Month: March 2025

For users of AI tools, the discussion often focuses on the newest version of ChatGPT or Gemini or DeepResearch. But the point of contact of the AI industry with users is of course just one part of the AI industry as a whole. as a whole. Leonardo Gambacorta and Vatsala Shreeti lay it out in “The AI supply chain” (Bank of International Settlements Papers No. 154, March 2025). Here is their version of the AI supply chain:

Based on this structure, here are some of the key player in the overall AI ecosystem:

Of course, these graphs are an overview rather than a detailed presentation, but that’s often a useful way of gaining perspective. Gambacorta and Shreeti provide more detail in the text. For example (citations and references to graphs omitted):

Hardware. Consider the most important hardware for AI applications, namely microprocessors like GPUs. Nvidia – headquartered in Santa Clara, California, in the United States – serves most of the market for GPUs, with its market share reported to be larger than 90%. It has gross margins of over 70% and has seen its revenues increase by 405% between 2023 and 2024. Initially serving the video game market, Nvidia had a head start in leveraging the parallel computing capacity of its GPUs for AI models. Over time, it has built-up substantial intellectual property and a significant reputation, solidifying its position as the market leader for GPUs. Apart from the GPUs themselves, Nvidia also produces complementary software. Nvidia’s GPUs come in an exclusive bundle with CUDA, its parallel computing platform, which enables programmers and software developers to simplify the process of using GPUs and to enhance their performance. CUDA has become the industry standard for programmers and can only be used with Nvidia’s GPUs. …

To be sure, several other firms, including startups and big techs, are also active in the market for AI hardware. Advanced Micro Devices (AMD), Intel and big techs like Microsoft, Google and Amazon are all producing AI microprocessors to compete with Nvidia’s GPUs, both for training AI models and for inference. Chinese companies like Alibaba, Baidu and Huawei are also starting to produce their own microprocessors, especially in light of geopolitical constraints. …

Cloud computing layer. Globally, the cloud computing market is dominated by three big tech companies: Amazon Web Services (AWS) with a market share of 31%, Microsoft Azure with 24% and Google Cloud Platform with 11%. In the European Union (EU), the estimated combined market share of AWS and Azure in 2020 was over 80% and their profit margins were also reported to be high, at 30% and 38%, respectively. In the case of the IaaS segment – the most relevant one for AI models the market is even more concentrated. In 2023, AWS, Microsoft Azure and Google Cloud Platform together accounted for nearly 74% of the global market. …

Training data. So far, frontier AI models have been trained using vast troves of publicly available data. However, as the stock of public data rapidly declines, firms are turning to other data sources. …

Foundation models. At first glance, the market for foundation models is dynamic and rife with competitors. There are over 300 foundation models in the market, provided by 14 different firms. There are also competing business models – while some firms choose to offer proprietary foundation models (like OpenAI and Google DeepMind), others have adopted a relatively more open approach (notably Meta with its open source Llama models and, more recently, DeepSeek). … Nevertheless, the market for foundation models is currently dominated by only a handful of firms like OpenAI, Google DeepMind, Anthropic and Meta. In 2023, despite numerous competing foundation models, OpenAI’s GPT-4 accounted for 69% of the market for generative AI in terms of global revenue. Given the dynamic nature of the market and the potential to realise efficiencies, the hierarchy may shift rapidly. …

AI applications and user-facing layer. The last stage of the AI supply chain, the user facing layer, follows the playbook of digital platforms and mobile applications. Since the “ChatGPT moment” of AI, applications built on top of foundation models have been proliferating in various sectors of the economy including health, education, backend processing and compliance, software development and others. Nonetheless, and as with digital platforms, there can be a risk of “winner takes all” dynamics emerging in the markets for AI applications. While it is a Herculean task to trace the market for AI applications in every sector, the market for chatbots can be instructive. … [D]espite a flurry of similar interfaces, ChatGPT still accounted for 60% of the chatbot market (measured by the total number of monthly visits) in 2024, highlighting the importance of being first to market.

One lesson from this overview is that if one believes that AI tools are going to be important drivers of productivity going forward, then a leading economy like the United States should be interested in strengthening the entire ecosystem for developing and using AI. A harder lesson is that with a very rapidly evolving technology, the best approaches for firms, users, and policymakers are not at all obvious in advance. Firms that seem to have a dominant market position in early 2025 may not have one in six months or a year or two, whether policymakers take any action or not. Lessons are learned about a technology as firms and strategies rise and fall, and to a substantial extent, policymakers should let the process play out so tha these lessons can be learned. In particular, policymakers have a tendency to listen to the loudest complainers, rather than giving equal weight to the beneficiaries who are too busy building and using these tools to spend time lobbying about them.

Industrial policy can be defined as government policies that seek to shift the sectoral structure of an economy: for example, toward a certain high-technology industry, or toward a manufacturing industry, or in a lower-income country away from a heavy production of raw materials and agricultural goods and toward a greater emphasis on processing those outputs. Tools for industrial policy can involve direct subsidies, indirect subsidies (like low-interest loans or building key infracture), subsidies based on export performance, reducing import competition via tariffs and other methods, reducing barriers to trade to make it cheaper to purchase key inputs and goods, rules requiring domestic content in production of certain goods, direct government purchases of goods, and others. As one might expect, such policies have sometimes been successful, and sometimes have resulted in pouring resources down a rat hole.

A group of IMF economists–Sandra Baquie, Yueling Huang, Florence Jaumotte, Jaden Kim, Rafael Machado Parente, and Samuel Pienknagura–have published a Staff Discussion Note attempting to summarize some lessons in “Industrial Policies: Handle with Care” (IMF, SDN/2025/002, March 2025). Here are a few of their themes that caught my eye.

The standard examples cited as success stories for industrial policy tend to be countries from East Asia: Japan, Korea, Taiwan, China. The standard examples cited as failures of industrial policy are often countries from Latin America. One intriguing distinction here is that the success stories were outward-focused, requiring that industries meet export targets in the global economy to receive subsidies, while the failures often involved import substitution, in which imports were blocked to support domestic consumption for domestic use. The IMF economists write:

The debate around the effectiveness of IPs [industrial policies] has centered around two somewhat opposing narratives. On the negative side, there is the experience of Latin American countries with import-substitution, which, after two decades of favorable economic performance between the 1950s and the 1960s, have struggled to achieve high-productivity growth. On the positive side, there is the experience of Asian economies, such as Hong Kong SAR, Japan, the Republic of Korea, Singapore, and Taiwan Province of China, which focused on export-led growth and provided a blueprint for proponents of IPs. The divergent paths of Latin American and Asian economies have led many observers to stress the importance of design. For example, Cherif and Hasanov (2019) contrast the limitations of IPs focused on the development of domestic markets through import protection, which was the Latin American model, with the virtuous cycle fostered by the export-driven IP model pursued in Asia.

Another theme is that if one compares broad-based “structural” policies to more targeted “industrial” policies, the structural policies often have substantially larger and longer-lasting effects. Indeed, structural reforms may be a precondition for industrial policies to succeed. the IMF economists write:

Structural reforms have, on average, much larger effects than IPs [industrial policies], pointing to their fundamental role. IPs are accompanied by smaller economic benefits than “horizontal policies” focused on lowering corruption, improving governance and enhancing access to credit. Even when IPs may be desirable, horizontal policies are key. IPs are more effective in countries with better institutions, business environment, and financial market conditions, and a more educated workforce. Good institutions limit the capture of IPs by interest groups and facilitate their successful implementation. A strong business environment eases the flow of factors of production to targeted, fast-growing, firms, and pushes them to remain competitive. Efficient financial markets allow targeted firms to get a double boost to unlocking their potential, as IP support can be combined with private credit to seize profitable projects.

Industrial policies that reduce trade barriers tend to produce greater gains than import tariffs that limit trade. The IMF discussion note says:

Relatedly, trade-liberalizing IPs—those that reduce trade restrictions—are associated with higher firm productivity and value added in the medium term, with negligible change in the stock of capital. … An additional liberalizing policy is associated with improved medium-term performance of firms: 1.6 percent higher productivity, 1.2 percent higher value added, 0.8 percent more payroll (a proxy for wages and employment), and 0.4 percent more capital stock although the latter is not statistically significant (Figure 8, panel 3). The positive association between liberalizing trade conditions and firm productivity and value added relates to a long-standing literature on how lower trade barriers can strengthen competition in the liberalized sectors, inducing firms to leverage economies of scale, improve efficiency, and innovate (Helpman and Krugman 1985; Melitz 2003, Aghion and others 2005). Differently from export incentives and domestic subsidies, which are targeted in nature, liberalizing trade barriers yield a uniform impact across firms within the targeted sector. The results are in line with the finding that industrial subsidies targeting high-externality sectors yield smaller welfare gains compared to trade liberalizing measures such as broad-based tariff reductions (Bartelme and others 2019).

Lowering import barriers also favors technological transfers in the medium and long term. Although well-targeted protectionist IPs may temporarily boost received technological transfers, lifting trade-restricting policies unlocks larger and potentially longer improvements. Lifting an additional import barrier increases the number of received patent applications by 5 percent on average after four years …

Finally, industrial policies aren’t free, but impose both monetary and nonmonetary costs. The IMF economists write:

On the fiscal front, IP [industrial policy] expenditures in the 2019–21 period in a sample of OECD countries amounted to about 1.4 percent of GDP (Criscuolo and others 2023). Thus, in the context of high debt levels, IPs can limit governments’ ability to save and/or redeploy resources to tackle other challenges. IPs can also affect sectors or firms that are not targeted, through the reallocation of sales or resources to the supported entities. This action may not be welfare enhancing if IPs are not well targeted and this reallocation harms more productive sectors or firms. Moreover, the current geoeconomic landscape adds to the complexity. IPs can lead to cross-border spillovers, raising the risk of retaliation by other countries, which can ultimately weaken the multilateral trading system and worsen geoeconomic fragmentation. This, in turn, can also limit global welfare by stifling innovation incentives and the flow of new technologies across countries.

From my own perspective, arguments in favor of industrial policy often devolve into a list of injustices purportedly faced by US firms in global markets. The injustices sometimes seem real to me; other times, not so much. Also, I am perplexed by anyone who expects global markets to reflect concerns of justice–especiallly if these concerns somehow lead to cutting off support for the trade dispute resolution mechanisms of the World Trade Organization.

But perhaps more to the point, the arguments over what is “fair” in global trade often seem to me a way of blaming others while failing to tackle domestic policy issues. For example, the US economy has real challenges with K-12 education and worker training, with its oversized budget deficits and growing levels of government debt, with keeping its research and development at the cutting edge of technological progress, with rules and regulations that give small pressure groups a way to stifle both good and arguably harmful development, and with other issues as well.

It’s crystal-clear that decision-making by algorithm can be imperfect. It’s also crystal-clear that decision-making by humans can be pretty imperfect, too. However, the imperfections across these two types of decisions are probably not the same. How and when should society make use of decision-making by algorthm? Cass R. Sunstein provides a thoughtful overview in “The use of algorithms in society” (Review of Austrian Economics, December 2024, 37: 399–420).

Sunstein emphasizes two considerable advantages of decision-making by algorithm: it can avoid bias and noisiness. Bias arises when a certain judge or doctor is overly influenced by certain factors, like whether they saw what seemed like a similar case recently. Noisiness is just the human attribute of inconsistency, where decision rules might be applied different ways if the weather is good or bad, or before or after lunch. (For an overview of decision-making in bail cases, with reference to bias, noisiness, and a possible role for algorithms, a useful starting point is “The US Pretrial System: Balancing Individual Rights and Public Interests,” by Crystal S. Yang and Will Dobbie, in the Fall 2021 issue of the Journal of Economic Perspectives, where I work as Managing Editor.)

Here, I want to focus on some of the reservations that Sunstein summarizes about algorithmic decision-making–some of which might be overcome with better systems over time, some perhaps not.

1) Even if the algorithm does a better job than most humans, some humans will do better than the algorithm. Sunstein writes:

Some important work suggests that while algorithms outperform 90% of human judges in the context of bail decisions, the top 10% of judges outperform algorithms. The reason appears to be that the best judges have and use private information to make better decisions. They consider factors that algorithms do not. They appear to have something like local knowledge – an understanding of the defendant or the circumstances that algorithms lack. We could easily imagine a similar finding for doctors. It is possible that the best doctors know whom to test for heart disease, because they see something, or intuit something, that algorithms do not consider.

2) The gain from using algorithms in many contexts is relatively small in percentage terms, although a small percentage gain applied to a large number of people can certainly be meaningful. Sunstein writes:

As I have noted, algorithms do better than people do, but they do not do spectacularly better. The impressive aggregate figures, in terms of welfare gains, come from the fact that very large populations are involved. If algorithms show a modest percentage increase in accuracy as compared with human beings, we might find seemingly major improvements. If an algorithm can produce a slight increase in the accuracy of screening for heart disease, we might see a significant reduction in deaths. But a slight increase in accuracy remains slight. I have said, for example, that formulas do better than clinicians. But in the median study, formulas are right 73% of the time, while clinicians are right 68% of the time.

3) People are more unforgiving of algorithmic error than of human error. Sunstein:

In short, people are less forgiving of algorithms than they are
of human beings. … Is that rational? If people want to make the correct decision, it is not. If their goal is to make money or to improve their health, they should rely on the better decider. But one more time: if people enjoy making decisions, a preference for making one’s own decisions might be perfectly rational. Perhaps people find the relevant decisions fun to make. Perhaps they like learning. Perhaps decision-making is a kind of game. Perhaps they like the feeling of responsibility. Perhaps they like the actuality of responsibility. If so, algorithm aversion is no mistake at all.

4) Greater complexity will limit the benefit of an algorithm. For example, the algorithm behind a dating app might offer recommendations that are slightly more likely to be successful than those who went unrecommended–but certainly no guarantee of true love. Sunstein tells an interesting story of a prediction competition called the Fragile Families Challenge., which teaches humility about the predictions of algorithms.

There’s a dataset called the Fragile Families and Child Wellbeing Study. It collected data on thousands of families with unmarried parents, where the mother gave birth to a child around the year 2000. The study collected a LOT of data about these families at the time of birth, and then followed up at ages 1, 3, 5, and 9. The challenge was whether this earlier data could be used to predict certain outcomes about the child or the family when the child turned 15. Hundreds of analysts applied to this contest, and 160 teams were selected. As it turned out, many of the predictions were similar to each other, but literally none of them were very accurate. Sunstein writes:

The central question was simple: Which of the 160 teams would make good predictions? The answer is: None of them. True, the machine-learning algorithms were better than random; they were not horrible. But they were not a lot better than random, and for single-event outcomes – such as whether the primary caregiver had been laid off or had been in job training – they were only slightly better than random. The researchers conclude that “low predictive accuracy cannot easily be attributed to the limitations of any particular researcher or approach; hundreds of researchers attempted the task, and none could predict accurately.” Notwithstanding their diverse methods, the 160 teams produced predictions that were pretty close to one another – and not so good. As the researchers put it, “the submissions were much better at predicting each other than at predicting the truth.” A reasonable lesson is that we really do not understand the relationship between where families are in one year and where they will be a few years hence. … You can learn a great deal about where someone now is in life, and still, you might not be able to say very much at all
about specific outcomes in the future.

5) Algorithms are not good, and perhaps cannot be good, at what are sometimes called path-dependent events. The question when a certain political movement will rise or fall, or whether a certain musical act or movie will become popular, will rely on the unfolding of a chain of events. Algorithm rely on patterns from the past events, and may not be good at predicting the timing or probability of future chains of events.

The challenge is consider both the benefits and tradeoffs of algorithms in different settings. If someone I love is facing a bail decision, and I don’t know who the judge will be, I personally would prefer that an algorithm make the decision. When it comes to the algorithms that govern self-driving cars, many people are clearly much less forgiving of errors and accidents caused by an algorithm than they are of errors and accidents of human drivers. In many contexts, from love to future possibilities, the guidance from following an algorithm may be positive, but quite small.

To me, algorithms are always interesting because they specify reasons for an underlying decision. Sometimes the reasons will expose bias and noise in human decision-making; sometimes the algorithm itself can display bias. But when you know the reasons, you can evaluate the decision more clearly.

For those not familiar with the work of Henry George (1839-1897), he is best-known today for the thesis of a book called Progress and Poverty, which after its original publication in 1879 became a best-seller in the late 1880s and into the 1890s. He argued for a land tax as a practical method of financing government in a way that would also be more fair and efficient.

For a modern take, there is a Henry George Foundation today, and its website offers an overview of a proposal for a land tax in the United Kingdom: “LVT [land value taxation] is an annual, nationally determined, nationally collected, percentage tax, paid by the freeholder, on the open market value of all land with no exceptions.” Notice that a land value tax is not identical to a property tax—which includes both the value land but also the value of the housing or commercial buildings, or land improved in various ways (say, for agriculture or recreation purposes). Thus, a land value tax does not rise when you build something on a given property; conversely, a plot of land with nothing built on it, right next to a similar plot of land with a house or factory built on it, would be taxed the same amount.

In the late 19th century, when George was writing, very large amounts of land in the United Kingdom were owned by those who were noble or rich or both. They could block this land from being developed, and thus limit the ability of towns to expand, for either housing or industry. By taxing what that undeveloped land would sell for on the open market, there would be an incentive to sell off some of that land. Moreover, if the government built, say, a railroad link through a certain area, then the value of the undeveloped land close to the railway would rise–thus providing an even greater incentive to sell off some of the land.

A land tax would work somewhat differently today, of course. But one can imagine a situation where a suburb has very restrictive zoning–say, one house per acre. However, if it was possible to develop that land with, say 16 small houses or an apartment building on that acre, that property could be be taxed on what the underlying land was worth–not just on the value of the single home on the property.

I will not try here to argue the case for and against a land tax in any detail. Instead, I’ll point to an historical episode that illustrates some of its practical difficulties. Samuel Watling tells the story in “The failure of the land value tax” (Works in Progress, Issue 18, March 13, 2025).

In the UK circa 1900 , the national government was primarily funded by an income tax, which paid for the military and the civil service. About one-quarter of the income tax revenue was passed along to local governments, which were responsible for “poverty relief, the police, education, and sanitation.” In what we would today call “unfunded mandates,” the central government has passed laws requiring that the local governments provide certain levels of poverty relief, police, education (for ages 5-12) and sanitation, but without sufficient funding to do so.

Local UK governments of this time had property taxes available to them as an option. A main use of property was renting the land, either for housing or business. Thus, a tax on property was largely a tax on rental income: according to Watling, “three quarters of funding for local government activities – poverty relief, the police, education, and sanitation – came from taxing rental income. Since urban rents added up to about 10 percent of GDP at the time, this meant that one tenth of the economy was responsible for financing almost the entirety of local authority budgets.” As one would expect, a tax on rental income is largely passed along to the renters. For well-to-do cities, these taxes on rental income raised enough money for the public services they were obligated to provide. For poorer cities, with greater needs for welfare spending and less valuable property to tax, the situation was more difficult.

In the predictions of Henry George, a land tax could address these issues, and in a sweeping way. Watling explains:

But these marginal improvements – reducing the disincentive to improve land and providing more funds for urban councils – were only part of the reason Liberal Georgists favored land value taxation. George had promised his followers nothing short of Utopia. George argued that since all production needs land, competition would push labor and capital returns down to minimum levels, leaving all remaining economic surplus to accumulate as land rent. Therefore, he concluded taxing land rent alone could fund all government activities since it captured society’s total surplus value.  Intercepting this entire social surplus with the land value tax, he argued, would provide not just all the money the government needed but enough to end poverty and create a harmonious society in which all humans could fully satisfy their innate needs and desires.

Events happened, as they do. Watling provides details. For my purpose, the key fact is that the Liberal Party ended up enacting a set of land taxes in 1911, and set about the task of placing a value not on a given property–which could be valued based on the rent paid or by comparison with similar properties nearby–but only on the land. For properties that provided revenue via mining, this calcualation was reasonably straightforward. But a land tax based on what the land would be worth, if it was developed more fully or at all, was a harder calculation. Watling again:

There were close to ten million properties in the country that needed valuing, and for the majority of these properties, the land and structure had been traded together, meaning that there was no distinct market valuation of land to draw from. What’s more, in line with Georgist theory, the tax was supposed to credit owners for improvements they made to the land. But this meant calculating several hypotheticals, many of which had never been measured or recorded, including building and structure value and value contributed from plumbing, access to railways, and other infrastructure contributions. The process was beyond the capacity of the government. In August 1910, the Liberals sent out 10.5 million copies of the notorious ‘Form 4’, which required owners to submit specific details on their income and the use and tenure of their properties. It also required them to estimate the site value themselves. Failure to return the document carried a fine of £50, about £7,500 in current prices. 

The pushback was extreme. There were lawsuits galore, until one of the land taxes (there were several) was invalidated entirely. Setting aside the tax complianc costs being imposed on landowners, the costs to government of implementing the tax were considerably greater than the additional revenue raissed. Moreover, although the promise of a land tax was that it would encourage use of underutilized land, it instead imposed immediate costs on underutilized land, so that owners and potential builders of that land lacked funding for construction. Rates of building dropped, rather than rising.

Perhaps all of this could have been worked out over time, but events contineud to happen, as they do–in this case, World War I. By the time the Great War was over, the Liberal party had lost its appetite for pursuing land taxes further, and abolished them in 1922. The difficulties for local governments across the UK to raise funding for their required activities continued–and in some ways continue to the present day. For the land tax in particular, Watling notes:

[T]he pure land value tax is chimerical. Those countries that raise substantial amounts of tax from land, such as Japan and the USA, do so through taxes on property … A pure tax on the unimproved value of land has never been successfully implemented anywhere. Land value taxes introduced in Australia and New Zealand have been repealed. Denmark’s land value tax is a minor quirk of the system, comprising less than two percent of total revenues. 

Henry George was writing in the context of his time and place–the United Kingdom at the tail end of the 1800s. One can argue that a pure land tax could have been a sensible approach in that time and place. In other times and places–and with the property tax as a proven and practical alternative–it’s harder to make the case.

The “multipolar” world has been a reasonably popular framework for thinking about the global economy. The idea was that the world economy was sorting itself into three geographically defined regions: a North American group, a European group, and an East Asian group. Each group of countries had a combination of traits: a sufficient amount of industry and technological leadership in many areas, a mix of skilled and unskilled labor, access to natural resources. In addition, the three regions had the advantages of geographic proximity, and relatively free trade within the area, so that goods and services could flow back and forth with some ease.

For some earlier discussions of this idea here, see “NAFTA in a Multipolar World Economy” (August 11, 2017) and “A North American Vision” (November 5, 2014).

From this perspective, a main concern with President Trump’s threats to set off a trade war with Canada and Mexico is that it threatens to fracture the North American bloc. However, Germany and western Europe will remain at the heart of the European bloc, while a combination of China, Japan, and Korea will remain at the heart of the Asia bloc.

After all, President Trump negotiated and signed the United States-Mexico-Canada Agreement (USMCA) in 2020 during his first term, to address his concerns about the earlier North American Free Trade Agreement. It’s worth quoting some of Trump’s comments during the signing ceremony in 2020:

The USMCA is the largest, fairest, most balanced, and modern trade agreement ever achieved.  There’s never been anything like it.  Other countries are now looking at it, but there can’t be a border like that because, believe it or not, that is by far the biggest border anywhere in the world, in terms of economy, in terms of people.  There’s nothing even close.

This is a colossal victory for our farmers, ranchers, energy workers, factory workers, and American workers in all 50 states … The USMCA is estimated to add another 1.2 percent to our GDP and create countless new American jobs.  It will make our blue-collar boom — which is beyond anybody’s expectation — even bigger, stronger, and more extraordinary, delivering massive gains for the loyal citizens of our nation.

For the first time in American history, we have replaced a disastrous trade deal that rewarded outsourcing with a truly fair and reciprocal trade deal that will keep jobs, wealth, and growth right here in America.  And, in a true sense, it’s also a partnership with Mexico and Canada and ourselves against the world.  It’s really a trade partnership, if you look at it that way.  And it’s a day of great celebration in all three countries.

For some perspective on what can be lost by fracturing the North American trading bloc, the Brookings Institution has published USMCA Forward 2025, a collection of seven essays and additional short comments about some of the positive gains from Trump’s trade agreement. As Joshua P. Meltzer and Brahima Sangafowa Coulibaly point out in their introduction to this volume, U.S. exports to Mexico and Canada have increased by 46% since the USMCA agreement was signed in 2020.

This set of essays focuses on a few key areas of interest for the North American Trade bloc. One is critical minerals. Meltzer and Coulibaly note:

Critical minerals and rare earths are key inputs into the production of many technologies, such as batteries, mobile phones, and semiconductors and needed for defense purposes. … The challenge for North America is the heavy dependence on many of these minerals from China, particularly when it comes to processing. The Trump administration has also made secure supply chains a focus, and this will require addressing the heavy reliance on China for critical minerals. China’s recent announcement that it will restrict exports of various critical minerals to the U.S. in response to U.S. tariffs further underscores the strategic need for the U.S. to reduce this dependency. … [T]he U.S. is 100% reliant on imports of 16 critical minerals such as graphite and more than 50% reliant on imports for another 29 critical minerals, including rare earths, zinc, and nickel. About 40% of U.S. import of critical minerals come from Canada and Mexico. Moreover, the U.S., Canada, and Mexico have largely complimentary resources, meaning that U.S. support for the development of critical minerals and rare earths in Canada and Mexico does not compete with U.S. production but can replace existing dependencies on China. 

In short, if the US is going to be a world leader in technologies like batteries, mobile phones, and semiconductors, it needs easy access to minerals across the North American trade bloc. Moreover, if US manufacturing in these and other areas is to set the standard for global productivity, the US business sector needs to be able to locate different pieces of the goods and services supply chain across the US, Canada, and Mexico in ways that can boost efficiency.

I was pleased back in 2020 by the passage of the USMCA, since it seemed to assuage President Trump’s worries about the NAFTA agreement, while still supporting the North American trading bloc. Import tariffs aimed at China at least have the rationale of being aimed at a country where the US is involved with economic and geopolitical competition. But in Trump’s words from 2020, USMCA is “a partnership with Mexico and Canada and ourselves against the world” and “a truly fair and reciprocal trade deal that will keep jobs, wealth, and growth right here in America.”

Chad Bown interviews Paul Krugman at the Trade Talks podcast on a range of subject related to trade and industrial policy (Trade Talks, March 16, 2025, “Paul Krugman talks trade, industrial policy, and Trump”). Here are a few comments that caught my eye:

If you are worried about dependence of the US economy on foreign supply chains for certain products, the appropriate answer might be industrial policy, but not tariffs.

Max Corden’s 1974 book Trade Policy and Economic Welfare remains relevant. And what Corden and others said was, if there’s something that you think you need to be producing, then encourage production. The answer is industrial policy. The answer is to subsidize or otherwise promote. In general, a tariff has side effects that may not be what you want. If you were worried that too many of the world’s semiconductors are being produced within striking range of China, then you want to subsidize production of high-end semiconductors in the United States. But that’s not a good reason to raise the cost of semiconductors to the U.S. downstream industry. So, there’s a really pretty strong case for industrial policy here. That’s the generic principle.

Now actually implementing it is tricky, by the way. The thing about these agglomeration economies is that, once they’re well established, they’re really hard to break. … And so if you want to develop rival agglomerations to the existing agglomerations that you think are in the wrong place, it’s going to be expensive and hard, which doesn’t mean you shouldn’t do it, but you should realize that it’s not something you do by throwing a few dollars at the problem.

On Europe’s competitiveness problem:

I would suspect that the Europeans would be feeling relatively okay about their economic performance if it weren’t for the comparison with the United States. The old Eurosclerosis of persistent high unemployment is gone. In general, prime age workers are more likely to be working in Europe than they are in the U.S. In a lot of ways, the quality of life is decent. Their life expectancy is several years longer than ours. So Europe looks pretty good, except that they have clearly fallen behind in some advanced technologies and a significant productivity gap has opened up. …

A significant part of that gap in productivity between the U.S. and Europe is really very localized. It’s a reasonable guess that roughly half of the U.S.-European productivity differential reflects very high value-added per worker in Silicon Valley and also Seattle, which operates in somewhat the same way, on one side of the continent, and greater New York on the other. That it’s really the agglomerations of the tech industry in Silicon Valley and the agglomeration of the financial industry, on the East Coast, that are the difference.

On why tariffs aren’t the answer to reducing the US trade deficit:

It’s also probably not the case that tariffs will do much to reduce the trade deficit. There are some subtleties there, but the basic point in textbook economics says that the trade deficit is really determined by the capital account. It is the fact that foreigners want to invest in the United States – so there’s a net inflow of capital – and just as a matter of accounting that means that we have to have a trade deficit on the other side.

If you ask, “So what happens if you put on tariffs?” The answer is, even if other countries don’t retaliate, what happens is that the dollar rises. And we have lower imports, but we also have lower exports because we have a stronger dollar. And of course, if other countries do retaliate we don’t need as strong a dollar. But one way or another, exports fall to pretty much offset the effect on imports. …

Within the range that we’re talking about, tariffs are really unlikely to have an impact on the trade deficit. At the same time, they will raise costs. What’s really striking … was this disproportionate concentration of tariffs on intermediate goods rather than consumer goods, which meant that even manufacturing was probably not benefiting. You were probably actually reducing manufacturing employment. And we’re doing it again. As we’re holding this conversation, the tariffs that have already gone into effect are on steel and aluminum. That’s good for steel and aluminum manufacturer, I guess, and apparently lawn furniture, which for some reason is covered by this as well. But it’s pretty bad for everybody else who’s downstream. These are not tariffs that look like they’re going to achieve even their ostensible goals.

In discussions of how businesses borrow money, there used to be essentially two choice: firms could either issue bonds or borrow through banks. But during the financial crisis of 2008-09, as well as other episodes, it seemed that a number of banks were taking too much risk, which in bad times meant that their solvency was threatened and sometimes emergency action from the Federal Reserve was needed to keep them going. A wave of additional regulations on bank lending followed, to limit risky loans and bolster bank safety.

But as banks pulled back from lending, a number of businesses found it useful to borrow money elsewhere–specifically, via “private credit.” Fernando Avalos, Sebastian Doerr and Gabor Pinter tell the story in “The global drivers of private credit” (BIS Quarterly Review, March 2025, pp. 13-30).  For additional background, the IMF devoted Chapter 2 of its semiannual Global Financial Stability Report in April 2024 to “The Rise and Risks of Private Credit.”

Avalos, Doerr and Pinter write: “Private credit funds have increased their assets under management (AUM) from about $0.2 billion in the early 2000s to over $2,500 billion today.” My personal rule is that quantities measured in trillions of dollars deserve some attention.

“Private credit” is usually set up as an investment fund, where investors put money in and borrowers–typically medium- and small-sized firms–get funding. The funds are “closed end,” meaning that they raise a fixed amount of money and then stop. Some of these funds just do direct lending; some offer more complex loans, which can include provisions for converting the loan into equity ownership in the firm; and some invest in the debt of “distressed” companies, which they can buy at a low price. The firms that borrow through private credit are, as one would expect, firms that aren’t able to borrow what they want through banks or bond markets. Thus, they can often be younger firms that don’t yet have the steady profits that risk-averse bank lenders are looking for.

Avalos, Doerr and Pinter write:

Most funds operate as closed-end structures that lock in capital for their life cycle, which typically ranges from five to eight years. They do not trade on exchanges and are not available to retail investors, which makes them illiquid and subject to lighter regulation. The life cycle of funds usually matches the average maturity of their loan portfolios … Some fund structures, however, offer investors more frequent redemption windows. An important example is BDCs in the United States, many of which list their shares on stock exchanges and are accessible to retail investors. They are subject to federal regulation and have disclosure requirements similar to those of mutual funds, providing transparency and investor protection. With over $300 billion in AUM [assets under management], BDCs represent 20% of the private credit market in the United States today. Attempts to bring retail investors into the fold have been a general trend in the private credit space.

If you aren’t a finance wheeler-dealer, private credit may seem like just another exotic fact about the economy. The lenders in private credit have, to this point, mostly been big firms with long time horizons, like pension funds, insurance companes, and sovereign wealth funds. The private credit funds often specialize in a certain type of industry or firm, and the managers of the fund often have deep knowledge about the industry and firms to which they are lending. In a way, the purpose of stricter bank regulation was to get riskier loans out of the banking system, so it shouldn’t be a surprise when such loans end up being organized in an alternative form.

But of course, financial regulators and international organizations like the International Monetary Fund and the Bank for International Settlements stay awake nights thinking not about how loan arrangements are working just fine in the present, but what the effects would be if such arrangements took a bad turn for the worse in the future–especially as retail-level investors with the ability to zoom in and out of markets become more common in this area.

For example, what if broader economic or financial conditions lead to a much higher risk of default in these funds? As a result, holders of these funds start trying to sell these not-very-liquid investments, and panicky selling drives down the price. Regulated pension funds and insurance companies–even some banks that invest in these funds–see that the value of their investment in these funds is falling. They start to draw on lines of credit and other sources of short-term funding, but with the increased risk and falling prices, those other sources of short-term funding start drying up. Yes, all of this is not at all a likely near-term scenario. But it’s why the IMF wrote last year:

Given the potential risk private credit poses to financial stability, authorities could consider a more proactive supervisory and regulatory approach to this fast-growing, interconnected asset class. … Several jurisdictions have already undertaken initiatives to enhance their regulatory framework in order to more comprehensively address potential systemic risks and challenges related to investor protection. The US Securities and Exchange Commission (SEC) is making substantial efforts to enhance regulatory requirements for private funds, including enhancing their reporting requirements. The European Union has recently amended the Alternative Investment Fund Managers Directive—commonly referred to as AIFMD II—to include enhanced reporting, risk management, and liquidity risk management. … Regulatory authorities in other countries (such as China, India, and the United Kingdom) have also enhanced the regulation and supervision of private funds. With the growth of the private funds sector in general, supervisors have also increased scrutiny over various aspects of private funds, particularly on conflicts, conduct, valuation, and disclosures.

Thus, the race between financial innovation and regulation continues to evolve.

I saw some mention that there has been a sharp rise in federal civilian employment in 2023 and 2024. When I tracked down the numbers at the FRED website run by the Federal Reserve Bank of St. Louis, it looked like this:

Just to be clear, this is civilian employees only, and doesn’t count the 1.3 million or so in the armed forces. However, it does count about 600,000 postal workers, although the US Postal Service is a semi-autonomous agency. The spikes in federal employment every 10 years are temporary employment related to the decennial Census. If you squint a little, you can also see a pattern where government employment tends to rise in times of recession (shaded areas) or immediately after.

However, it also looks as if federal employment had settled at under 2.8 million workers during non-recession, non-Census periods in the late 1990s, the first decade of the 2000s, and from about 2013-2016. From this view, the increase of about 140,000 federal jobs from the start of 2023 to the present does look like a break with past patterns.

I know that the rise in federal employment since 2023 is not about additional post office workers. I’ve seen comments in the press that the higher federal employment relates to implementation of infrastructure and green energy grants. But I confess that I haven’t done the work of tracking the rise in federal employment in the last couple of years back to individual agencies. Someone who wants to spend the time digging around at the at the Office of Personel Management website could do so.

But if you focus on the 2.4 million non-postal but civilian employees, the breakdown across agencies looks like this, according to Drew Desilver at the Pew Research Organization (“What the data says about federal workers,” January 7, 2025). One example of a small-employment agency among the small boxes at the bottom right of the figure would be the US Department of Education, with fewer than 5,000 employees. But truly substantial cuts in federal employment would require truly substantial cuts from the big boxes.

I would not expect federal employment to be a constant share of the US workforce. After all, a substantial part of government work involves working with information, and the leaps and bounds of information technology should make it possible, in a broad sense, to accomplish similar tasks with fewer workers. Indeed, that seems to be the pattern over time. The figure below takes the number of federal employees and divides by total employees in the US economy. Back in the early 1990s, federal employees were almost 3% of the workforce. It’s now about 1.9% of the workforce–basically the same as before the federal employent spikes from the Census and the pandemic recession. Also, when you look at federal employment relative to total employment, the recent jump in federal jobs goes away; in other words, the last two years are a time when federal employment has been rising at about the same rate as total employment, but not faster.

Of course, these kinds of overall numbers don’t offer evidence that certain parts of the US government should have fewer workers, more workers, or the same number. The rise in federal employment in 2023-24 suggests that more attention might be paid to who was hired in what departments. The ratio of federal employment to total employment suggests that we have not seen (yet) seen a radical break with past federal employment patterns.

It is common in current US political discourse to hear it asserted, as an incontrovertible truth, that the US economy is smaller because of the US trade deficit–or equivalently, that tariffs to reduce imports will cause the US economy to grow. Such claims are generally not well-founded. But here, I want to point out one of the arguments for this claim that reflects a more fundamental misunderstanding.

As you will learn from any introductory economic text, there are several ways of measuring the size of an economy, and one of the standard approaches is:

GDP = C + I + G + X – M.

This “national income accounting identity,” as it is sometimes called, is based on the idea that a nation’s economic output can be used in several main ways: it can be consumed (C), it can be invested (I), and it can be part of government consumption (where this term includes only government use of goods and services, not government spending that only represents a pass-through of income to households or firms). The final two terms cover international trade. Some portion of a nation’s economic output can be exported to other countries, but we also need to take imports into account, which wereproduced elsewhere.

This equation is not a “theory” about how the economy works. Instead, an “identity” is a statement that is true by the definition of the terms. This is one of the ways in which GDP is defined. If you go to the website of the Bureau of Economic Analysis and look at a press release for recent estimates of GDP, these are the categories that you see being estimated.

The problem arises when someone takes an accounting identity and believes you can just move the numbers around to achieve a goal. Maurice Obstfeld explains the issues in “Mistaken Identities Make for Bad Trade Policy” (Peterson Institute for International Economics 24-13, October 2024). He writes:

The national income and product (NIP) identity is often the basis of claims that a trade deficit—an excess of import spending over export earnings—causes reduced economic growth and job losses. The identity reflects that a nation’s total production output (gross domestic product, or GDP) must be consumed by households, invested by businesses, purchased by the government, or exported abroad.

GDP = consumption + investment + government purchases + net exports.

The last term on the right is net exports (export receipts minus import expenditures), the balance of trade. It is included because some parts of national consumption, investment, and government purchases are imported from abroad, and these components (which add up to total imports) must therefore be subtracted from the right-hand side above to make the identity a true representation of how GDP is allocated among its possible uses. The preceding relationship is an identity because every product within GDP that is sold on the market is purchased for some use: double-entry bookkeeping.

The claim that trade deficits (negative levels of net exports) cost production and jobs follows immediately from a superficial application of the NIP identity. Suppose net exports fall further, causing the trade deficit to grow, but nothing else on the right-hand side changes. Then the identity implies that GDP must be lower by the same amount. This opens a faulty line of reasoning through which bigger trade deficits are necessarily a drag on output and employment.

Perhaps the key phrase in that explanation is “but nothing else on the right-hand-side changes.” To be more specific, say that imports fall (set aside for the moment why they fall). Say that 100% of that decline in imports is matched by increased domestic output, so that GDP rises. However, “nothing else on the right-hand-side change”–that is, even though domestic production goes up, neither private nor government consumption rises, nor does investment, nor exports. Obstfeld puts it this way:

The prediction that implicitly underlies their calculations, however, is that if imports fall by some amount (for example), an equal amount of consumption or investment demand will automatically be redirected toward domestic products, leaving the sum of total consumption and investment spending unchanged. In terms of the NIP [national income and product] identity, they argue that net exports on the right-hand side will rise without any accompanying changes in the other right-hand side quantities, necessarily leading to higher GDP in precisely the amount of the trade balance improvement.

The flaw in this argument is that the trade deficit rarely if ever changes without some accompanying movement in consumption, investment, or government spending—and the way in which the trade balance interacts with other economic activity depends critically on why it is changing.

Notice the rhetorical shift that often happens here. We started with a statistical definition of GDP, which will always be true, because it’s the definition. It is true that if imports fall, something else in that definition of GDP will have to change, to preserve the identity. But the statement that the change will entirely happen in the form of greater domestic production is a specific theory about what will change–and it’s not at all obvious that the theory is correct. Here are some alternative theories about effects of import tariffs:

If the US imposes tariffs on imported goods, US imports will decline. However, other countries will retaliate with tariffs on US exports, so US exports will decline as well. If these two effects exactly offset one another, so that the lower US imports and lower US exports are the same, the trade deficit does not change and GDP does not change. Instead, there is a dislocation and reallocation in the US economy in which export-oriented industries take a hit, while US production for the US domestic market rises.

Or say that the US imposes tariffs on imported goods, so that US imports decline. This will necessarily mean that foreign producers who used to be selling into the US market are earning fewer US dollars. In the foreign exchange market, the supply of US dollars declines, and the exchange rate of the US dollar rises. As a result, US exports become more expensive in global markes, and exports end up falling as well.

Or say the US imposes tariffs on imported goods, so that US imports decline. Many of those imported goods are used by US firms as inputs to production. The reason US firms import these inputs is that they are either less expensive or higher quality (or both) than the same product would be if produced in US borders (if indeed they are even produced at all within US borders). Thus, all the US firms that have been depending on imported inputs (which includes most large and successful US multinationals) are facing a rise in their costs. As a result, they may decide to cut back on their levels of investment.

Or say the US imposes tariffs on imported goods, so that US imports decline. Many of these imports are purchased by consumers, who choose these items because that they are either less expensive or higher quality (or both) than the same product would be if produced in US borders (if indeed the product is even produced at all within US borders). Thus, when these consumers face the necessity to purchase alternative goods,they will be buying something that they would have preferred less–based either on higher price or a difference in quality. As a result, consumers may decide to cut back.

I want to emphasize two points here.

One is that all of these possibilities, as I have laid them out here, remain all consistent with the basic definition of GDP. The basic definition of GDP does not tell you which of these outcomes are more or less likely–it only tells you how GDP is calculated. The definition of GDP does not tell you that if tariffs are imposed on imports, GDP will rise, or that it will fall, or that it will remain the same. The definition doesn’t tell you whether a change in tariffs will affect exports, or consumption, or investment. It’s just a definition, not a theory of how the economy will react. Anyone who starts with the statistical definition of GDP, and then asserts that lower imports will necessarily lead to equivalently higher domestic production, is pulling a fast one. In Obstfeld’s phrasing, they are assuming that “nothing else on the right-hand-side changes.” It’s a whale of an assumption.

The other point is that to distinguish between possible theories, one needs to look at evidence. Obstfeld goes into considerably more detail about what theories are likely to play out in response to restrictions on imports, and why. For example, the “theory” that other countries will respond to tariffs by retaliating is happening in real time. The “theory” that tariffs on imports lead to a stronger exchange rate, and thus depress sales of exports, has happened in practice. Firms and households do suffer when their access to the imported goods they would prefer to have purchased is restricted.

There are lots of other arguments about import tariffs: I’ve discussed some of them in the past, and am sure to discuss more in the future. But the argument that import tariffs will increase total domestic production, when based on the definition of GDP and the national income and product identity, should be an embarrassment to anyone making it, and it should be ridiculed and laughed down wherever it is encountered.

In the US economy, since the modern pattern of economic growth started back in the early 19th century, average annual growth has been remarkably close to 2% per year on a per capita basis (as I have noted here and here). It would be an extraordinary reversal of fortune for this process to stop and then to reverse itself. But polling data suggests that about 75% of Americans believe that when today’s children grow up, they will be worse off than their parents.

It isn’t just America, either. Across high-income countries of the world, and a number of middle-income countries as well, majorities or near-majorities believe that when children in their country grow up, they will be worse off than their parents. Marta Doroszczyk compiles some polling data for a short article on “Generational Concerns” in the March 2025 issue of Finance & Development from the IMF. Here’s an illustrative figure:

What to make of this?

1) Polling data is rarely simple to interpret. My guess is that when many people are asked about economic prospects for the long-run and the next generation, they have a tendency to react based on medium-run or even short-run concerns–and probably not just economic concerns, either. My guess is that few people have recently looked up the per capita growth statistics before answering.

2) It’s not hard to understand why people in Japan, which has had an economy stuck in slow growth since the early 1990s, or Greece, which has been struggling through one economic crisis after another for a couple of decades now, might be pessimistic about the economic future.

3) But even those types of issues and patterns duly noted, there’s a widespread economic pessimism here, which reaches beyond the particulars of any single country.

4) Back in 2019, the OECD put out a report on what it means to be “middle class.” A central theme in the report was that, across many countries, “middle class” referred to a sense that access to consumption were available in three main areas spending on housing, health care, and higher education. In a lot of countries, those are areas where prices have been rising rapidly.

5) It’s interesting to consider some of the more optimistic countries at the bottom of the figure: India, Bangladesh, Indonesia, Israel, Philippines, Poland.

6) If someone is determined to be pessimistic, it can be hard to talk them out of it. But pessimism affects politics. If we are heading into a time when future generations are actually worse off, we are in a zero-sum or negative-sum economy, in which the only way to benefit some group–or to pursue objectives like enviromental protection–is to cause equivalent losses for other groups. The underlying politics of that setting will be full of bitterness and suspicion. in a US political context, and perhaps a European one as well, it feels to me as if there is room for a politics of optimism and abundance, but it needs to be backed up by actual public and private investments, accomplishments, and observable progress.