AI and Supercomputing are Powering the Next Wave of Breakthrough Science – But at What Cost?
Stefano Bianchini, Aldo Geuna, Fazliddin Shermatov, [1em] BETA, Université de Strasbourg, France · 2025
Read AI and Supercomputing are Powering the Next Wave of Breakthrough Science – But at What Cost? by Stefano Bianchini, Aldo Geuna, Fazliddin Shermatov, [1em] BETA, Université de Strasbourg, France (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-12686.txt.
Introduction
Stefano Bianchini, Aldo Geuna, and Fazliddin Shermatov measure what happens when artificial intelligence meets supercomputing in science. Both tools are spreading fast, but their joint effect on discovery, and who gets access to them, was poorly documented. The authors draw on metadata from more than five million publications across twenty-seven fields, published between 2000 and 2024. Track where AI and high-performance computing were used together, separately, or not at all in the work behind each paper.
They find the combination is strongest at the frontier: work using both is more likely to introduce novel ideas and reach top-cited status than work using either alone. Access, though, is concentrating in a few regions, led by the United States and China, with the European Union still competitive in combined output. The authors argue that future discovery depends on policy that widens access, not only on better algorithms and machines.
Catalog record
| CREATOR | Stefano Bianchini, Aldo Geuna, Fazliddin Shermatov, [1em] BETA, Université de Strasbourg, France |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 79 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-12686
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AI and Supercomputing are Powering the Next Wave of Breakthrough Science – But at What Cost? Stefano Bianchini, Aldo Geuna, Fazliddin Shermatov, [1em] BETA, Université de Strasbourg, France Abstract Artificial intelligence (AI) and high-performance computing (HPC) are rapidly becoming the engines of modern science. However, their joint effect on discovery has yet to be quantified at scale. Drawing on metadata from over five million scientific publications (2000–2024), we identify how AI and HPC interact to shape research outcomes across 27 fields. Papers combining the two technologies are up to three times more likely to introduce novel concepts and five times more likely to reach top-cited status than conventional work. This convergence of AI and HPC is redefining the frontier of scientific creativity but also deepening global inequalities in access to computational power and expertise. Our findings suggest that the future of discovery will depend not only on algorithms and compute, but also on how equitably the world shares these transformative tools. Science has long advanced through the interplay of theory, observation, and computation. Yet, in recent decades, signs of fatigue have begun to set in: despite the rapid growth in publications and data, progress seems to be slowing down. Ideas are becoming harder to find [ 1 ] , research outputs are becoming less disruptive [ 2 ] , and productivity in research and innovation has declined across most fields [ 3 ] . This deceleration has revived an age-old concern – that science may be about to enter a phase of diminishing returns, the most accessible problems having already been solved [ 4 ] , while the complexity of knowledge continues to rise [ 5 ] . Yet just as science is readying itself to face these diminishing returns, a new class of computational instrument has begun to transform research practices. In recent years, artificial intelligence (AI) and high-performance computing (HPC) have underpinned many of the most remarkable advances made. AI models trained on vast scientific datasets are capable of identifying regularities and generating hypotheses that would have otherwise escaped human reasoning, while HPC systems enable model training, testing, and large-scale simulations to be carried out. From the prediction of protein structures [ 6 ] to the discovery of new materials [ 7 ] and the control of fusion plasma [ 8 ] , the combination of AI and supercomputing is redefining what science can achieve. This convergence has not happened overnight. Earlier waves of information and communication technologies (ICTs) had already accelerated data collection and analysis, but they did so without fundamentally altering the logic of scientific inference [ 9 ] . AI and HPC, by contrast, have shifted science from automation to autonomy: from machines that assist computation to systems that learn, generalize, and explore enormous combinatorial spaces [ 10 ] . In the process, they have blurred the boundaries between theory, modeling, and experimentation, creating new forms of “computational empiricism” in which discovery emerges as much from pattern recognition as from deductive reasoning. AI continues to redefine performance boundaries and, in its own quiet way, to reshape the way science gets done. No longer simply a tool for data crunching, it now permeates virtually every step in the scientific pipeline – from hypothesis generation to experimental design to publication [ 11 ] . For some, deep learning represents much more than mere progress in computation; it is a new engine of invention, redefining how science operates and where its current frontier stands [ 12 ] . In 2024, the Nobel Prizes in Chemistry and Physics recognized breakthroughs made possible by AI, with DeepMind capturing the prevailing sentiment with its forecast that we may be entering “ a new golden age of discovery ” [ 13 ] . Behind the emergence of AI lies an equally important, if less celebrated, revolution in hardware. The massive gains in model performance have not resulted from better algorithms alone, but rather from the development of ever more powerful computers. Moore’s Law may be stalling, yet computing power continues to grow, fuelled by the global arms race in silicon. Nations are investing hundreds of billions of dollars in supercomputing, and semiconductors have become one of the world’s most coveted trade commodities [ 14 ] . The European Union alone has committed roughly €7 billion ($7.5 billion) to its EuroHPC Joint Undertaking (2021–2027) to build a world-class supercomputing ecosystem [ 15 ] . Meanwhile, the United States and China continue to push back the frontier, in a race to develop and deploy successive generations of exascale-class systems (albeit that fewer details have been revealed about their consolidated budgets) – an escalating competition that has come to define the infrastructure of modern science. These trends signal the emergence of a new mode of discovery and raise a fundamental question: Can the convergence of AI and HPC actually deliver the leap forward that science has been waiting for? In this article, we look beyond the anecdotal evidence and test, for the first time at scale, the joint effect of AI and HPC on scientific discovery. We show that their combined impact on science is both real and substantial. A growing share of high-impact, high-novelty research is now powered by these two technologies working in tandem, a trend that cuts across nearly every scientific discipline. But the shift also raises a number of red flags. As access to supercomputing and AI expertise becomes concentrated in a handful of countries [ 17 , 16 ] , elite institutions, and private firms [ 18 ] (see also Figure 3 ), there is a growing risk of a more stratified scientific ecosystem, one in which a few players increasingly determine the direction of discovery. We return to the implications for policy and for the scientific community in the concluding section. Results The convergence of AI and HPC in scientific research We compiled metadata for over five million AI-related publications, drawing on a curated list of keywords to search titles, abstracts, and author keywords in Scopus [see Supplementary Materials (SM) ]. To detect the use of high-performance computing (HPC), we mined acknowledgment sections for explicit mentions of computational resources (terms such as “HPC” or “supercomputer” ) – a strategy that reliably captures the actual use of large-scale computing in scientific research. This approach was validated through case studies conducted in collaboration with HPC facilities at the Universities of Strasbourg and Turin [see SM ]. In total, we identified over 200,000 publications acknowledging HPC use. Papers that fell at the intersection of our AI and HPC corpora were classified as AI+HPC, denoting research that explicitly combines the two technologies. AI+HPC research can take two broad forms: the use of HPC to either train AI models or to run large-scale inferences. To better understand this intersection, we applied a Large Language Model (LLM)-based classifier to full-text papers (when available) [see SM ], distinguishing between two categories. The first, AI-development, refers to research that builds or improves AI models, algorithms, or theoretical foundations. The second, AI-use, includes studies that apply existing AI tools and techniques to problems in other domains. Representative examples of each category are provided in the SM . Figure 1: Growth and convergence of AI and HPC in scientific