Economies of Scale, Part I: Artificial Intelligence

Adam Smith’s 1776 book An Inquiry into the Nature and Causes of the Wealth of Nations opens with his famous description of a pin factory. He notes that previously, pins would have been hand-made by a single craftsman; by his day, each step in making the pin from pulling the wire, attaching the head, sharpening the point, and readying the box of pins for market would be done by a different specialist. With a division of labor, ten people could produce 48,000 pins per day, compared to the craftsman who could scarcely produce a single pin in that same time period. 

But who needs 48,000 pins? In Chapter III, Smith notes “That the Division of Labour is Limited by the Extent of the Market.” The scale efficiencies produced by the division of labor will only be incentivized if the producer can sell to a sufficiently broad market. Economies of scale and the size of markets were therefore intimately related. On the eve of the industrial revolution, European manufacturers had greatly increased their productivity by selling to markets that had been vastly widened through the advent of water-borne transportation. The continent’s burgeoning wealth did not originate in the manor houses and countrysides where the aristocracy lived, but in cities with ports located on oceans or large rivers where commerce could support large populations. 

The interplay between the size of markets, the division of labor, and economies of scale remain as important today as they did in Smith’s time.  The globalization that has occurred over the past couple of generations has been driven by stunning decreases in the costs and time required to move goods from one part of the world to another.  The contemporary manifestation of Smith’s division of labor are global supply chains involving thousands of companies spread across multiple time zones and jurisdictions. Added to this are communications technologies that allow coordination of all these activities.  Ports and waterborne transportation remain as important to this system as in Smith’s day; try getting the contents of a Panamax container ship to Central Asia or to the interior of sub-Saharan Africa and you'll see one of the sources of poverty there. The economies of scale that modern markets support then have large consequences for politics and the broader society. 

One manifestation of the enduring importance of scale is artificial intelligence. The dawn of the personal computer and internet in the 1990s raised hopes that digital technology would be made available at a small scale to millions of ordinary users, and would consequently democratize the distribution of power in societies. In some respects this proved true; information became much cheaper and more widely available, while the internet permitted social mobilization across borders.___STEADY_PAYWALL___

But with the passage of time, economies of scale reasserted themselves. In particular, network externalities gave advantages to large social networks. Instead of dispersed communities of interest, the internet supported a small number of huge tech platforms like Meta, Google, and Amazon. These companies acquired enormous amounts of data about their users, which in turn allowed them to entrench their economic and political power.

Artificial intelligence only increases the importance of scale. Early on, big data required large populations on which to train algorithms, which gave advantages to large countries like the United States and China. But cutting edge generative AI requires even greater scale. Fei-Fei Li, director of Stanford’s initiative on Human-Centered Artificial Intelligence, recently wrote that her university lab simply could not afford the machines on which research depended. These existed only in the private sector, which had the resources to hire away her graduate students and faculty at an alarming rate. The U.S. government’s laboratories are in no better a position to conduct this work either, in terms of equipment or human resources. 

The importance of scale then raises some significant issues for democracy itself. Concentrated economic power inevitably leads to concentrated political power. This power can be exercised in traditional ways, such as lobbying Congress over regulation to protect a company’s economic self-interest. But with the advent of social media, political power could be wielded directly through control of political discourse on large platforms. We saw a vivid example of this with Elon Musk’s takeover of Twitter: He didn’t like what he saw as the leftward bias of content moderation on that platform, so he bought the company and made it much more hospitable to rightwing views, including extremist voices formerly regarded as outside the Overton window. He has also been able to run his own foreign policy, denying use of his Starlink satellite system to the Ukrainians seeking to attack Russian forces in Crimea. 

Platform political power has created pushback and multiple efforts to control the sector. The left was initially mobilized by what it saw as the platforms’ enabling of Donald Trump’s rise in 2016, crystalized by the Cambridge Analytica scandal. As these companies sought in response to moderate what they saw as hate speech and disinformation, the right reacted against what they saw as liberal bias and censorship of conservative voices. Polarization led to the standoff we currently see, where both sides agree that platform power should be limited, but can’t agree on how or for what purposes. 

The advent of generative AI will change these political considerations once again. The technology is seen as so foundational and potentially powerful that it will quickly become a critical component of national competitiveness. It will also have direct strategic value by enhancing military capabilities in a myriad of ways.  There has been talk and some tentative efforts to regulate AI in both the United States and Europe. But lurking in the background is the geopolitical competition with China. If it is the case that only large private companies can rapidly push the frontiers of the technology, will governments want to stand in their way? They certainly have the means to do so: they could nationalize the companies, break them up through antitrust, or strictly regulate them. But each of these approaches generates significant costs in terms of innovation, speed, risk-taking, and ultimately the ability to achieve scale economies. 

We certainly don’t want to accept the lazy conclusion that geopolitical competition forces us to welcome scale and that we should allow large tech companies to grow as large and powerful as they want. But there may be a critical difference between social media and generative AI. Large scale in the former realm is, on the whole, a bad thing for democratic political discourse. The world would likely be a better place if the large social media platforms were broken up and political discourse was less subject to their ability to amplify or suppress certain points of view. This was the essence of the middleware proposal put forward by the Stanford working group on platform scale that I chaired. But greater diversity may not be as desirable if we are to make progress with the latest forms of AI, given what seems to be the intrinsic nature of the underlying technology. Only time will tell if this proves true.

Francis Fukuyama is chairman of the editorial board of American Purpose and Olivier Nomellini Senior Fellow and director of the Ford Dorsey Master’s in International Policy program at Stanford University’s Freeman Spogli Institute for International Studies.