Amparo Alonso-Betanzos

Original version in English

Artificial Intelligence (AI) has been evolving at an extraordinary pace. In just a few years, systems that once seemed distant ambitions, and I say this as someone who has been working in the field since the 1980s, have not only become real but have woven themselves into the fabric of everyday life, shaping how we communicate, work, and access information. As a researcher, I find this progress both exciting and unsettling. The capabilities of modern AI are genuinely remarkable, but they also force us to confront a question that goes well beyond the technical: what kind of AI do we want to build, and what role should it play in our societies?

One concern that deserves more public attention is the growing computational footprint of these systems. Training a large AI model requires substantial amounts of energy, making efficiency in both software and hardware a critical factor. ¹ AI Index Report 2025, carbon emissions from AI training have been rising sharply with each new generation of models: GPT-3 emitted 588 tonnes of CO₂; GPT-4, released three years later, produced 5,184 tonnes; and Meta’s Llama 3.1 405B, in 2024, reached 8,930 tonnes. For context: the average person generates around 5.5 tonnes of CO₂ per year. Training a single frontier model can therefore represent the lifetime emissions of hundreds of individuals. And beyond carbon, the water footprint of these systems is equally striking: training GPT-3 alone required an estimated 700,000 litres of water for datacenter cooling.

However, focusing only on the training phase is a common mistake. Every time a user interacts with the system (what is called inference), the model also burns energy, and that cost often ends up being the largest share of a system’s total emissions over its lifetime. To give you a sense of scale: a single query to a large language model can require up to ten times more electricity than a standard Google search. The hit on water reserves is no smaller: studies suggest that a short exchange of 20 to 50 prompts effectively uses half a litre of water just for cooling. Multiplied across millions of daily users, what looks like a quiet background task turns into an enormous, relentless strain on global resources. On top of that, data ages fast. Models need to be retrained repeatedly to stay current, and until learning architectures shift toward more efficient approaches, we’ll remain stuck in that dynamic: consuming ever more resources just to keep our digital tools from going stale.
The underlying driver is scale. The volume of data used to train AI systems has been roughly doubling every eight months, while the computational resources required for training double approximately every five months. Larger datasets demand longer training runs, and longer training runs mean larger energy bills and a larger environmental toll. There is, however, at least one notable counterexample. DeepSeek’s V3 model, released by the Chinese lab in late 2024, reported CO₂ emissions of around 597 tonnes, comparable to GPT-3 from 2020, despite being a far more capable system. This suggests that efficiency gains are possible, and that the current trajectory is not entirely inevitable. It also raises important questions about how we measure and compare the environmental cost of AI across different development approaches. The consequences of this trajectory, however, extend beyond the ecological. The resources required to train and deploy state-of-the-art models are not within the reach of all research institutions, universities and organizations, especially in the Global South. This is not merely a sustainability challenge. It is equally a challenge of equitable access. When the capacity to develop and shape AI is concentrated in a handful of well-resourced actors, the technology’s benefits (and the values embedded within it), reflect a narrow set of perspectives.
That concentration of power also has a sovereignty dimension that deserves to be named directly. When the systems that mediate how we access knowledge, generate text, or interact with public services are built and controlled elsewhere, we are not merely consumers of a technology. We are, to a significant degree, dependent on choices made by others about what these systems know, what they prioritise, and whose languages and cultures they serve. This is not an abstract concern. Large language models trained predominantly on English-language data perform systematically worse in other languages, and not just in terms of fluency. They reflect cultural assumptions, historical framings, and implicit value judgements that are far from universal. For languages spoken by smaller communities (in Europe languages such as Galician, Basque, Welsh or Maltese, among many others) the risk is not only one of poorer performance, but of gradual marginalisation: a world in which the most capable AI systems simply do not speak your language, or do not speak it well enough to be genuinely useful and reflect your values and culture.
Strategic autonomy in AI therefore requires more than regulation. It requires infrastructure. Investing in supercomputing capacity and datacenter ecosystems is not a matter of technological pride; it is a prerequisite for ensuring that foundational models can be trained on our soil, under our standards, with our data, and with linguistic and cultural diversity genuinely represented. Without that capacity, sovereignty remains a declaration of intent rather than a practical reality.
This is where the frugal AI agenda becomes strategic, not only environmental. Rather than assuming that progress requires ever- larger and more computationally expensive models, our research group has been exploring what is called frugal or sustainable AI: approaches that aim to achieve strong performance with more modest resources, less data and lower energy consumption, both during training and at inference time. Efficiency, in this sense, is not merely a resource optimisation problem; it is a design philosophy, and we would argue, also an ethical one. This philosophy becomes particularly relevant in contexts where computational constraints are not a choice, but a given. Consider the Internet of Things (IoT), where models must run on small edge devices with limited processing power and battery life; or robotics, where agents must learn to perform tasks with limited time and few training examples; or scenarios defined by data scarcity or privacy requirements, where access to large, curated datasets simply is not possible. In all these settings, the dominant paradigm — scale up, collect more data, train longer — is not a viable answer.
Hardware efficiency is improving (roughly 40% annually, according to current estimates), but this alone will not be sufficient to offset the exponential growth in computational demand. What is needed is a broader shift in research culture, one that values doing more with less, not as a constraint to be overcome, but as a virtue to be cultivated. Frugal AI makes it more feasible to train competitive models without requiring the infrastructure of a hyperscaler. It opens a realistic path to building capable, culturally grounded systems for lower-resource languages, where the available training data is inherently limited. And it reduces the energy and water demands of the entire development pipeline, aligning technological ambition with environmental responsibility.
One of the most promising directions in sustainable AI research is rethinking not just what models learn, but how they learn. Classical machine learning assumes access to large, labelled datasets and substantial computational budgets. But human intelligence does not work this way — we learn from a few examples, generalise rapidly, and adapt to new situations without restarting from scratch. Replicating these properties in artificial systems is both a scientific challenge and a practical necessity.
Several research directions are converging on this goal, and they form the core of our group’s current work. A thread common to all of them is that we do not treat sustainability as a separate objective: reducing CO₂-equivalent emissions is built into how we design and evaluate our methods, alongside predictive performance. In continual learning, we explore how systems can acquire new knowledge over time without forgetting what they have already learned — something essential in dynamic, real-world environments, and a property that also reduces the need for costly full-model retraining each time circumstances change. In the area of recommendations, our work on more efficient algorithms has achieved reductions in CO₂-equivalent emissions of around 80% without sacrificing performance.
We are also addressing scenarios where labelled data is scarce — a far more common situation in practice than the clean, fully annotated datasets that benchmarks tend to assume. Using techniques such as PU Learning, which exploits a small set of positive training examples alongside a large pool of unlabelled ones, we have been able to improve the performance of models identifying genes associated with dietary restriction and longevity. This is something we encounter repeatedly: constraints do not weaken science. Often, they sharpen it. We are now extending this direction into few-shot and zero-shot learning, where the goal is to build useful models from very limited labelled data, opening the door to applications in domains where annotation is expensive, expertise is rare, or privacy makes large-scale data collection impossible.
Frugality in AI is not only a matter of computation. Data itself carries a cost that is too often overlooked in energy, infrastructure, human labour, and, increasingly, in questions of privacy, consent, and cultural representation. Who decides what data is collected, in which languages, reflecting whose experiences? These are not purely technical questions. Research into data-efficient methods, such as active learning, self-supervised learning, or synthetic data generation, seeks to reduce dependence on massive curated datasets without sacrificing model quality, and to do so within the kind of privacy frameworks that regulations such as the European AI Act rightly demand. The
ability to learn from less data, and from data that never leaves its source, is in this sense both a technical advance and an expression of values.
Underlying all of this is a cultural question: how do we measure progress in AI systems? For much of the past decade, the dominant benchmark has been raw performance: accuracy on standard datasets, scores on reasoning benchmarks, fluency in language generation. These metrics have driven remarkable advances, but they have also created perverse incentives, rewarding scale above all else and obscuring the true cost (environmental, economic, and geopolitical) of the systems being built. What we are not measuring matters, and new metrics are needed. Without them, neither better tools nor meaningful regulations will follow. A more complete picture of AI progress would account for performance per unit of energy consumed, per litre of water used, per tonne of CO₂ emitted. It would ask not only whether a model can solve a problem, but whether it can do so across different resource environments and in the full diversity of human languages and cultures, treating linguistic breadth not as an afterthought but as a core criterion of quality.
Europe has both the regulatory tradition and, increasingly, the infrastructure to lead on this redefinition. The question is whether we choose to. Frugal AI, efficient, culturally grounded, and built on sovereign infrastructure, is not a consolation prize for those who cannot compete at the frontier. It is, we would argue, a more honest and more ambitious vision of what AI should become.