Language models process human questions by converting words into "tokens," which are smaller units—sometimes parts of words—encoded numerically for machine comprehension. These tokens form the fundamental operational currency driving LLMs’ computation. However, every operation to generate or process tokens consumes energy, and given the scale of modern AI applications, this translates into meaningful quantities of carbon dioxide emissions. Despite increased awareness of AI’s impressive capabilities, the ecological footprint of simply interacting with these models remains poorly understood by most users.

The German research team examined fourteen large language models ranging from seven billion to over seventy billion parameters — the parameters being critical factors that define how a model learns and processes information. Their approach involved asking each model one thousand benchmark questions spanning diverse academic and practical subjects, allowing an apples-to-apples comparison of energy consumption normalized against performance and reasoning complexity. Crucially, the study differentiates between models that generate concise answers and those designed to engage in more elaborate, step-by-step reasoning processes.

The results demonstrate that reasoning-enabled LLMs, which produce extensive intermediate "thinking" tokens before delivering final answers, can be up to fifty times more carbon-intensive than their concise-answer counterparts. On average, reasoning models generated approximately 543 ‘thinking’ tokens per question, whereas concise models relied on just 38 tokens to provide their responses. These additional tokens accumulate computational demand and thus escalate CO₂ emissions. Yet, this token density does not assuredly yield greater accuracy; rather, it often introduces verbose details, some of which are extraneous to the correctness of the answer.

Among the models tested, the Cogito model—a reasoning-enabled LLM with seventy billion parameters—emerged as the most accurate, attaining nearly 85% correctness over the thousand questions. However, this high performance came at a significant environmental cost: the Cogito model emitted three times as much CO₂ as similarly sized LLMs generating succinct answers. This finding highlights a critical accuracy-sustainability trade-off in current AI systems: models that maintain carbon emissions below approximately 500 grams of CO₂ equivalent struggle to surpass 80% accuracy on these benchmarks.

Another notable insight is the subject-dependent variation in emissions. Queries involving complex reasoning—such as abstract algebra problems or philosophical dilemmas—induced up to six times more CO₂ emissions than straightforward topics like high school history. This discrepancy is linked to the increased number of reasoning tokens required by the model when addressing conceptually demanding questions, further exacerbating energy expenditure and carbon footprint.

These findings underscore the broader implications for AI deployment and responsible usage. By selectively employing models optimized for concise responses when exhaustive reasoning is unnecessary, users can dramatically reduce their environmental impact without sacrificing meaningful accuracy. Researchers emphasize that awareness is paramount; knowing the carbon cost tied to specific AI tasks enables individuals and organizations to make more informed decisions regarding model choice and usage frequency.

The study also highlights how hardware differences and regional energy grid variations may influence these emission estimates. The carbon intensity of powering AI systems depends on local energy sources, data center efficiencies, and the underlying infrastructure, all of which introduce variability into sustainability assessments. Consequently, while the reported metrics provide a compelling benchmark, they should be considered within the context of these fluctuating parameters.

A striking calculation presented by the researchers compares the CO₂ emissions of question-answering at scale to familiar everyday activities. For instance, having the DeepSeek R1 model (with seventy billion parameters) answer 600,000 questions produces carbon emissions comparable to a round-trip transatlantic flight from London to New York. Contrastingly, the Qwen 2.5 model, slightly larger in size but more efficient, can deliver nearly twice as many answers at equivalent accuracy while generating the same carbon footprint. Such comparisons contextualize AI’s environmental costs alongside other human activities, making the impact more tangible.

Ultimately, this research calls for greater transparency in AI’s environmental footprint and suggests that integrating emission metrics into user interfaces could foster more sustainable AI consumption habits. By informing users of the carbon cost of specific interactions—whether that be generating a lengthy philosophical essay or transforming a casual photo into a stylized action figure—platforms can encourage prudent and environmentally mindful usage. This aligns with broader efforts across technology sectors to quantify and mitigate the ecological ramifications of burgeoning digital tools.

The energy expenditure linked to AI communication is an emerging but pressing consideration in the debate surrounding ethical and sustainable artificial intelligence. As AI becomes increasingly ubiquitous in research, education, industry, and entertainment, understanding and managing its energy demands is paramount to balancing innovation with planetary stewardship. This study provides a crucial empirical foundation and invites future work to further refine these assessments and develop greener AI architectures.

The environmental trade-offs highlighted by this analysis remind us that technological progress is inseparable from ecological responsibility. The choices developers and users make today, from model design to everyday prompts, collectively shape the carbon trajectory of AI’s future. Thoughtful stewardship, powered by rigorous data like that provided by this research, is essential to ensure that AI advancements serve humanity without compromising our planet’s health.

Subject of Research: Not applicable

Article Title: Energy Costs of Communicating with AI

News Publication Date: 19-Jun-2025

Web References: http://dx.doi.org/10.3389/fcomm.2025.1572947

References: Dauner, M., et al. (2025). Energy Costs of Communicating with AI. Frontiers in Communication. https://doi.org/10.3389/fcomm.2025.1572947

Image Credits: Not specified

Keywords

Large Language Models, AI Energy Consumption, Carbon Footprint, Artificial Intelligence, Environmental Impact, Reasoning Models, Tokenization, AI Accuracy, Sustainability, Machine Learning, CO₂ Emissions, Green AI

Titled “Neuromorphic Computing at Scale,” this review meticulously examines the current state of neuromorphic technology and proposes a robust strategy for the development of large-scale neuromorphic systems. Neuromorphic computing, which seeks to mimic the architecture and functionality of the human brain, has gained immense traction as a transformative approach in computing, applying insights drawn from neuroscience. The findings within this article are poised to reshape our understanding and approach to computational processes, particularly in fields where computational efficiency and energy consumption are of utmost concern.

At the heart of the research lies the imperative to enhance the scalability of neuromorphic systems. As the electricity consumption associated with artificial intelligence technologies escalates, projected to double by 2026, the need for energy-efficient computing solutions has never been more urgent. Neuromorphic chips are conceived to outstrip traditional computing frameworks, not only in terms of energy consumption and physical space optimization but also in overall performance across a myriad of domains, including artificial intelligence, healthcare, and robotics.

Kudithipudi emphasizes that neuromorphic computing is reaching a “critical juncture,” with scalability acting as a litmus test for the progress and viability of the field. Notable advancements have already been observed, with Intel’s Hala Point demonstrating the integration of an astounding 1.15 billion neurons into its neuromorphic architecture. However, the research findings suggest that there is still significant growth necessary in this sector to tackle intricate, real-world computational challenges effectively.

The insights presented by the authors resonate with the sentiment that neuromorphic computing is currently experiencing a pivotal moment akin to previous watershed moments in the development of technologies, such as the advent of AlexNet in deep learning. This period presents a remarkable opportunity to design new architectures and frameworks that can find applications in commercial settings. Central to this endeavor is the need for collaborative efforts bridging academia and industry—an aspect echoed throughout the collaborative nature of the research team comprised of varying institutions and corporate partners.

Kudithipudi is no stranger to the domain of neuromorphic computing. Her extensive contributions include securing a substantial $4 million grant from the National Science Foundation last year aimed at launching THOR: The Neuromorphic Commons. This groundbreaking initiative seeks to establish a collaborative research network that provides open access to neuromorphic computing hardware and tools, fostering interdisciplinary partnerships and innovation.

In addition to scaling up access to neuromorphic resources, the authors advocate for developing a diverse range of user-friendly programming languages. Such a shift would lower barriers to entry, fostering a richer collaborative environment across various disciplines and industries. The aim is to cultivate a community capable of addressing complex problems by leveraging the strengths of neuromorphic computing.

Among the co-authors is Steve Furber, an emeritus professor at the University of Manchester, who has an illustrious history in neural systems engineering. Furber highlights the significance of this research paper, noting that it captures the current landscape of neuromorphic technology at a moment when it is poised for expansive commercial applications, moving beyond mere brain modeling into broader AI applications capable of managing large-scale, energy-intensive AI models.

The research aims to identify key features that must be honed to achieve the desired scale in neuromorphic computing. Notably, the concept of sparsity, a characteristic inherent to biological brains, surfaces as a focal point. Biological brains develop by forming extensive neural connections before selectively pruning those that are redundant or less effective. This strategy not only conserves space but optimizes information retention, yielding a model for neuromorphic systems to emulate. If replicated successfully, such a feature could significantly enhance the energy efficiency and compactness of these systems.

The collaboration resulting in this research paper represents a noteworthy convergence of various key research groups, uniting to share critical insights regarding the current and future states of the neuromorphic computing field. The authors express optimism that this concerted effort will pave the way towards making large-scale neuromorphic systems more mainstream, amplifying the discourse surrounding their potential benefits.

Tej Pandit, a doctoral candidate at UTSA and a co-author on the project, focuses his research on training AI systems to learn continuously without compromising prior knowledge. His recent publications contribute significantly to the evolving narrative of neuromorphic systems and their potential implementations. The research project exemplifies UTSA’s commitment to fostering knowledge within this transformative field, believed to be a catalyst for addressing pressing challenges concerning energy waste and the trustworthiness of AI outputs.

The widespread collaboration involved in this article extends beyond academic institutions, encompassing partnerships with national laboratories and industrial stakeholders. Collaborators include the University of Tennessee, Knoxville, Sandia National Laboratories, Rochester Institute of Technology, Intel Labs, and Google DeepMind, among others. This extensive network of partnerships embodies the interdisciplinary approach essential for driving the future of neuromorphic computing.

In a world increasingly dependent on advanced technologies, the implications of neuromorphic computing transcend mere computational efficiency. As researchers strive to create systems that mimic the intricate workings of the human brain, the potential for breakthroughs in energy consumption, AI dependability, and healthcare solutions is vast. With each step forward, the dialogue surrounding neuromorphic computing broadens, inviting researchers, industry leaders, and policymakers to engage in a shared vision of a more efficient and sustainable technological future.

As we move forward, the epochal research published today stands as a beacon for what the future may hold—not just for the field of computing but for our interactions with technology at large. The merging of academia and industry, coupled with a renewed focus on collaboration and innovation, holds the promise of transformative advancements that could redefine our understanding of intelligence, both artificial and human, in the years to come.

Subject of Research: Neuromorphic Computing
Article Title: Neuromorphic Computing at Scale
News Publication Date: 22-Jan-2025
Web References: Nature, MATRIX: The UTSA AI Consortium for Human Well-Being, THOR: The Neuromorphic Commons
References: None provided
Image Credits: The University of Texas at San Antonio

Keywords: Neuromorphic Computing, AI, Scalability, Energy Efficiency, Interdisciplinary Collaboration, Neuroscience, Artificial Intelligence, Computational Innovation