based mechanism that endogenizes efficiency as economic value, making "achieving the same inference with fewer FLOPs" a tradable asset, while avoiding the blunt approaches of Pigouvian taxes ("AI leakage" risk) and outright bans (which stifle innovation).

Key Insight: Directly transplanting the mature carbon cap-and-trade mechanism (EU ETS, California, China, Korea) to AI—the core unit shifts from "carbon emission" to "AI Allowance" (electricity/FLOP quota for inference); allocation uses benchmarking instead of grandfathering, avoiding AI analogues of "carbon leakage."

Core Idea: The government allocates AI Allowance for free to each company based on their FLOP output × industry watts-per-FLOP benchmark × company-specific assistance factor (\(A_i = O_i \cdot B \cdot C_i\)). Companies can buy/sell/store quotas. Theoretically, after introducing trading constraints, the optimal FLOP usage \(x^\ast\) for rational companies is strictly less than in the absence of the mechanism.

Method

Overall Architecture

The paper first quantifies "why market incentives are needed" (Sec 1–3), then reviews existing market-based methods (Sec 4: Pigouvian tax, user fee, credit/subsidy, deposit-refund, tradable permits), and finally proposes an AI cap-and-trade (Sec 5): capping only inference FLOPs, benchmarking-based allocation + secondary market trading, KKT-based proof of strict FLOP reduction, and numerical experiments under two buy/sell price settings.

Key Designs

  1. AI Allowance Allocation Mechanism (Benchmarking + Assistance Factor):

    • Function: Localizes the benchmarking method from carbon ETS to AI, avoiding historical unfairness from grandfathering and preventing one-size-fits-all allocation from squeezing out large companies.
    • Mechanism: For each regulated company \(i\), the quota is \(A_i = O_i \cdot B \cdot C_i\). \(O_i\) is its FLOP output (e.g., a two-year rolling average with 15% adjustment as in EU ETS), \(B\) is the industry benchmark watts-per-FLOP (modeled after "top 10% best/average 90%" in carbon markets), \(C_i\) is the company-specific assistance factor: \(C_i > 1\) for clean energy use, \(C_i < 1\) for fossil fuel use or violations. Given company efficiency \(E_i\) (watts/FLOP), the actual allowed FLOPs are \(F_i = A_i / E_i\).
    • Design Motivation: Uniform allocation would force companies like OpenAI with hundreds of millions of users to artificially throttle, causing "quota-induced regression." Benchmarking gives large companies reasonable FLOP space but requires them to buy quotas if inefficient, while efficient small companies can sell surplus quotas—neither crowding out incumbents nor locking out startups.

  2. Secondary Market + Allowance Banking:

    • Function: Makes efficiency itself a tradable cash asset and allows companies to smooth fluctuations across years.
    • Mechanism: The government is the primary market (free allocation), while companies freely trade quotas on the secondary market (like the European Energy Exchange/Korea Exchange); companies exceeding quotas must purchase or face heavy penalties. Surplus quotas can be banked for the next year. The paper frames this new cash flow as a "breathing room revenue stream" for startups, helping them survive the burn rate period.
    • Design Motivation: Empirical evidence from carbon markets (e.g., EU ETS) shows that the secondary market turns "emission reduction" into a profit opportunity, serving as the most effective lever for incentivizing efficiency innovation. Transferring this logic to AI, "training efficient small models" itself gains market value, potentially reversing the current incentive distortion of "compute sufficiency is all that matters."

  3. Rational Company Equilibrium & FLOP Reduction Proof (KKT Conditions):

    • Function: Rigorously proves within a game-theoretic framework that introducing trading leads to lower equilibrium FLOP usage, providing mathematical backing for policymakers.
    • Mechanism: Models single-company utility as \(u(x) = -x^{-k} - ax\) (\(x\) is FLOPs, \(-x^{-k}\) reflects diminishing returns on performance, \(a\) is cost-per-FLOP). Without the mechanism, \(\nabla u = 0\) yields \(x^\ast = (k/a)^{1/(k+1)}\). Introducing trading variable \(y\) (>0 sell, <0 buy) with price \(b\) and quota cap \(F_i\), the constrained problem is \(\max u(x,y) = -x^{-k} - ax + by\quad\text{s.t.}\ x+y \le F_i,\ x \ge 0\). The Lagrangian first-order condition gives \(\mu_1 = b\), and complementary slackness yields \(x^\ast = (k/(a+b))^{1/(k+1)}\) and \(y^\ast = F_i - x^\ast\). Since \(b > 0\), \(x^\ast_{\text{cap}} < x^\ast_{\text{no cap}}\)—the trading price \(b\) adds the "opportunity cost" of buying/selling quotas to the cost-per-FLOP, strictly reducing optimal FLOP usage.
    • Design Motivation: Many AI governance proposals are political arguments lacking economic validation. Providing a closed-form solution allows policymakers to see the sensitivity of "how much reduction" to \(b\), and to identify which parameters (e.g., \(b\) set too low, \(a\) too high) may cause the reduction target to fail.

Loss & Training

Not applicable. The paper is position + economic modeling; "training" only occurs in numerical experiments—scanning \(x^\ast_{\text{no cap}}\) vs \(x^\ast_{\text{cap}}\) for different \(a\) (cost-per-FLOP), plotting curves under two price settings \(b=10^{-2}\) (fixed) and \(b = \sqrt{a}\) (cost-scaled) (Fig 1).

Key Experimental Results

Main Results

The authors plot equilibrium FLOP usage \(x^\ast\) for companies under different cost-per-FLOP \(a\) (Fig 1):

Scenario Key Parameters \(x^\ast_{\text{no cap}}\) vs \(x^\ast_{\text{cap}}\) Conclusion
Fixed buy/sell price \(b=10^{-2}\) Scan \(a \in [10^{-4}, 10^{-1}]\) \(x^\ast_{\text{cap}} < x^\ast_{\text{no cap}}\) always holds Any \(a > 0\) reduces FLOPs
Cost-scaled \(b=\sqrt{a}\) Same as above \(x^\ast_{\text{cap}} < x^\ast_{\text{no cap}}\), with greater reduction ratio Coordinated increase of \(b\) with \(a\) intensifies efficiency pressure

Ablation Study

Configuration Key Findings Interpretation
No cap (baseline) \(x^\ast = (k/a)^{1/(k+1)}\) Companies only weigh performance vs direct cost
Cap added, \(b \to 0\) \(x^\ast \to x^\ast_{\text{no cap}}\) Mechanism fails if quotas are worthless
Cap added, large \(b\) \(x^\ast\) significantly decreases Higher prices yield more reduction, but risk industry impact
Benchmarking vs grandfathering Benchmarking better incentivizes efficiency Consistent with real carbon market observations (Yang 2020, Wang 2022)
Cap only training vs cap only inference Capping inference is more practical Training cap stifles frontier research; inference cap reduces emissions and shifts cost pressure to the largest cash flow source

Key Findings

  • The closed-form solution clearly shows \(b\) (trading price) is the most sensitive parameter: too low and the mechanism is "toothless," too high and it suppresses the industry.
  • Choosing to "cap only inference FLOPs" is a key policy tradeoff—because inference dominates emissions (Schmidt 2021, De Vries 2023, Jegham 2025), and it protects innovation on the training side.
  • DeepSeek is a natural empirical case: US chip export restrictions to China effectively create a cap, driving efficiency innovations like MoE and MLA; this paper formalizes the causal path "market constraint → efficiency innovation" in economic terms.
  • Learning from carbon leakage, free + benchmarking allocation is key to reducing "AI leakage" (companies relocating to unregulated regions).

Highlights & Insights

  • Accurately "translates" the mature carbon cap-and-trade mechanism to AI—retaining the three core mechanisms of benchmarking, secondary market, and banking, and adding an assistance factor for clean energy in the AI context, making it highly actionable.
  • Publishing a governance + economics paper at a top ML conference is rare; providing a closed-form KKT solution gives policymakers direct "mathematical evidence," turning typical policy debates into quantifiable models.
  • The naming and analogy of "AI leakage" enables economists and policy researchers to immediately apply existing carbon leakage toolkits—this kind of conceptual "interface design" is valuable for interdisciplinary progress.
  • Monetizing efficiency itself—small companies' "efficiency surplus" becomes sellable cash, reversing the current trend of "compute gap → irreversible industry concentration." The approach could inspire applications in cloud compute/model hosting and other domains.

Limitations & Future Work

  • The model simplifies the performance-FLOP relationship as \(-x^{-k}\), lacking the multi-break-point structure of real LLM scaling laws; in reality, \(k\) also varies by task, possibly making conclusions unstable in some regimes.
  • Does not model "strategic games between companies" (e.g., incumbents hoarding quotas to suppress entrants); only single-company KKT is analyzed.
  • Price \(b\) is endogenously determined in the secondary market, but the paper treats it as an exogenous constant, not modeling supply-demand equilibrium; real emission markets often see sharp price fluctuations.
  • Regulatory costs (FLOP counting, third-party audits, cross-border quota recognition) are barely discussed; implementation may be even more challenging than carbon markets.
  • The paper does not address how to unify quota measurement for emerging modalities (e.g., video generation FLOPs are orders of magnitude higher than text).
  • vs Pigouvian tax / token tax (Hebous & Vernon-Lin, Korinek & Lockwood): These directly tax electricity or tokens; this paper uses cap-and-trade, as it has been proven in carbon markets to reduce emissions more effectively at lower cost and without causing geographic relocation (leakage).
  • vs User-side fee (e.g., UN UNEP 2025): User-side fees only incentivize users to reduce usage, not companies to optimize efficiency; cap-and-trade incentivizes both.
  • vs Insurance / Certification (Lior 2021, Ball 2025): Those approaches focus on misuse liability; this paper focuses on efficiency/sustainability.
  • vs AGI safety market (Tomašev 2025): They discuss agent-to-agent markets to mitigate AGI risk; this paper is about a company-level quota market, which is complementary.

Rating

  • Novelty: ⭐⭐⭐⭐ Complete transplantation of cap-and-trade to AI with KKT proof is a rare perspective in ML, though the mechanism itself is a mature borrowing.
  • Experimental Thoroughness: ⭐⭐⭐ Only toy-level numerical experiments (two curves in Fig 1), no calibration with real company energy data; lacks governance simulation and game-theoretic modeling.
  • Writing Quality: ⭐⭐⭐⭐ Clear structure, solid citations, and concise terminology analogies (AI leakage, AI Allowance).
  • Value: ⭐⭐⭐⭐ As AI governance becomes increasingly urgent, this provides a concrete proposal that legislators can immediately reference—such interdisciplinary manifesto papers often have more real-world impact than technical novelty.