At a public event on AI for science last month a young researcher asked what is left for academics to do and how they can compete with industry. A question that would have raised eyebrows a decade ago is now met with concerned nods in the audience. Training a large foundation model or building a 100-qubit quantum computer are no longer tasks within the reach of most academic institutions. So far in both AI and quantum technologies, industry and academia have travelled together on the same path, helping each other to push forward, but academics feel they have been left behind and worry that the paths are starting to diverge.

This anxiety is not unfounded. “Compared to two decades ago, quantum computing is no longer dominated by academic research groups, but by commercial providers,” write Tsubasa Ichikawa and colleagues in a Down to Business article in this issue. They look at the numbers of qubits available on commercial quantum computing services and what type of applications they are used for by researchers. Despite the democratization of access to quantum computers, one can already see that users affiliated with companies have access to more qubits than standard users. This is probably because they can use larger processors and have know-how on optimizing applications for the specific hardware. IBM devices are most widely used, dominating in terms of the number of publications produced. One of the contributing factors to this trend is likely to be the popularity of the quantum programming language Qiskit released by IBM in 2017.

The quantum computing industry is at an early stage, but in AI research and AI for science the dependence on hardware and software developed in industry is more dramatic. Even without considering high-performance computing, at the level of any lab regardless of size, physics research depends on GPUs for computation. The GPU market is dominated by Nvidia, which has a tradition of working closely with and investing in academia, but historically, the semiconductor industry has seen alternating periods of rapid growth followed by slumps — boom-and-bust cycles1. Sudden price hikes or shortages are likely to affect academia more, making hardware unaffordable. In terms of software, libraries such as PyTorch and TensorFlow are widely used in academia. These tools are free, but are developed with heavy investment from industry.

Last year, Nur Ahmed and colleagues highlighted the fact that industry is becoming more influential in AI research in terms of academic publications, cutting-edge models, and key benchmarks2; a similar situation likely exists in AI for science at least in terms of top publications. This dominance stems from access to more computing power, data and human talent2. Whereas in some areas of physics there is no shortage of data, the availability of computing power, especially for smaller research groups, is a problem. The brain-drain from academia to industry documented in computer science2 is also increasingly lamented in physics, as we heard at recent conferences.

As much as academia relies on industry for hardware and software, industry depends on academia for training the next generation of scientists and pushing the knowledge boundaries in unexpected directions. Joel Klinger and colleagues found that “private sector AI researchers tend to specialize in data-hungry and computationally intensive deep learning methods”3; a trend that can result in a premature narrowing of AI research and a slow-down of innovation.

“industry’s strength and academia’s agility. Both are needed for a sustainable future of AI research and AI-empowered scientific discovery”

In a Comment in this issue, Petros Koumoutsakos argues that there is plenty of room between AI and traditional computational science and researchers from the two fields would benefit from more interactions. To illustrate this point Koumoutsakos uses the legend of Richard the Lionheart’s big sword that can cut a big rock yet fails to slash a feather pillow that Saladin’s nimble blade cuts easily. Lionheart’s big sword stands for the AI brute force approach and Saladin’s one for the more subtle methods of computational science, but researchers need both. This metaphor also extends to the relationship between industry’s strength and academia’s agility. Both are needed for a sustainable future of AI research and AI-empowered scientific discovery, but the question of how to create a system that brings the best of both worlds does not have an easy answer.