AI Training Incentives
As AI systems increasingly adopt distributed and collaborative training, all participants that offer data and computation also contribute to the energy footprint of these processes. Each type of participants faces trade‑offs among its effort, performance, use of energy, and CO2 emissions. While technical optimisation of distributed AI services improves sustainability, greater benefits arise when appropriately designed incentive mechanisms motivate participants to adopt energy‑ and carbon‑aware behaviours, thus ensuring sustained engagement and more environmentally conscious operation. While the AI-Native Sustainability focuses primarily on incentive design and user behaviour, the proposed framework extends this perspective by explicitly situating these mechanisms within a resource-constrained system environment. The incentive mechanisms can be viewed as operating on top of a resource orchestration layer that determines feasible operating conditions, such as energy availability, timing constraints, and network capacity.
This work introduces an incentive‑driven framework that actively motivates participants to promote energy and carbon efficiency during collaborative AI training. The framework aligns individual decisions with system‑wide sustainability goals, thus ensuring that contributors remain engaged in the collective reduction of the environmental footprint of training. By means of lightweight coordination mechanisms, participants can adjust their behaviour – for example, by shifting computation to periods of higher renewable energy availability or by selecting more energy efficient training configurations – without compromising performance of the AI model. By rewarding energy‑aware actions, the framework gives rise to conditions in which self‑interested behaviour naturally supports energy efficiency and carbon reduction together with own performance.
To formalise these interactions, we model the training process as a finite normal‑form game with incomplete information, corresponding to a federated setting where each client observes only its own data. Within this structure, we investigate and evaluate two decision and coordination mechanisms. The first mechanism is a heuristic phase, where clients make decisions independently based on their own reasoning. This setup allows the server to observe and subsequently exploit these choices by adjusting penalties, to study how users behave when interacting autonomously, and to establish a benchmark against which the coordinated (correlated‑equilibrium) mechanism can be compared. The second mechanism provides signals (namely, suggestions for the scheduling the training activities) that guide clients towards a correlated equilibrium, thus enabling more coordinated and energy‑efficient behaviour. These mechanisms allow us to assess how different forms of decision‑making influence participation, energy use, and alignment with sustainability objectives.
Coordinated decision‑making leads to significant improvements of energy efficiency in collaborative training. When participants follow correlated signals, the system uses green energy more effectively – leaving a smaller surplus and adapting smoothly to fluctuations – whereas using only the heuristic phases leads to larger deviations due to fully decentralised choices. Overall, well‑aligned incentives and even light coordination led to considerable improvements of the training energy footprint without compromising learning quality or participation.
AI Inference Incentives
AI inference is currently the largest source of carbon emissions in deployed AI systems. It is responsible for up to 90% of system costs and for as much as 70% of operational emissions, depending on the energy mix. Its energy demand is escalating as larger models require more computation per query, while the emergence of Agentic AI generates many additional inference calls through autonomous, continuous task execution. Energy use per query varies widely – from 0.23 Wh for typical prompts to 33 Wh for long ones – thus justifying the need for mechanisms that would lead to the reduction of energy consumption inference, while preserving acceptable QoE.
To address this important challenge, we introduce a framework that models the user’s valuation of inference accuracy, latency, and environmental consciousness, and links these preferences to the carbon emissions associated with different inference configurations. This framework answers two key questions: 1) What is the optimal balance between accuracy reduction and latency increase when a specific type of incentives is offered? 2) What incentive should be provided to users in order to achieve a target carbon emissions reduction, and what QoE deterioration is associated with this target?
Based on our model, we take that the “green” user accepts 9% higher latency and 7% lower accuracy than the “QoS‑sensitive” user. Building on these insights, a practical two‑tier subscription model is proposed. The first tier offers optimum quality inference, while the discounted sustainable tier provides users with a price reduction in exchange for allowing the provider to serve a portion of inference requests at lower accuracy and higher latency during periods of high carbon intensity. This structure offers AI providers the capability to flexibly reduce emissions while still accommodating diverse user preferences and maintaining service availability and user satisfaction.
References
- V. Siris, A. Stamou, G. D. Stamoulis, K.Varsos, and R. Khalili, “Greening AI Inference with Accuracy and Latency-aware User Incentives,” Submitted for publication, Dec. 2025.
- Wiesner et al,FedZero:Leveraging Renewable Excess Energy in Federated Learning, e- Energy’24, 2024.
Authors
Athens University of Economics and Business–Research Center (AUEB)
Adamantia Stamou (Dr. Eng., MBA) is a senior researcher at STEcon research group, AUEB. Her duties include research and project management in European Research Projects related to 6G networks and Blockchain, performing state-of-the-art multidisciplinary research in the intersection of economics, digital information and technology. Adamantia combines the fields of Informatics & Telecommunications Engineering with Strategic Management & Coaching Leadership to provide exceptional Research & Development results and Innovative solutions in both Academia and Business Sectors.
Athens University of Economics and Business–Research Center (AUEB)
Vasilios Siris is a Professor at the Department of Informatics of the Athens University of Economics and Business. His research interests include wireless and mobile communication systems, blockchain technology, and digital identities.
Athens University of Economics and Business–Research Center (AUEB)
George D. Stamoulis is a Professor in the Department of Informatics in AUEB, and Head of the STEcon research group. He received the Diploma in Electrical Engineering (1987, with highest honors) from the National Technical University of Athens, Greece, and the MS (1988) and PhD (1991) degrees in Electrical Engineering from the Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. His research interests are in energy efficient 6G communications, economic and business models for networks, clouds, smart energy grids and blockchain environments, demand response and flexibility management in smart energy grids, auction mechanisms for scarce goods, and reputation mechanisms for electronic environments.
Athens University of Economics and Business–Research Center (AUEB)
Konstantinos Varsos is a postdoctoral researcher at STEcon research group, Athens University of Economics and Business, Department of Informatics, and Adjunct Lecturer at the University of Crete. His research focuses on game theory, computational mathematics and learning.