EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Payment-Coupled Carbon Accounting for LLM Inference: Governing Emissions Reporting Through Settlement Incentives

Carlo Segat, Thomas Cory

International Conference on AI Research · 18.11.2026 · 7 (1) · 369-377

DOI: 10.34190/icair.7.1.5564 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Conference on AI Research, online erschienen am 18.11.2026. Verfasst von Carlo Segat und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Over a deployed model’s lifetime, the energy and CO2-equivalent cost of large language model (LLM) inference now rivals the cost of training it. Yet reporting that cost is voluntary: inference providers self-disclose, most publish no per-inference figures at all, and each is commercially incentivised to understate the volume it serves. Existing remedies treat measurement and trust as separate problems. Per-token carbon estimators quantify emissions, while Distributed Ledger Technology (DLT) carbon markets tokenise verified offsets. Neither ever observes a served request, blindly trusting the provider’s token counts and recording only the equivalent CO2, not the delivered service. Declaring fewer requests is therefore free. This paper’s research question is whether carbon accounting can be made resistant to underreporting. To this end, the paper puts forward a design that cryptographically binds the accounted and paid party. A crucial element of this design is a settlement mechanism that forces paying for an inference to also record its emissions. The design is instantiated over the x402 micropayment standard on an EVM Ledger and evaluated against a baseline without carbon logic, measuring the marginal on-chain cost and the added end-to-end latency the coupling introduces. The coupling roughly doubles settlement gas while adding tolerable latency. The contribution is an economic anti-underreporting guarantee: the ledger acts as a tamper-evident record of served volume, from which a verifiable lower bound on a provider’s emissions follows. The inherent limitation is that converting served volume into carbon still relies on declared inputs (e.g. region, model and grid intensity). The client can inspect those inputs before signing, though the prototype leaves such screening to future work. This paper is a first step towards joining sustainable-AI measurement and on-chain settlement, 2 fields that have so far developed in isolation. It reframes emissions reporting as an incentive problem, not a measurement problem. The effect is not only conceptual: users of LLM inference under carbon-disclosure duties can rely on a trustworthy record of their activities.

Zitieren

APA

Segat, C., & Cory, T. (2026). Payment-Coupled Carbon Accounting for LLM Inference: Governing Emissions Reporting Through Settlement Incentives. International Conference on AI Research, 7(1), 369-377. https://doi.org/10.34190/icair.7.1.5564

Vancouver

Segat C, Cory T. Payment-Coupled Carbon Accounting for LLM Inference: Governing Emissions Reporting Through Settlement Incentives. International Conference on AI Research. 2026;7(1):369-377. doi:10.34190/icair.7.1.5564

BibTeX
@article{Segat2026,
  author = {Segat, Carlo and Cory, Thomas},
  title = {Payment-Coupled Carbon Accounting for LLM Inference: Governing Emissions Reporting Through Settlement Incentives},
  journal = {International Conference on AI Research},
  year = {2026},
  volume = {7},
  number = {1},
  pages = {369--377},
  doi = {10.34190/icair.7.1.5564}
}
Kostenlos starten →