I study the effects of technology adoption on firm productivity in the French cloud computing services market, a setting with falling adoption costs yet a high concentration of suppliers. I study how IT-using firms make technology bundling choices when faced with these adoption frictions, and analyze the effects of competition policies on downstream adoption and welfare. I find that adopting cloud computing increases firm productivity from 0.2% to 1.6%, with heterogeneous effects across sectors and a larger impact for firms that purchase services from multiple cloud providers. To estimate the effects of competition policies to increase supplier switching, I estimate a model of industry dynamics, in which downstream firms produce output and make computing input bundle choices and suppliers compete to set prices. I find substantial entry and switching costs, of which egress fees are at most 8%. There is substantial complementarity between Microsoft products. A simulated merger between Microsoft’s and Google’s cloud platforms lowers average annual welfare by 0.28%, driven by an average price increase of 26%. By comparison, simulating the effects of a ban on data egress fees from the EU Data Act and an increase in software interoperability, I find these policies produce annual welfare gains of 0.06% and 0.16% (€183-491 million), respectively.

What is the impact of artificial intelligence adoption on firm productivity and labor automation? Using data from Spanish manufacturing firms during 2018-2024, I find that adoption of deep learning increases labor-augmenting productivity by 15.5%-17.2%, depending on sector, with no effects on Hicks-neutral productivity. This is roughly 3-5 times that of robots, which increase labor productivity by 2.8%-6.6%. Preliminary results suggest that AI is skill-biased, unlike robotics.

A large literature has argued that offensive advantage makes states worse off because it can induce a security dilemma, preemption, costly conflict, and arms races. We argue instead that state welfare is U-shaped under offensive advantage. We assess the offense-defense balance by considering a model where two states choose arms levels and decide whether to attack. High defensive advantage is first-best because the arms burdens required to deter attacks and maintain peace are low. High offensive advantage is comparatively worse because war is likely, but war tends to be smaller in scale, quicker, and less costly. Intermediate offensive advantage is worst because high arms burdens are required to deter attacks while wars, when they occur, are larger, longer, and more destructive. We discuss historical examples of this phenomenon, including the Warring States periods in China and Japan, the Imjin War, the Federalist Papers, Napoleonic Europe, and the World Wars.

Since its emergence around 2010, deep learning has rapidly become the most important technique in Artificial Intelligence (AI), producing an array of scientific firsts in areas as diverse as protein folding, drug discovery, integrated chip design, and weather prediction. As scientists and engineers adopt deep learning, it is important to consider what effect widespread deployment would have on scientific progress and, ultimately, economic growth. We assess this impact by estimating the idea production function for AI in two computer vision tasks that are considered key test-beds for deep learning and show that AI idea production is notably more capital-intensive than traditional R&D. Because increasing the capital-intensity of R&D accelerates the investments that make scientists and engineers more productive, our work suggests that AI-augmented R&D has the potential to speed up technological change and economic growth.

A formal model reveals how the information environment affects international races to implement a powerful, dangerous new military technology, which may cause a “disaster” affecting all states. States implementing the technology face a tradeoff between the safety of the technology and performance in the race. States face unknown, private, and public information about capabilities. More decisive races, in which small performance leads produce larger probabilities of victory, are usually more dangerous. In addition, revealing information about rivals’ capabilities has two opposing effects on risk: states discover either that they are far apart in capability and compete less or that they are close in capability and drastically reduce safety to win. Therefore, the public information scenario is less risky than the private information scenario except under high decisiveness. Finally, regardless of information, the larger the eventual loser’s impact on safety relative to the eventual winner’s, the more dangerous is the race.

Do workers benefit from improved computer processors? Using anonymized trace data from over 1 million users, we study the impact of more powerful CPUs on task performance, quantifying resulting speed gains and changes in misallocated resources. We are currently analyzing software and creative tasks.

Why have the impacts of ICT to US productivity growth declined since the 1990s? Using data from the Annual Survey of Manufacturers and Census of Manufactures, we plan to estimate production functions with IT-augmenting productivity in order to decompose effects into 1) computing technology changes and 2) non-ICT bottlenecks.

New technologies with military applications may demand new modes of governance. In this article, we develop a taxonomy of technology governance forms, outline their strengths, and red-team their weaknesses. In particular, we consider the challenges and opportunities posed by advancing artificial intelligence, which is likely to have substantial dual-use properties. We conclude that subnational governance, though prevalent and mitigating some risks, is insufficient when the individual rewards from societally harmful actions outweigh normative sanctions, as is likely to be the case with AI. Nationally enforced standards are promising ways to govern AI deployment, but they are less viable in the “race-to-the-bottom” environments that are becoming common. When it comes to powerful technologies with military implications, there is only one multilateral option with a strong historical precedent: a non-proliferation plus norms-of-use regime, which we call NPT+. We believe that a non-proliferation regime may, therefore, be the necessary foundation for AI governance. However, AI may exhibit characteristics that would make a non-proliferation regime less effective than it has proven for nuclear weapons. As an alternative, verification-backed restrictions on AI development and use would address more risks, but they face challenges in the case of advanced AI, and we show how these challenges may not have technical solutions. Perhaps more importantly, we show that there is no clear example of major powers restricting the development of a powerful military technology when that technology lacks a ready substitute. We, therefore, turn to a final alternative, International Monopoly, which was the preferred solution of many scholars and policymakers in the early nuclear era. It should be considered again for governing AI: a monopoly would require less-invasive monitoring, though at the possible cost of eroding national sovereignty. Ultimately, we conclude that it is too soon to tell whether a non-proliferation regime, a verification-based regime, or an International Monopoly is most feasible for governing AI. Nonetheless, a variety of policies would yield a high return across all three scenarios, and we conclude by identifying some of these steps that could be taken today.