Tjai Yu Hui, National University of Singapore
Introduction
The artificial intelligence (AI) investment boom has been attributed as one of the key factors contributing to United States (US) and global growth this year. AI-related capital expenditure, which is capital spending on data processing equipment and software needed for AI, now surpasses consumer spending, which has historically been the largest contributor to US gross domestic product. This surge in investment is driven by the “hyperscalers,” namely Meta, Google, Microsoft, and Amazon, as they build data centres to meet future demands for computing power. With AI companies hitting valuations in the trillions, investments in AI research and development have truly exploded. However, is the current AI wave a sign of sustainable long-run growth with broad-based economic gains, or is it a bubble driven by speculative competition? To analyze this trend, we use Aghion and Howitt’s – the 2025 Economics Nobel Prize winners – model of creative destruction and Mokyr’s work on knowledge institutions.
Creative Destruction in Aghion-Howitt’s Framework
Aghion and Howitt (1992) build on Schumpeter’s (1942) theory of creative destruction, whereby old innovations, such as old products and production methods, are destroyed, and new innovations are created, resulting in profit. This economic reward is transitory as the technology diffuses to imitators, causing the profit to be competed away. The cycle then repeats with new innovations.
In Aghion and Howitt’s extension to Schumpeter’s theory, they analyse whether the private incentives to innovate generated by creative destruction are socially optimal. They identify the replacement effect, which causes under-innovation. Under the replacement effect, the prospect of future innovation reduces current innovation. Future innovation would replace current technologies, thereby destroying the monopolistic profits currently enjoyed by the incumbent. If firms believe that future research will rapidly erode their profits, then there is a risk of a no-growth trap, where no firm has the incentive to innovate.
Why AI Firms Keep Innovating
The heavy AI investment today seems to contradict the replacement effect because rapid innovation turnover has not discouraged innovation. However, Aghion and Howitt predict a fall in innovation only when future innovation is so large and rapid that the firm expects no profit. The current AI environment has rapid innovation replacement, but interim profits remain high.
The supernormal monopoly profits and the ability to capture market share through patents, hardware dominance, and first-mover advantages incentivises firms to innovate. The productivity benefits and global scalability of AI in various industries means that market size is large. The market for AI is expected to further expand as many organisations are still experimenting with AI usage or scaling up their AI usage. These result in potential profits of such a large scale that even if firms expect rapid AI innovation in the future, the monopoly profits that they can obtain in the short interval between their innovation and future innovations are sufficiently large for them to embark on research and development today. Besides profit, the strategic advantages in data and ecosystem control might not be fully destroyed by new innovation, thereby also preserving incentives to innovate.
Over-innovation and the AI Arms Race
However, more innovation is not always better. Aghion and Howitt identify another effect – the business-stealing effect – which causes over-innovation. Under the business-stealing effect, firms only consider their private benefit, which is the profits they can capture, and ignore the profit of the incumbent that their new innovation would destroy. For a social planner, this destroyed profit is a social cost. A social planner considers the benefits of the entire sequence of innovations, including the benefit from the old innovation that could have continued. Today’s AI race mirrors the prediction. Excessively rapid turnover of AI technologies has replaced innovations prior to their efficient diffusion to society and full utilisation of their social value. When AI models are frequently released and duplicated research is carried out across firms, firms are investing heavily to dominate in the next generation of technology though displacing, instead of complementing, their competitors.
Because firms do not internalise the profits destroyed, the model predicts excessive innovation “arms races”. In these “arms races”, firms invest aggressively not primarily to increase social return such as productivity or output, but to stay ahead of rivals in capturing transient private profits. This is reinforced by circular financing, where firms’ growth and valuations are due to strategic rivalry and mutual investment instead of proportional productivity gains. The gap between private and social returns of AI widens when computing resources are concentrated within a handful of firms. The entrance of low-cost competitors such as Deepseek could shorten the monopoly interval and reduce the reward to innovation. Yet, the increased competition could push firms to race even harder to avoid being knocked out of the frontier. The innovation race intensifies, more capital is poured into research, yet the profits are smaller, and social gains are modest at best.
The Role of Knowledge Institutions for Sustainable Growth
To avoid an inefficient innovation race, Mokyr’s work shows that long-run growth does not only depend on the innovations themselves, but also culture and institutions that influence knowledge creation and diffusion. In explaining growth during the Industrial Revolution, Mokyr emphasises that cultural attitudes that encouraged knowledge sharing and inquiry, which were supported by institutions such as universities, scientific academies, and governments, allowed innovations to be widely adopted and accumulate into sustained productivity growth. Applying this to AI, there needs to be institutions that promote openness to new ideas and collaboration in research, for example between firms, universities, and public research labs, thereby reducing the concentration of technological capacity currently present in the private industry. Efforts in research and development become less fragmented, so broad-based productivity gains are achieved, and duplicative business-stealing effects are reduced. Industry standards and norms, such as for interoperable systems and shared benchmarking practices, encourage cumulative downstream innovation as institutions can build on existing innovations. Additionally, human capital investments are needed for more efficient diffusion and adoption of technology across sectors in society. Without these cultural and institutional supports, the AI boom risks being dominated by business-stealing dynamics, where diminishing returns are clustered among a few firms, instead of AI contributing to sustained growth.