Matias Covarrubias

So far the AI boom has been an investment boom in data centers and their equipment. We add two AI sectors, built from the engineering recipe of an AI data center, to the US production and investment networks, and let every industry buy AI services for the tasks that AI performs more cheaply than labour. We then ask what AI must deliver to justify the investment that firms plan. The answer is 2.1 percentage points of extra GDP growth a year from 2026 to 2040, first from AI itself and then from the capital that accumulates around it. The build-out raises the interest rate, and other industries make room by investing less, most of all where capital lasts longest. To the rest of the economy the boom is a demand shock to a few suppliers and a cost shock to everyone. Its purchases travel up the supply chain, the higher cost of capital travels down it into prices, and hours move from the services AI can do to the industries that build AI. Finally, we run an array of scenarios: a larger build-out, AI that disappoints, a geopolitical shock to the price of imported servers, and a regulatory pause.

In dynamic multisector economies, the planner's optimal capital allocation can serve to minimize the aggregate impact of shocks cascading through nonlinear production networks. We show analytically that (i) optimal capital allocation under uncertainty involves deliberately over-investing in upstream sectors in order to mitigate severe economic downturns; (ii) this efficient strategy reduces the average level of consumption and gives rise to a high welfare cost of business cycles. Deploying novel deep-learning techniques in a general environment we show quantitatively that (iii) the ergodic distribution of the simulated nonlinear economy features higher mean capital levels in key upstream sectors, lower mean levels of macroeconomic aggregates, realistic aggregate volatility and a welfare cost of business cycles one order of magnitude larger than in standard linear models.

We study the propagation of minimum wage shocks through the firm network, focusing on the 2016 hike in Hungary's minimum wages. The unique aspects of this case include the presence of two distinct minimum wages for skilled and unskilled workers, each increasing at different rates, and the fact that about half of Hungary's regular workforce was earning the minimum wage at this time. By integrating detailed employer-employee social security data with firm-to-firm value-added tax information, we are able to evaluate both the direct and indirect impact on firms. The indirect effects are further categorized into upstream (exposure through suppliers) and downstream (exposure through buyers) effects. We obtain three main results: (i) both minimum wage increases had a negative direct impact on employment by affected firms, such that, ceteris paribus, a firm that had average exposure to both minimum wages would have decreased headcount by 3.6% and hours by roughly 3.2%; (ii) upstream exposure to the skilled minimum wage increased headcount and hours, while downstream exposure to the skilled minimum wage decreases headcount and hours; and (iii) we find strong substitution effects between skilled and unskilled workers, that is, firms that had more exposure to the skilled minimum wage increased their number of unskilled workers, and vice-versa.

We evaluate the effect of dynamic oligopolistic strategies on the transmission of monetary policy and conclude that concentration increases the effectiveness of monetary policy mainly because fewer firms collude more often, and collusive prices react less to monetary policy. The main innovation of the paper is that we solve the strategic problem of firms by using state-of-the-art artificial intelligence algorithms called Deep Reinforcement Learning. These algorithms learn by experimenting and improving their strategies over time, using only information about their period rewards and the evolution of the state. Additionally, they approximate the policy and value functions using neural nets. This methodology allows us to study market structures with any number of firms, and we generalize the set of strategies to repeated game strategies and non-Markovian strategies. Using this approach, we construct a monetary economy with menu costs and oligopolistic competition at the sector level. After benchmarking with known results in the literature in a game with a finite (but large) horizon, we allow for repeated game strategies by solving the infinite horizon version. We get a range of converging equilibria, from the Markov equilibrium calculated in recent studies to high-markup equilibria that we interpret as tacit collusive equilibria. We find that monetary policy effectiveness, measured by its contribution to the volatility of log consumption, is up to 40% larger with collusive strategies versus non-collusive strategies. By inspecting the converging policies, we observe that collusion is supported by strong reactions to the price of the competition that ensure that deviations are costly. This stronger price complementarity accounts for the increased effectiveness. Then, we vary the number of firms per sector and show that fewer firms imply more effective monetary policy, mainly because collusion is easier.

This paper quantifies the welfare costs associated with market power in Spain. To this end, it examines markups in the product market using a dynamic model with heterogeneous firms, variable markups, and endogenous firm entry. Adapting the methodology of Edmond, Midrigan, and Xu (2023) to the Spanish context using data from the Bank of Spain's Central Balance Sheet Data Office, we identify three distortion channels: the aggregate markup as a uniform tax, markup dispersion generating inefficient resource allocation, and inefficient firm entry. Our results indicate that, for an average markup of 20% over marginal cost, the welfare cost amounts to 16.94% of the representative consumer's annual consumption, increasing nonlinearly to 30.83% when the average markup is 25%. The main source of welfare loss stems from the aggregate markup, followed by resource misallocation, which reduces total factor productivity by between 1.3% and 2.9%. Distortions in firm entry have comparatively smaller effects. These estimates are conservative due to the limited capture of markup heterogeneity and the use of moderate parameters for the materials share in production. The results reaffirm the importance of promoting competition in Spain and indicate that even moderate markup reductions can have a notable impact, particularly if initial markups are high.

This paper introduces a novel monetary policy tone index that leverages a Large Language Model (LLM) to measure the policy stance conveyed in ECB communications. Using this index, we assess the market reactions to dovish and hawkish speeches, where our main focus is on inter-meeting speeches by ECB Governing Council members. Our methodology involves asking an LLM to: (i) process each public intervention, (ii) fill out a form with questions about the speech's content, and (iii) justify its answers. We create three measures: stance (how hawkish or dovish the speech is), discrepancy, and surprise (distance from the last two speeches by the same member). We verify the accuracy of the index by using human annotators on a subset of speeches. Finally, we evaluate the high-frequency market reaction of the one-year Overnight Index Swap (OIS1Y) rate to understand the effectiveness of central bank communication and the underlying reasons behind these reactions. We find that for all three measures, hawkish (dovish) speeches increase (decrease) the short end of the yield curve. Among the three measures, surprise has the largest effect, followed by discrepancy and stance. Additionally, we find that hawkish speeches have an outsized effect, more than three times larger than dovish speeches. Finally, the impact of speeches made during recessions is more than double the impact in the full sample.