From Micro to Macro via Production Networks
GAICIL Insights
Calling the economy a network can sound like a buzzword dressed in a lab coat.1 From Micro to Macro via Production Networks, an essay by Vasco M. Carvalho, takes that objection head on. Viewing the economy as a complex production network, the essay concedes, may seem at first hearing to be yet another fuzzy analogy coated in big words. Then it shows the analogy has teeth: testable hypotheses about how local shocks propagate and give rise to aggregate fluctuations.
The essay maps production networks onto a standard general equilibrium setup, explores the empirical properties of a large-scale production network using detailed US input-output data, and discusses how theory and data combine to explain comovement and aggregate volatility. Through stylized examples it demonstrates that the propagation of sectoral shocks, and hence aggregate volatility, depends on the network's shape. Some arrangements dissipate shocks. Others concentrate them. The essay concludes that the immediate implications of the networked view cannot be reasonably refuted, and that theory and empirics together challenge a long-standing irrelevance argument holding that sectoral shocks must average out.
Apply the lens to artificial intelligence and a new map of the economy appears. AI capabilities come from a few specialized sectors, semiconductors, cloud computing, model development, and are used as inputs everywhere else. In network terms these sectors are becoming hubs. The essay implies that shocks to hubs have outsized aggregate effects. A large leap in the efficiency of AI models would propagate widely. So would a disruption in chip supply. The economy is growing a new set of load-bearing walls.
The richest part of the essay is its agenda. Nodes in the analysis are sectors, yet input sourcing decisions really happen at the level of the plant or firm. Choosing the relevant node, firms, sectors or both, raises three complications. At the firm level it is harder to brush aside market structure. It becomes essential to distinguish easily substitutable inputs from crucial, hard-to-substitute inputs where firms are locked in and switching costs are large. And firm-level input-output information is in very short supply.
Each complication describes AI almost exactly. Market structure in AI inputs, from chips to foundation models, is concentrated in places, and its dynamics shape the price and quality of AI downstream. The line between substitutable and hard-to-substitute inputs decides how exposed downstream firms are to their AI suppliers' choices. Lock-in and switching costs are recurring worries in cloud and AI services. And the data shortage limits anyone's ability to map dependencies. A firm that cannot switch its model provider in a week sits on a different branch of the network than one that can.
The essay points to recent advances in the theory of firm-level networks and to novel data sources as first steps. AI can help here too, by assembling firm-level production networks from administrative records and transaction data. Better data would let policymakers pinpoint critical nodes, evaluate supply chain resilience and anticipate how innovations spread.
The policy implications point toward openness and substitutability. If hard-to-substitute inputs with high switching costs amplify shocks, then interoperability, multiple sources of supply and contestable markets for key inputs reduce aggregate vulnerability. The same policies intensify competition among suppliers, which drives innovation and lowers downstream prices. Resilience and dynamism turn out to be served by the same medicine.
A quieter lesson concerns portability. Policies that make it easy to move workloads, data and models between providers shrink switching costs, and shrinking switching costs turns hard-to-substitute inputs into substitutable ones. That single design choice changes the network's shock-absorbing capacity.
The essay's broader contribution is to show that an economy's architecture shapes its behavior. Aggregate outcomes emerge from the links between producers. For a technology like AI, entering production through countless links, that architecture will set both the scale of benefits and the channels through which risks travel. The fuzzy analogy, it turns out, is a blueprint, and reading it is the first step toward policies that maximize gains and manage risks.
Introduction
A modern economy is an intricately linked web of specialized production units, each relying on the flow of inputs from their suppliers to produce their own output which, in turn, is routed towards other downstream units. In this essay, I argue that this network perspective on production linkages can offer novel insights on how local shocks occurring along this production network can propagate across the economy and give rise to aggregate fluctuations. First, I discuss how production networks can be mapped to a standard general equilibrium setup. In particular, through a series of stylized examples, I demonstrate how the propagation of sectoral shocks—and hence aggregate volatility— depends on different arrangements of production, that is, on different “shapes” of the underlying production network. Next I explore, from a network perspective, the empirical properties of a large-scale production network as given by detailed US input-output data. Finally I address how theory and data on production networks can be usefully combined to shed light on comovement and aggregate fluctuations.
Takeaway
Viewing the economy as a complex production network may seem, at least at first hearing, as yet another fuzzy analogy coated in big words. In this essay, I have attempted to show that this perspective can indeed offer testable hypotheses and insights by mapping it to a standard general equilibrium setup and showing how this provides guidance for empirical explorations of input-output data. Looking at sectoral comovement from this vantage point, I have shown that the immediate implications of this networked perspective cannot be reasonably refuted. Furthermore, as I have discussed, theory and empirics together provide a challenge to a long standing "irrelevance" indictment in the literature. To go beyond these suggestive possibility results, a small but fast expanding literature on production networks is hard at work on a number of important challenges. First, while throughout this essay I have kept equating nodes to sectors, input sourcing decisions actually take place at the level of the plant or of the firm. The question therefore concerns what constitutes the relevant node: firms, sectors or both? Relative to sectors, progress on firm-level production networks needs to deal with three added complications. First, on the theory side, it is more difficult to brush aside the complexities of market structure (as I have done here by appealing to identical, perfectly competitive firms inside each sector).1
Second, at this level of disaggregation it is clear that we have to distinguish between easily substitutable inputs and crucial, hard-to-substitute, inputs where firms are locked-in and switching costs are large. Third, relative to sector-level data, input-output information at the firm-level is in very short supply. Recent advances in developing a theory of firm-level networks (Oberfield 2013) and the availability of novel data sources provide important first steps in this direction: for examples, see Bernard et al. (2014) and Carvalho et al. (2014) for data on Japan and Atalay et al. (2011) for US data. Second, the quantification and empirical validation of the network viewpoint is another active area of research. Working with calibrated dynamic extensions of the simple multi-sector model set forth here, Carvalho (2010) and Atalay (2014) find a far from negligible role for idiosyncratic shocks, echoing the earlier findings of Horvath (1999). The results in Carvalho and Gabaix (2013) regarding the dynamics of aggregate volatility, are consistent with these findings albeit working under a much simpler setting.
Notes
- Vasco M. Carvalho, From Micro to Macro via Production Networks, 28 J. Econ. Persp. 23 (2014), https://doi.org/10.1257/jep.28.4.23. Takeaway source: Looking Ahead, reproduced from a free author or repository copy (working-paper version). ↩
Cite as: Vasco M. Carvalho, From Micro to Macro via Production Networks, 28 J. Econ. Persp. 23 (2014), https://doi.org/10.1257/jep.28.4.23
Header image: Coastal Highway Bridge at Night.jpg by Andrew Parlette, CC BY 2.0, via Wikimedia Commons.
