AI and Electricity: A System Dynamics Approach - Explained (7/10) - "Sustainable AI" Scenario

Welcome to the seventh series on "AI and Electricity Scenarios: A System Dynamics Approach".

Today, we'll explore the results of the "Sustainable AI" scenario, a perspective developed by the School of Thought of Sustainable AI Advocates, which envisions a future where technological advancement and environmental responsibility coexist harmoniously.

This vision is driven by the belief that AI-driven efficiency advancements, coupled with frugality and resource-conscious use, can lead to substantial improvements in data center operations. Furthermore, the integration of AI with the energy system, guided by a long-term vision, can foster a symbiotic relationship that supports broader economic development and sustainability goals.

The Genesis: AI with a human face

This modern concept of Sustainable AI finds resonance with ideas proposed decades earlier by British economist E.F. Schumacher in his influential 1973 book, Small is Beautiful.

Schumacher's work, often overlooked in contemporary debates on machine learning and AI, offers insights that are increasingly relevant for developing a humanist approach to technology in the age of computation. His emphasis on Technology with a human face provides a thoughtful framework to think the challenges posed by AI development and offers a powerful way to consider alternatives to the gigantism of current tech industry ideologies.

Central to Schumacher's analysis is the idea that modern technology serves to 'degrade' both environmental and social structures. This concern extends to the impact on human cognition and creativity, as he was particularly worried about the negative effects of mechanization on human sensibility and the life of the mind – a concern that resonates strongly with current debates about AI's trajectory.

Blueprints for Balance: Architecting Sustainable AI

This scenario is characterized by an efficiency balancing mechanism, where data center electricity consumption is effectively equilibrated.

At the heart of this vision is an efficiency balancing mechanism that ensures energy consumption remains manageable. This mechanism relies on:

• Systemic improvements in hardware and software efficiency
• Frugal design principles that guide AI model development and operation
• A symbiotic relationship between AI-driven demand and the energy system

General Analysis of Energy Consumption (2025-2030)

From 2025 to 2030, the Sustainable AI scenario reveals a steady increase in energy consumption, reflecting the expanding demand for AI applications and the enhancement of AI capabilities.

Annual generative AI training energy consumption rises from 47 TWh in 2025 to 228 TWh by 2030, a more than threefold increase which highlights the growing complexity and scale of generative models.

Generative AI inferencing grows stronger starting in 2027 as deployment increases, from about 15 TWh to 310 TWh within the same period. This substantial increase underscores the growing importance of AI technologies and highlights the need for efficient energy management strategies to support this expansion sustainably.

Large language models (LLMs) and industrial generative AI training also show significant growth. Consumer LLMs training increases from 40 TWh in 2025 to 167 TWh by 2030, reflecting their expanding applications in sectors such as customer service and content generation. Industrial generative AI training rises from 7 TWh to 61 TWh, indicating the rapid integration of generative AI into industrial processes for enhanced automation and operational efficiency.

Traditional AI electricity use maintains a relatively stable energy consumption pattern until 2028, then observes a slight decrease as it becomes more industrialized and as open-source and low-code model communities(133) gain momentum. This trend suggests a shift towards more efficient traditional AI models. Inferencing and training for traditional AI evolve from 18 TWh to 20 TWh in 2025 to 16 TWh and 29 TWh by 2030, respectively, showing modest growth in training requirements while inferencing remains stable.

Generative AI inferencing is emerging as the dominant electricity consumer

Scenario projections indicate that Gen AI inference will become the primary driver of electricity consumption within the AI sector by 2027-2028. This trend could lead to electricity consumption exceeding 200 TWh within three years, highlighting the intensifying computational requirements of generative AI models.

Despite this, the Sustainable AI scenario is evolving positively, largely due to three factors.

First, the 2023 seminal work by Luccioni et al., “Power Hungry Processing: Watts Driving the Cost of AI Deployment?”, offers stronger evidence for both public perception and industry practices. Their comprehensive study of 88 diverse AI models revealed that generative AI models can consume significantly more energy for inference tasks than conventional algorithms, with some models requiring up to 30 times more energy than traditional search engines for real-time processing. A key insight from this research - that using multi-purpose models for discriminative tasks can be more energy-intensive than employing task-specific models - is being seriously incorporated into industry sustainable AI roadmaps. This awareness is prompting a positive reevaluation of AI model deployment strategies, emphasizing the need for more energy-efficient approaches in AI application design.

Second, significant hardware efficiency breakthroughs are materializing, offering substantial benefits to companies embracing Sustainable AI school of thought. Advancements in AI hardware and cooling technologies are driving significant improvements in performance and energy efficiency. The NVIDIA GB200 NVL72 system, utilizing multiple GB200 superchips, exemplifies this progress with substantial performance gains over previous-generation systems based on H100 GPUs. System-level comparisons demonstrate up to a 30-times performance increase for LLM inference workloads, attributed to the new Blackwell architecture, improved cooling solutions, and system-level optimizations. At the chip level, the B200 GPU offers a 127% improvement over the H100, delivering 2,250 TFLOPS of FP16/ BF16 compute compared to the H100’s 989 TFLOPS. These advancements are complemented by server rack densification and widespread adoption of liquid cooling systems, supporting up to 132kW rack power density. As awareness of AI’s energy implications grows, these combined innovations are facilitating substantial chip evolutions and driving significant changes in datacenter design and efficiency. Hence, in the Sustainable AI scenario, while direct chip-to-chip comparisons may initially suggest performance gains, companies are prioritizing system-level optimizations through energy-efficient technologies to achieve a competitive advantage, maintain sustainability, and mitigate potential performance and cost-related challenges in the AI landscape.

Third, the Sustainable AI scenario is characterized by a symbiotic relationship between AI infrastructure and demand, where efficiency and resource conservation are mutually reinforced. This synergy extends from user-centric applications to broader systemic impacts. Research from the 2020s is now translating into real operational gains. While AI inference may initially increase direct energy consumption, its potential to optimize energy usage across multiple sectors can lead to net positive energy outcomes. Applications in HVAC systems, microgrids, electric vehicle-to-grid integration, and waste management demonstrate this potential. This symbiosis provides real-life effects, emphasizing the importance of considering AI’s indirect effects on energy consumption alongside its direct impacts.

Traditional AI will continue to play a crucial role in decarbonization efforts across various end-use applications

As a key influencer of the efficiency balancing mechanism in the model, Traditional AI plays a crucial role in the Sustainable AI scenario. It is confirmed as an effective tool for decarbonization efforts across various sectors. Its steady growth demonstrates that machine learning models such as decision trees, random forests, and support vector machines are more energy-efficient than generative models. This provides a crucial counterbalance to the rising energy demands of generative AI. In this scenario, Traditional AI training energy consumption is projected to increase modestly from 38 TWh in 2025 to 45 TWh by 2030. These models are currently deployed at scale across numerous industries, actively contributing to decarbonization efforts in ways that generative AI has yet to match in impact and widespread adoption.

In this scenario, Traditional AI is widely used in energy grid optimization, smart building management, industrial process optimization, transportation and logistics, and precision agriculture, among other areas. These applications have been refined and optimized over years of deployment, making them highly effective in reducing carbon emissions across various sectors of the economy. As highlighted by Rolnick et al.’s comprehensive study, AI, particularly traditional models, holds significant potential for addressing climate change challenges across various domains.

While the growth of generative AI inferencing is inevitable (projected to exceed 200 TWh by 2027-2028) and may offer new possibilities for decarbonization, traditional AI remains the primary tool for immediate and large-scale impact due to its efficiency, widespread adoption, and proven effectiveness. However, ongoing optimization and innovation in traditional AI models are crucial to keep pace with evolving demands, with the development of Small Language Models (SLM) and Tiny AI representing promising directions that combine the strengths of traditional AI and machine learning.

Resource-conscious generative AI training approaches intensify their focus on less energy-intensive models

Gen AI training strikes a delicate balance between technological and infrastructure progress and frugality. This balance is achieved through advancements in three key areas: infrastructure improvements, hardware and software innovations, and frugality strategies.

On the infrastructure side, multi-datacenter training may revolutionize Gen AI infrastructure, making it systemically more efficient and environmentally friendly. This approach to sustainable AI development encompasses several key aspects, integrating hardware, software, and infrastructure optimizations. It leverages distributed computational load and network optimization through load balancing and time zone optimization, while utilizing emerging memory technologies like STT-RAM and Memristors to enhance server performance and energy efficiency.

Training strategies prioritize renewable energy utilization and adapt to seasonal variations, aligning with efforts to improve data center PUE. Advanced cooling technologies, exemplified by NVIDIA’s Blackwell GB200 family, complement these efforts. Hardware efficiency is boosted through flexible scaling across data centers, supported by Moore’s Law and projections of increasing transistor counts, with NVIDIA targeting over 200 billion transistors by 2024 and TSMC aiming for 1nm processes with over a trillion transistors by 2030. Advanced network technologies optimize communication, reducing latency for real-time processing. On the software front, distributed learning algorithms, hyperparameter tuning, transfer learning, and AutoML are employed to reduce model sizes and improve resource utilization.

This system approach combines hardware advancements, energy-efficient infrastructure, and software optimizations to create a more sustainable AI ecosystem.

Moreover, in this scenario, initiatives such as Luccioni’s work using CodeCarbon to measure carbon footprints(171, 172) and Schwartz et al.’s proposal to emphasize computational cost reporting and prioritize efficient hardware and algorithms(174) are integrated. Frugality strategies, drawing on the first specification standard about frugal AI (AFNOR), offer actionable levers to reduce AI’s impact.

Thank you for engaging with the "Sustainable" scenario!

In our next exploration, we will delve into the outcomes from this scenario and share our recommendations towards Sustainable AI.

Looking forward to sharing this soon.