A Nobel Economist's View on AI and the Case for Maintenance-First Thinking

The dominant narrative around AI investment has been expansionist — build more, release faster, scale further. But a quieter counter-argument is gaining traction in academic and policy circles: that the infrastructure underpinning existing systems may deserve as much attention as the systems being built on top of it. Two related perspectives are shaping this debate, one from an economist who has studied inequality and technology for decades, and one from a broader argument about how societies allocate effort between creation and upkeep.

Daron Acemoglu, the MIT economist who received the Nobel Prize in Economics in 2024, has been one of the more measured voices on AI's productive potential. His position is not that AI lacks impact, but that the scale of transformative economic effect attributed to current systems is likely overstated. He has argued that the tasks AI automates well are a narrower subset of economically significant labor than the industry's projections imply, and that productivity gains will be more concentrated and slower to diffuse than headline forecasts suggest.

This framing matters because it directly challenges how enterprises and investors are currently allocating resources. If AI's near-term productivity contribution is more modest than assumed, the calculus around wholesale workforce restructuring and infrastructure replacement looks different than it does under optimistic scenarios.

The second thread of the argument concerns maintenance. There is a long-standing structural bias in how institutions — whether governments, corporations, or research bodies — reward new construction over the repair and upkeep of what already exists. This applies to physical infrastructure, but it extends equally to software systems, data pipelines, institutional processes, and organizational knowledge. New builds generate visibility and credit. Maintenance does not.

For AI adoption specifically, this dynamic creates a compounding problem. Organizations rushing to deploy new AI capabilities frequently do so on top of data infrastructure, integration layers, and operational processes that have not been adequately maintained. The result is that AI systems are being asked to perform reliably on foundations that were not designed for the task and are not being updated to meet it. Garbage-in conditions persist, and AI systems amplify rather than resolve the underlying disorder.

Acemoglu's skepticism and the maintenance argument converge on a shared implication: the value extraction from AI will depend heavily on whether organizations build the discipline to manage and sustain systems over time, not just deploy them. Companies that treat AI as a procurement decision rather than an operational capability will find the gap between expected and realized value difficult to close.

The broader institutional implication is that the AI sector's incentive structure — which rewards releases, announcements, and benchmark performance — may be misaligned with the kind of sustained, less visible work that determines whether these systems actually function in production environments. Evaluation frameworks, audit processes, data governance, and model maintenance cadences are less photogenic than capability launches, but they are where operational AI either succeeds or degrades.

From an execution standpoint, this suggests that organizations with a genuine maintenance culture — those that have already built practices around system reliability, data stewardship, and process review — are better positioned to realize durable AI value than those operating on a continuous-deployment-without-upkeep model. The differentiator in AI adoption may not be which tools are deployed first, but which organizations have the operational infrastructure to keep them working.

Sources: — MIT Technology Review (https://www.technologyreview.com/2026/05/12/1137103/the-download-nobel-winner-ai-maintenance-of-everything/)