That anxiety is understandable. Last summer’s bump up in unemployment among recent college graduates was viewed by some as a sign that AI was beginning to shut young workers out of the labor market.

In my previous post on this topic, I argued that the rise in unemployment for young, college-educated workers looked more like a seasonal spike than a structural break. Graduation season always brings a wave of new entrants into the labor market, and that predictable influx tends to push up unemployment temporarily among 20‑ to 24‑year‑olds with a bachelor’s degree. In other words, given the data available at the time, it looked like “a familiar seasonal pattern, not a white‑collar recession.”

Advances in AI may eventually impact the entry‑level labor market, but it hasn’t happened yet in a way that is visible in aggregate unemployment data. With more data points now available, last summer’s sharp rise in unemployment among recent college graduates looks less like the beginning of a sustained deterioration and more like what it probably was all along: a predictable summer spike that has since largely waned.

Unemployment Has Fallen Sharply for Young Workers

The unemployment rate for 20‑ to 24‑year‑olds rose into late summer and early fall of 2025, which helped fuel commentary about AI beginning to hurt the job prospects of young workers. But with several more months of data now available, that interpretation looks much less convincing.

The unemployment rate for young workers has fallen markedly from its highs last summer, while the overall unemployment rate has remained relatively stable (Figure 1). In fact, the unemployment rate for 20‑ to 24‑year‑olds now sits at its lowest level in more than two years. The gap between the overall unemployment rate and the unemployment rate for 20‑ to 24‑year‑olds has also narrowed considerably since last summer and is now at its smallest level in more than two years.

For students graduating this spring, these trends should be at least mildly reassuring. Based on the aggregate data, the job market has not suddenly become inhospitable to young workers. In fact, the data suggest that young workers are doing better relative to the overall labor market than at any point in the past few years.

Last Summer’s Spike Was Most Likely Seasonal

The more relevant measure, however, is the unemployment rate for recent college graduates. Here too, the new data fit the same pattern: last summer’s spike was most likely just a typical seasonal fluctuation. As Figure 2 shows, the unemployment rate for 20‑ to 24‑year‑olds with a bachelor’s degree rose markedly last summer and then came back down sharply in the fall and winter, just as it has many times before. Last summer’s increase was thus likely just a predictable seasonal pattern rather than a lasting deterioration in the job market for recent college grads.

That does not make last summer’s spike in unemployment meaningless. For those who graduated into what was a particularly congested labor market, it might have felt very real, like something permanent had changed. It is hard to maintain objectivity when you are the one living through it: sending out countless applications, waiting for responses, and watching friends land jobs. Starting a career is stressful even in good times, and headlines filled with stories of AI-related layoffs likely amplified that stress. But based on the aggregate data, it looks like the labor market for new college grads is largely behaving as expected.

The comparison with similarly aged workers without a college degree is also revealing. In my earlier piece, I pointed out that the unemployment rate for recent college graduates can sometimes rise above that of similarly aged workers without college degrees during the summer months, precisely because so many graduates are entering the labor force at once. But that reversal tends to be brief: by the fall, college graduates usually have a lower unemployment rate than their less‑educated peers.

As we can see in Figure 2, that same pattern played out again last year. As of this spring, 20‑ to 24‑year‑olds with a bachelor’s degree once again have a meaningfully lower unemployment rate than similarly aged workers without college degrees. If the labor market were becoming hostile to recent college graduates, this is not what one would expect to see.

The Gender Split Looks Less Alarming Too

As I argued in my earlier piece, last summer’s spike in unemployment among recent college graduates seemed to be driven more by men than by women. While male unemployment in this age and education group is generally noisier and tends to be higher than female unemployment, both series have declined markedly from their highs last summer, as shown in Figure 3.

Thus, even for the subgroup that was most affected by the rise in unemployment (male, recent college grads), the data do not point to mounting distress. If anything, they suggest the opposite: the worst of the unemployment spike has faded, and the gender gap has narrowed considerably.

Putting It All Together

If AI is already taking a meaningful bite out of job opportunities for new college graduates, it is not yet visible in aggregate unemployment data. The job market for recent college graduates appears to be holding up reasonably well. Last summer’s spike in unemployment appears to have been a predictable seasonal fluctuation, not the beginning of a secular deterioration in the job market for entry-level workers.

For the Class of 2026, this does not guarantee a smooth transition into the job market. Some will still struggle to land their first job, and for them the search may feel every bit as discouraging as the headlines suggest. But based on the aggregate data available today, while it is possible that AI may eventually impact the employment prospects of new grads, it is not visibly weakening the labor market for young workers in the way some have argued. If anything, the data suggest that job prospects for the Class of 2026 may be somewhat better than they were for the cohorts that preceded them.

More broadly, this is a reminder that economic narratives are often built on a handful of observations that happen to arrive at just the right time. Plausible stories about how AI may affect us in the future are not the same thing as evidence that it is impacting us today. And when fluctuations in the data are highly seasonal, as is the case for the unemployment rate of recent college grads, this piece demonstrates that a few more observations can make the difference between accepting a compelling narrative and getting an accurate interpretation of the data.

Share

If you enjoyed this mildly efficient and occasionally rational update on the labor market for recent college grads, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

That observation raises a natural follow-up question. If recent advances in AI lead to a sustained increase in productivity, and if that translates into faster real GDP growth, could that by itself materially improve the U.S. fiscal outlook?

To a point, yes. If economic growth were to rise while everything else remained unchanged, then growth in the U.S. debt-to-GDP ratio would slow, potentially stabilizing the fiscal outlook without the need for spending cuts.

But while it is reasonable to think that faster growth need not directly affect the primary deficit, basic economic theory suggests that the real interest rate and the growth rate of the economy are not independent variables. If AI meaningfully raises productivity growth, it should also raise expected returns to investment, which in turn affects the equilibrium price of capital. Once that mechanism is taken into account, the extent to which faster growth improves debt dynamics is likely to be smaller than the most optimistic versions of the AI story imply.

The central question, then, is not whether recent advances in AI will increase economic growth. It is whether economic growth will accelerate by more than the real interest rate on government debt rises.

A Brief Recap of Debt Dynamics

In my last piece, I showed that debt sustainability does not depend on the rate of economic growth alone, but on the gap between the rate of economic growth and the real interest rate on government debt. This can be seen by examining the following equation which characterizes the change in the debt-to-GDP ratio, Δ(D/Y):1

where 𝑟 is the real interest rate on government debt, 𝑔 is the growth rate of real GDP, and PD/Y is the primary deficit-to-GDP ratio.

This equation implies that the debt-to-GDP ratio can be stabilized so long as the primary deficit is not too large and 𝑔 > 𝑟. The intuition behind this result is that when economic growth exceeds the real interest rate, the economy’s taxable capacity is rising faster than interest is accruing on government debt, making any given level of government debt easier to manage.

In particular, if real GDP growth runs at 3 percent and the real interest rate on government debt is 2 percent, stabilizing today’s debt-to-GDP ratio of around 100 percent would require cutting the primary deficit roughly in half from its current level of around 2 percent of GDP to just 1 percent of GDP.

By contrast, if growth were to accelerate to 4 percent, holding 𝑟 fixed, the debt-to-GDP ratio could be stabilized at around 100 percent without the need for any meaningful fiscal tightening. In this sense, faster growth appears to offer a way out of an otherwise difficult fiscal situation.

This is what makes the prospect of AI-driven growth so appealing to those concerned about the U.S. fiscal outlook. If AI can push the rate of real economic growth up by just a single percentage point, the U.S. debt trajectory becomes considerably more sustainable.

This conclusion assumes that the difference between 𝑔 and 𝑟 increases as economic growth accelerates. In other words, it is based on an increase in 𝑔, ceteris paribus. What happens if 𝑟 changes as well in response to an acceleration in the rate of economic growth?

An AI Productivity Boom Will Likely Raise Interest Rates

A useful way to think about the link between productivity growth and real interest rates is via the marginal product of capital.2 In a standard neoclassical model, the real interest rate is tied to the marginal product of capital (MPK) net of depreciation (δ):3

An increase in productivity raises the marginal product of capital, which in turn puts upward pressure on real interest rates throughout the economy.

The basic idea is that if capital markets are competitive, funds will flow to those projects which offer investors the highest risk-adjusted return on investment. An increase in productivity in one part of the economy will tend to attract investment flows, which puts upward pressure on required returns elsewhere as firms compete for capital. Thus, if AI raises productivity growth in a way that increases the marginal product of capital, it will likely put upward pressure on real interest rates throughout the economy, including on U.S. government debt.

This does not mean that the real yield on Treasury debt will move one-for-one with the return on private capital. Government debt is a safe and liquid asset, and its yield reflects more than just the marginal product of capital. Still, it would be surprising if the government’s cost of borrowing were completely insulated from a broad-based increase in real returns elsewhere throughout the economy.

We may already be seeing some version of this mechanism in the rush to build data centers, chip capacity, and power infrastructure tied to the AI buildout.

What This Implies about Debt Sustainability

An AI-driven productivity boom is likely to improve the fiscal outlook by accelerating long-run economic growth. A larger and faster-growing economy can support a larger stock of debt, and that was one of the main points of my earlier piece.

The key question, however, is not whether an AI productivity boom will increase the rate of economic growth. It is whether it raises growth by more than it raises the real interest rate on government debt.

The answer will depend on how large the productivity gains turn out to be, how investment‑intensive the AI transition is, how much of the increase in private returns feeds through into the real yield on Treasuries, and how monetary and fiscal policy respond along the way.

What economic theory suggests, at a minimum, is that it is too simplistic to treat faster productivity growth as a silver bullet for the debt problem. An AI productivity boom may well help slow the rise in the U.S. debt‑to‑GDP ratio, but the equilibrium response of real interest rates will likely limit how much it can do on its own.

That does not make the optimistic AI story wrong; it just makes it more nuanced than some recent commentary suggests.

Share

If you enjoyed this mildly efficient and occasionally rational take on whether an AI-driven productivity boom can solve the U.S. debt problem, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose.

In this piece, I’ll first review the data on debt and deficits in the U.S. Then I’ll explain how economists think about the sustainability of government debt in general before applying this framework to the U.S. fiscal situation today. As I show below, the sustainability of U.S. government debt ultimately depends on how economic growth, real interest rates, and fiscal deficits evolve over time. Under plausible assumptions about economic growth and real interest rates, stabilizing today’s debt-to-GDP ratio would require reducing the primary deficit by roughly $325 billion per year.

Rapidly Rising Debt and Deficits

Figure 1 shows the evolution of U.S. government debt held by the public relative to the size of the economy since 1970.

After declining during the late 1990s and early 2000s, the debt-to-GDP ratio rose sharply during the Global Financial Crisis (GFC) as collapsing tax revenues and large fiscal stimulus programs pushed deficits sharply higher. Debt shot higher again during the pandemic when emergency relief programs and the economic shutdown drove deficits to nearly 15 percent of GDP. As a result, U.S. federal debt held by the public now stands near its highest level in modern U.S. history at roughly 98 percent of GDP and is projected by the Congressional Budget Office (CBO) to reach nearly 120 percent by 2036.1

The rise in federal debt is being driven by persistent and increasing deficits. As shown in Figure 2, the federal deficit is currently running at nearly 6% of GDP. These are the largest federal deficits the U.S. has run since 1970 outside of the GFC and the COVID-19 pandemic. And there appears to be little relief in sight as the CBO expects deficits to remain at or above this level throughout the coming decade.2

What is driving these persistent deficits?

Figure 3 shows U.S. government net interest payments as a percentage of GDP since 1970. Rising debt levels combined with rising interest rates in the wake of the COVID-19 pandemic have significantly increased the total cost of servicing U.S. government debt. For much of the past decade, historically low interest rates kept federal interest payments relatively contained despite the growing stock of debt. That environment has changed. As older, low‑rate debt matures and is replaced with increased borrowing at higher rates, interest costs relative to the total size of the economy are increasing rapidly and are now at their highest levels since the mid-1990s. To put things in context, net interest payments of 3.2 percent of GDP now represent more than half of the total federal deficit, which currently stands at 5.8 percent of GDP.

Debt, deficits, and interest payments are all at or near historic highs relative to the size of the economy. How could this possibly be sustainable?

To answer this question, we need a framework for thinking about the dynamics of government debt.

A Framework for Debt Dynamics

Let’s start with a basic accounting identity that describes how government debt evolves over time:

where the primary deficit is simply the federal deficit excluding interest payments on existing debt, that is, the gap between non‑interest spending and revenues:

When non-interest spending exceeds revenue, the government is running a primary deficit.

There are two important implications of Equations (1) and (2). First, the only way to keep the level of debt from rising is if the government runs a primary surplus that is equal to or greater than its interest payments. Second, and perhaps more obvious, if the government runs a primary deficit, it must borrow to finance that gap, which increases the stock of debt. For a country like the U.S., which has run primary deficits for much of recent history, Equations (1) and (2) suggest that the level of debt will likely continue to rise.

But rising debt levels do not necessarily imply an unsustainable fiscal situation. To evaluate sustainability we first need some measure of “ability to repay debt.” While GDP is not literally “income” in the way wages are for a household, it is the broadest measure of the economy’s taxable capacity. And just as high income households can afford to carry a larger debt burden than low income households, a larger economy can support a larger stock of debt because the government has a larger tax base from which it can raise revenue to service its debt.

To incorporate this idea into our framework, we simply divide Equation (1) through by the level of GDP:

Now we have terms that are close to what we want: debt, interest payments, and the primary deficit all relative to the size of the economy (almost, as there is a timing mismatch in the first term!).

After some painstaking algebra and a few reasonable approximations, we arrive at the following equation which characterizes how the debt-to-GDP ratio changes from one year to the next:3

where r is the real interest rate on government debt, g is the growth rate of real GDP, and PD/Y is the primary deficit as a share of GDP.4 Equation (4) captures the key forces that determine whether the debt-to-GDP ratio rises or falls over time, and we will use it shortly to evaluate the current outlook for U.S. government debt.

There are two implications of Equation (4) that I want to point out. First, all else constant, primary deficits push the debt-to-GDP ratio higher, while primary surpluses push it lower. Second, the real interest rate (r) and rate of real economic growth (g) play crucial roles as well. In fact, Equation (4) tells us that if the rate of real economic growth exceeds the real interest rate on government debt r, the debt-to-GDP ratio can fall even if the government is running a (modest) primary deficit.

The intuition for this result is as follows: If the economy is growing faster than the real interest rate (i.e., g > r), then the government’s “ability to repay debt” is rising faster than interest is accruing. In that environment, the government can afford to run modest primary deficits while maintaining a stable, slowly rising, or even falling debt‑to‑GDP ratio.

Conversely, if the economy is growing slower than the real interest rate (i.e., g < r), then the opposite occurs: interest on the debt compounds at a faster rate than the government’s capacity to service its debt, and the debt-to-GDP ratio will rise. That is, unless the government implements austerity measures sufficiently large to generate a primary surplus.

Now onto the Sustainability of U.S. Government Debt

We are now in a position to apply the debt dynamics framework we have developed, Equation (4) in particular, to evaluate under what conditions the U.S. government debt is on a sustainable trajectory.

To proceed, we first need estimates for g, r, and D/Y.

Over more than two centuries of U.S. data, real GDP growth has averaged roughly 3 percent per year, despite substantial fluctuations across decades. So a reasonable estimate for the long-run rate of real economic growth g is about 3 percent.

The real interest rate on government debt is determined in financial markets. If the neutral nominal policy rate is roughly 4 percent, and all indications are that this is about right in today’s environment, and inflation stabilizes near the Federal Reserve’s 2 percent target, a reasonable long‑run estimate of the real interest rate r is about 2 percent.

Together, our estimates for g and r imply a growth-interest differential of roughly 1 percentage point (that is, g − r ≈ 3% - 2% = 1%). With the rate of real economic growth greater than the real interest on government debt, the federal government can afford to run modest primary deficits and still maintain control over its outstanding debt obligation. Thus, we are in a relatively forgiving environment for debt dynamics.

The key question, however, is how large of a primary deficit can the federal government run without causing a further increase in the debt-to-GDP ratio?

Setting Δ(D/Y) equal to zero in Equation (4), the primary deficit that would leave the debt-to-GDP ratio unchanged is:

Given that the debt-to-GDP ratio for the U.S. currently stands at around 98 percent, Equation (5) implies that:

Thus, our debt dynamics framework implies that the primary deficit consistent with a stable debt-to-GDP ratio for the U.S. is about 1 percent of GDP.

How does this compare to the actual primary deficit in recent years and that projected going forward? Figure 4 plots the primary deficit-to-GDP ratio for the U.S. since 1970, along with the CBO’s projections from 2026 through 2036.

Since the GFC, the primary deficit has exceeded the 1 percent threshold needed to keep the debt-to-GDP ratio from increasing. Looking forward, the CBO projects a primary deficit of ~2 percent of GDP in 2026. Stabilizing the debt-to-GDP ratio would thus require reducing the primary deficit by about 1 percent of GDP, or roughly $325 billion per year. This amounts to a reduction of roughly 6% of annual non-interest spending given that total annual non-interest spending is on the order of about $5.5 trillion per year.

Sensitivity to the Rate of Real Economic Growth

My estimates of the maximum primary deficit-to-GDP ratio needed to stabilize the debt-to-GDP ratio going forward depend heavily on the rate of real long‑run economic growth. To illustrate this sensitivity, Figure 5 plots the maximum primary deficit‑to‑GDP ratio consistent with a stable debt‑to‑GDP ratio as we vary the assumed long‑run real growth rate while holding r fixed at 2 percent.

The faster the economy grows, the higher the primary deficit-to-GDP ratio can be while maintaining a constant debt-to-GDP ratio. When debt levels are near 100 percent of GDP, each additional percentage point of economic growth allows the government to sustain roughly one additional percentage point of primary deficit without increasing the debt-to-GDP ratio.

It follows that the debt-to-GDP ratio could be stabilized at its current value given the current primary deficit-to-GDP ratio of about 2 percent if real economic growth going forward averages around 4 percent per year. While not impossible, this relatively rapid real rate of economic growth would exceed the historical long-run average of 3 percent by a considerable margin.

So Is U.S. Government Debt on a Sustainable Trajectory?

Our framework implies that whether the current trajectory of U.S. government debt is sustainable ultimately depends on the relationship between economic growth, interest rates, and the primary deficit.

If real growth averages somewhere close to its historical pace of roughly 3 percent, debt stabilization would require meaningful fiscal adjustment.

If growth were to rise toward 4 percent, however, the current debt-to-GDP ratio could be stabilized with little to no adjustments in fiscal policy.

The debate over fiscal sustainability is therefore not primarily about whether today’s debt level is “too high.” It is about whether economic growth continues to outpace real interest rates, and whether policymakers can keep primary deficits within the limits implied by the arithmetic of debt dynamics.

Share

If you enjoyed this mildly efficient and occasionally rational take on whether U.S. government debt is on a sustainable path, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

TL;DR: Everyone knows what they’re supposed to do in a drawdown: stay invested. So why is it so hard to follow the plan when markets fall quickly? The true constraint is your hold‑through risk tolerance—how much volatility you can endure without panic selling. Portfolios should be built around this, not for how bold you feel in calm markets.

During a downturn in the market, red on the ticker tape becomes hard to ignore. Positions that have already fallen sharply begin to feel intolerable. Selling feels rational. And re-entering the market often happens only after the recovery is well underway.

We saw a version of this last April. The selloff around Liberation Day was sharp and unsettling. When markets took a leg down, the emotional response was immediate. Investors who had calmly described themselves as focused on the long term suddenly reconsidered their equity exposure. Those who reduced risk often found themselves re-entering after much of the rebound had already occurred.

The puzzle is not knowing what to do. Most investors understand that volatility is the price of admission and that staying the course is usually the correct response to a downturn.

The real question is behavioral: why is it so hard to act rationally in the moment?

After laying out why staying the course is so difficult in a drawdown, I will use the recent SaaS selloff to illustrate how hard it is to distinguish temporary repricing from real impairment. Differentiating between the two is one of the most challenging parts about long-term investing.

Why Execution Breaks Down

We know from behavioral economics that losses hurt more than equivalent gains improve our wellbeing.1 A 30 percent drop does not feel like the mirror image of a 30 percent gain.

We also tend to evaluate outcomes relative to a recent peak.2 Suppose you invested $500,000 and your portfolio grew to $1 million. Then, after Liberation Day in April 2025, the value of your portfolio drops to $700,000. You don’t feel like someone who has gained 40%. You feel like someone who has lost 30%.

These two features — loss aversion and reference dependence — explain why drawdowns feel so different from the gains that precede them. Volatility is abstract in a rising market. In a falling one, it becomes personal. Once we enter the loss domain, avoiding further losses becomes more salient than maximizing expected return.

But psychology alone does not explain the full gap between knowing what to do and actually staying the course when markets are in freefall.

Risk Tolerance Is Not Fixed

Most questionnaires measure how much risk you think you can take; what matters is what I call your hold‑through risk tolerance—how much volatility you can endure without changing course.

We often treat risk tolerance as a stable trait. Investors are labeled conservative, moderate, or aggressive, as if they are permanent characteristics.

Yet research shows that measured risk tolerance declines during downturns.3 Risk aversion rises as markets fall.

The tolerance you believe you have in calm periods may not match the tolerance you reveal when markets are falling. Preferences shift with the state of the market.

The Economic Consequence of Panic Selling

Pulling the rip cord in a downturn can have significant economic consequences.

Selling after a 30 percent drawdown feels prudent. It feels like you are protecting yourself. But sharp declines have historically been followed by stronger subsequent long-run returns, as valuations fall and the expected compensation for bearing risk rises.4

This creates an uncomfortable asymmetry. The moment you feel the need to sell the most is precisely when expected returns tend to be at their highest. Loss aversion, reference dependence, and rising risk aversion can push investors to panic sell precisely when the risk-return tradeoff is most attractive.

The right answer is simple in theory. Stay invested. But doing so is far easier said than done.

If Execution Is the Weak Link, Design Around It

If behavior under stress is the constraint, portfolios should be designed around it. A strategy that maximizes expected return but cannot be held through the inevitable drawdown is not a strategy at all.

First, equity allocations should be sized such that a 30 percent drawdown is painful but not catastrophic. Investments that trade some upside for downside protection should also be on the table (e.g., options-based volatility-smoothing ETFs and mutual funds). By reducing overall volatility, both approaches aim to increase the probability that an investor stays invested.

Another thing to consider is how often we monitor our investments. There is a large literature on “myopic loss aversion” which shows that investors who review their portfolio more frequently perceive risk as higher and allocate less to equities.5

Investors who monitor their portfolios continuously also tend to trade more frequently, and research shows that this leads, on average, to underperformance relative to those who trade less. This is in part because those who trade more often tend to react to short-term price movements rather than genuine changes in fundamentals.6

Reducing how often you check your portfolio changes the psychological experience of volatility. Fewer evaluation points mean fewer perceived losses, which in turn means less often feeling the need to reduce exposure to equities. For many long-term investors, quarterly or semiannual check‑ins are enough.

More generally, anything that increases an investor’s hold‑through risk tolerance is an improvement over cookie-cutter portfolio designs and unchecked monitoring habits.

While each of these strategies is prudent, and staying invested is usually the correct approach, an important qualification is necessary.

When Selling May Be Rational

Not every decline is an innocuous case of Mr. Market feeling depressed and offering to sell otherwise fantastic securities at a discount. Sometimes a sharp decline in the market happens in response to rational fears about a genuine deterioration in fundamentals.

In the early days of the pandemic, for example, it was entirely plausible that the global economy might remain shut down for an extended period, leading to a worldwide recession. During the Liberation Day episode, the tariffs, as announced in the Rose Garden, had the potential to disrupt global trade and tip the U.S. economy into recession.

In such moments, investors are right to reassess current valuations and their long-run cash flow projections. If fundamentals have truly deteriorated, lower prices do not imply higher expected returns. And selling may be an entirely rational response.

The difficulty lies in separating these two possibilities in real time. Is this a temporary repricing of risk, or a permanent impairment of fundamentals? Markets fall in both cases. One creates opportunity. The other justifies caution.

Distinguishing between these two possibilities while losses are mounting is exceedingly difficult.

This uncertainty has again taken center stage in the recent SaaS meltdown.

The SaaS Meltdown as a Case Study

A number of publicly-traded software firms have fallen sharply from their recent peaks in late-December 2025 or January 2026. For example, Microsoft is down roughly 17%, Adobe about 28%, Salesforce around 30%, ServiceNow nearly 40%, and LegalZoom more than 30%. In many cases, years of gains have been erased.

At this point, investors are facing a critical question: Do rapid advancements in AI represent a real threat to the SaaS business model, or is this the investment opportunity of a lifetime?

For a company like Microsoft with a wide moat, diverse revenue streams, and a rapidly growing cloud infrastructure business, this feels more like an opportunity than a crisis.

For firms like LegalZoom, on the other hand, whose value proposition may soon be replaced by AI, impairment of the business model is a real possibility.

Conclusion

We only learn whether it was a true impairment or simply a temporary repricing with hindsight. In the moment, most drawdowns feel ambiguous. Expected returns may be improving, or fundamentals may be deteriorating, or both.

Even if we knew it was the former, staying the course when prices are falling and uncertainty is greatest may be the hardest part of being a successful long‑term investor.

Share

If you enjoyed this mildly efficient and occasionally rational take on why staying invested during a drawdown is so hard and what can be done about it, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

TL;DR: The recent surge in gold prices is best explained by a sharp rise in investor demand, visible in ETF and bar flows, rather than dollar debasement, de-dollarization, falling real rates, or central bank purchases.

Introduction

Gold prices have surged from about $2,050 per troy ounce in early 2024 to just above $5,000 this month, a gain of more than 140% (Figure 1).

Market observers have offered a range of explanations for this move, from dollar debasement and de-dollarization to falling real rates, central bank buying, and a surge in investor demand.

The goal of this piece is to interrogate the extent to which each of these potential explanations is (or is not) consistent with the facts. What I will argue is that the evidence points squarely to investment flows as the key driver, not dollar debasement, de-dollarization, falling real rates, or central bank purchases.

Let’s examine the facts.

1) Dollar debasement

If gold prices were rising primarily because investors expect a sustained loss of the dollar’s purchasing power, that expectation should appear in market-based inflation expectations.

It does not.

In fact, U.S. Treasury–TIPS breakeven inflation and forward inflation compensation have been remarkably stable over the last three years, remaining rangebound between about 1.8% and 2.7%, even as gold prices accelerated higher (Figure 2).1

If debasement were the dominant explanation, it would be difficult to reconcile the gold rally with stable inflation expectations. On this metric alone, dollar debasement is unlikely to be a primary driver of the recent surge in the price of gold.

Verdict: Not supported by market-based inflation expectations.

2) De-dollarization

A more defensible explanation is that gold is responding to a weaker dollar, rather than rising inflation expectations.

Because gold is priced in dollars, a weakening dollar mechanically raises the dollar price of gold, even if its real value remains unchanged. More broadly, a weakening dollar can induce portfolio rebalancing away from dollar-denominated assets without implying an imminent loss of reserve-currency status, or de-dollarization.

The question is not whether this channel is plausible, but how much it actually explains.

Figure 3 depicts the U.S. Dollar Index (DXY), a benchmark that measures the value of the U.S. dollar against a basket of six major foreign currencies: the Euro (EUR), Japanese Yen (JPY), British Pound (GBP), Canadian Dollar (CAD), Swedish Krona (SEK), and Swiss Franc (CHF). After a brief period of increased strength in late 2022 as the Federal Reserve hiked the Federal Funds Rate, the U.S. Dollar Index declined by about 14% between late 2022 and early 2026.

If we zoom in on the most recent two-year period during which the nominal gold price increased by more than 140%, the U.S. Dollar Index declined modestly from roughly 104 to around 97. A simple accounting decomposition shows that only about 7 percent of the rise in the price of gold over this period can be attributed to weakening of the dollar.2 More than 90 percent reflects an increase in the real price of gold.

So while dollar weakening matters, it is quantitatively insufficient to explain the surge in gold prices we have observed over the last two-year period. The rally is not primarily a mechanical FX-driven price effect. Something else must be going on.

One possibility is that expectations of future de-dollarization are influencing gold prices today. We would expect this to show up as increased investor demand (more on this below).

Verdict: Dollar weakening matters, but mechanically it explains only a small slice of the move higher in gold prices.

3) Real interest rates

Gold is often described as sensitive to real interest rates, which determine the opportunity cost of holding a non-yielding asset. Historically, the biggest gold rallies have come when real yields fall sharply, because the opportunity cost of holding gold collapses.

That is not what has happened in this episode.

There has been no sustained decline in real interest rates over the period in which gold prices accelerated. Indeed, since late 2023, 10-year U.S. TIPS yields, a proxy for real yields, have remained range-bound between roughly 1.6% and 2.4% (Figure 4). Real rates therefore do little to explain the timing or magnitude of the rally.

If anything, this episode is unusual precisely because gold has risen sharply in price without a corresponding decline in real yields.

Verdict: Real yields have held steady, making them an unlikely culprit.

4) Central bank buying

Central bank buying is often cited as a key driver of the recent surge in gold prices. Central banks, particularly those in developing countries or in countries which have close ties to Russia or China, fear U.S seizure of their dollar-denominated assets (e.g., U.S. Treasuries), so they have shifted instead to an asset that cannot be easily confiscated by the U.S. government. A shift toward physical gold might also be driven by fear of continued weakening of the U.S. dollar and therefore depreciation of dollar denominated assets like U.S. Treasuries.

Consistent with this story, central banks have indeed been net buyers of physical gold in recent years. But if central bank stockpiling of physical gold were the primary driver of the recent run-up in gold prices, we would have expected to see central bank net purchases accelerate during 2025.

They did not.

In fact, net central bank purchases declined by roughly 20 percent in 2025 compared to 2024, precisely when the rise in gold prices accelerated (Figure 5).3

Central banks clearly matter for the long-run level of gold demand, but they are unlikely to be the marginal driver of the recent surge in gold prices.

Verdict: Important in levels, not at the margin in 2025.

5) Investor demand

Figure 5 suggests that, rather than central bank stockpiling, a rapid increase in investor demand is the most plausible driver of the recent surge in gold prices. Between 2024 and 2025, net investment in physical gold increased by nearly 1,000 tons, representing a year-over-year increase of roughly 84%.

Figure 6 breaks down gold demand within the investment sector in both 2024 and 2025. Net investment in gold-backed ETFs increased sharply between 2024 and 2025, from approximately 0 tons of net investment in 2024 to more than 800 tons in 2025.

Production of gold was unable to keep up with this rise in net investment. Between 2024 and 2025, net gold production only increased by about 40 tons to roughly 5,000 tons. This means that a full 16% of all gold production last year was absorbed by flows into gold-back ETFs.

Investment in physical gold bars also increased modestly, from about 860 tons in 2024 to roughly 1,070 tons in 2025, meaning that 21% of all gold production last year went to investment in gold bars.

Given relatively stable net gold production, this dramatic rise in investment demand crowded out both central bank purchases and the use of gold for jewelry fabrication.

In liquid markets like this, prices are set at the margin, and ETF flows represent the most flexible and scalable source of marginal demand. It follows that increased investment demand for gold, primarily via ETFs but also through physical gold bars, is the most likely explanation for the recent surge in the price of gold.

Verdict: The only channel that actually moved enough to match the price action.

Conclusion

The recent surge in gold prices does not appear to reflect fears of imminent dollar debasement. Market-based inflation expectations remain well anchored, and real interest rates have been broadly stable throughout the period in which gold prices accelerated higher.

Nor is the rally primarily an exchange-rate story. Since February 2024, only about 7 percent of the increase in gold prices can be mechanically attributed to dollar depreciation. Central bank purchases, meanwhile, slowed in 2025 as gold prices continued to rise, making them unlikely to be the marginal driver of the move.

Taken together, these facts leave little room for macro-driven explanations. Gold is not responding to rising inflation expectations, falling real yields, or an abrupt breakdown in the dollar as the world’s reserve currency.

What the data do point to instead is a pronounced increase in investor demand, visible most clearly in flows into gold-backed ETFs and, to a lesser extent, physical gold bars. Put simply, ETFs and bar demand were large enough, and flexible enough, to absorb a big share of global supply, and prices adjusted accordingly.

What remains difficult to pin down is why investor demand has risen so sharply. The data do not allow us to cleanly distinguish between several overlapping motivations: hedging against the possibility of future de-dollarization, insuring against broader geopolitical or financial risks, or more purely speculative behavior, including momentum and trend-following strategies. All of these channels would manifest in similar ways—through increased investment flows—making them observationally hard to separate.

If you enjoyed this mildly efficient and occasionally rational take on the recent surge in the price of gold, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

2

Because gold is priced in dollars, its dollar price can be written as

Taking log differences between February 2024 and February 2026 yields

The fraction of the gold price increase mechanically attributable to dollar depreciation is therefore

The remainder reflects an increase in the real price of gold rather than weakening of the dollar.

TL;DR

Hyperscalers can keep expanding AI infrastructure, but only with shrinking financial slack. All are increasingly relying on long-dated leases that push the funding constraint into the future. At Oracle, operating cash flow no longer covers AI investment, making this shift a necessity rather than a choice. Nvidia’s valuation assumes AI demand grows fast enough to sustain a capital-intensive, lease-backed expansion.

Introduction

Over the past several years, the hyperscalers, including Microsoft, Alphabet, Meta, and Oracle, have embarked on one of the most aggressive infrastructure buildouts in modern corporate history. Collectively, these firms are now investing well over $200 billion per year in new data centers equipped with advanced semiconductors, networking equipment, and even on-site power generation to meet current and anticipated demand for AI model training and inference.

For Nvidia, this spending wave is not merely a tailwind for revenue growth. Its current valuation depends critically on the ability and willingness of the hyperscalers to continue expanding AI infrastructure investment for many years to come.

Nvidia itself has been explicit about how concentrated this demand is. In its most recent quarterly filing, the company notes that “our revenue is concentrated among a limited number of direct, indirect and cloud service purchasers,” and discloses that four direct customers each accounted for more than 10% of quarterly revenue.

This raises a fundamental question. Is the current pace of AI infrastructure investment by the hyperscalers financially sustainable? And if not, what does that imply for Nvidia’s current valuation?

Answering this question requires looking beyond headline financial metrics and focusing on how investments in AI infrastructure are actually being funded.

Measuring sustainability

Periods of rapid capital accumulation tend to distort conventional financial metrics, particularly headline free cash flow. Free cash flow often compresses at times like this as large investment outlays occur now while the cash flows they are expected to generate arrive much later. In such cases, free cash flow blends operating performance, investment intensity, and financing choices, making it a poor guide to whether current investment levels are financially sustainable.

Changes in cash balances are similarly uninformative in this environment, since firms can preserve cash by shifting investment into leases, reducing share buybacks, or raising funds through debt and other external financing, even as underlying funding slack erodes.

To evaluate sustainability, I instead focus on how much internally generated cash remains once AI infrastructure investment has been accounted for, and how quickly that margin is shrinking. This is summarized by the funding coverage ratio, which I define as follows:

Funding coverage ratio = Adjusted operating cash flow ÷ AI infrastructure investment

AI infrastructure investment is measured as capital expenditures plus finance lease principal payments, treating owned and leased assets symmetrically. Adjusted operating cash flow (Adjusted CFO) is cash flow from operations net of stock-based compensation, which removes equity-financed labor costs and isolates the cash actually available to fund investment.

This decomposition isolates the core sustainability question of the AI capex boom: how much operating cash remains after funding infrastructure, and how quickly it is being exhausted as investment continues to scale.

A funding coverage ratio above one indicates that a firm is able to fund investment internally out of operating cash flow. A ratio below one implies reliance on external financing through debt, leases, or other balance-sheet commitments. But even if the ratio is above one, a sustained decline signals diminishing financial slack as investment absorbs an increasing share of internally generated cash.

Figure 1 depicts the funding coverage ratio since late 2021 for Microsoft, Alphabet, Meta, and Oracle.

The broad pattern is clear. Microsoft, Alphabet, and Meta are currently able to fund investment internally, but they are doing so with materially less slack than just a few years ago. The AI infrastructure buildout has increasingly become the dominant use of internally generated funds.

Oracle’s funding coverage ratio is notably different. As of May 2025, it fell below one and has continued to decline, indicating that the company can no longer fund AI infrastructure investments internally. For Oracle, continuing to invest in AI infrastructure during the back half of 2025 required turning to external financing, both via debt markets and long-dated lease commitments.

While Microsoft, Alphabet, and Meta remain above the break-even threshold, their steady convergence toward it suggests that Oracle may not be an outlier, but a preview of what lies ahead if the AI infrastructure buildout continues at its current torrid pace (or accelerates further).

In the next section, I explore in more detail how these hyperscalers are adjusting their approach in response to declining excess cash from operations.

How hyperscalers are adjusting as funding slack erodes

As Figure 1 shows, AI infrastructure investment is absorbing a steadily growing share of internally generated cash flow. As funding slack erodes, how are hyperscalers adjusting in practice?

At Alphabet and Meta, the primary margin of adjustment has been share repurchases. Both firms remain internally funded on a cash basis, but only by allowing buybacks to absorb the pressure created by rising AI capex. As funding coverage ratios decline, discretionary uses of cash become the buffer that preserves balance-sheet stability. The result is fewer dollars returned to shareholders in order to sustain the infrastructure buildout.

On a pure cash-flow basis, Microsoft continues to generate ample operating cash relative to current AI investment. But this apparent slack increasingly reflects the timing of cash outlays rather than the absence of commitments. As I will demonstrate in the following section, a growing share of Microsoft’s AI buildout is being financed by long-dated leases that defer cash payments while locking in compute capacity today.

Oracle gives us a view into how hyperscalers respond when financial constraints start to bind. With a funding coverage ratio now well below one, the company no longer generates enough operating cash to fund AI infrastructure investment internally. Continued expansion has therefore required reliance on external financing, including debt issuance and long-dated lease commitments to secure compute capacity. The recent decline in Oracle’s stock price is consistent with growing investor concern about whether the company’s AI infrastructure buildout is financially sustainable going forward.

Despite differences in financial strength, the hyperscalers’ responses point to a common adjustment process. As internal funding capacity tightens, firms first reduce discretionary cash uses (e.g., share repurchases) and then shift toward external financing (e.g., debt and long-dated leases) that push cash outlays into the future. The AI buildout continues, but increasingly because hyperscalers are changing how it is paid for, not because internal cash generation is keeping pace with capex growth.

The growing importance of long-dated leases

As internal funding slack erodes, hyperscalers have increasingly relied on long-dated leases to sustain the pace of AI infrastructure investment. These arrangements span both finance and operating leases.

Finance leases are best understood as deferred capital expenditures financed over time. The lessee bears the economic risks and rewards of ownership, often with a purchase option that is reasonably certain to be exercised. Thus, finance leases are economically equivalent to purchasing infrastructure with borrowed funds, which is why finance lease principal payments are included alongside capex when assessing the sustainability of AI investment.

In contrast, operating leases transfer some risk away from hyperscalers onto data center owners since hyperscalers do not own the infrastructure but instead are committing to rent space within the data center itself.

While these structures differ in accounting treatment and risk allocation, their economic effect is similar.

Microsoft was the earliest and most visible adopter of this approach. Figure 2 shows a sharp rise in right-of-use (ROU) assets obtained via leases beginning in 2023, while cash principal payments associated with these agreements have increased much more gradually. The gap reflects a growing stock of future obligations that are not yet flowing through to today’s operating cash flow.

Recent disclosures indicate that this approach is spreading across the hyperscalers. In its most recent 10-Q filing, Alphabet reported tens of billions of dollars in data center lease commitments not yet reflected on its balance sheet, with terms extending well into the next decade.

Meta reports a similarly large pipeline of lease commitments, disclosing approximately $58 billion in leases that have not yet commenced, largely related to data centers, with payments scheduled to begin over the remainder of this decade.

Oracle’s disclosures are larger still, reaching into the hundreds of billions of dollars, reflecting the fact that leases have become the primary mechanism allowing the firm to continue expanding AI capacity despite binding constraints on operating cash flow.

More broadly, long-dated leases help explain how the AI infrastructure buildout can continue even as internal funding capacity tightens. They make the buildout more sustainable on the margin by deferring cash outlays today at the cost of increasing fixed claims on future operating cash flows. In effect, hyperscalers are betting that AI demand will materialize at sufficient scale before the bills come due to service the lease commitments being layered onto their balance sheets.

What this means for Nvidia’s current valuation

The shift toward long-dated leases explains how the AI infrastructure buildout can continue even as hyperscalers’ internal funding capacity tightens. By deferring large cash outlays, leases push the binding constraint from today’s operating cash flow to tomorrow’s realized demand for AI compute.

This makes the buildout more sustainable on the margin, but only by replacing liquidity risk with commitment risk. Discretionary capital spending is converted into fixed claims on future operating cash flows, leaving hyperscalers increasingly reliant on AI workloads generating sufficient revenue to justify those obligations over time.

Nvidia’s valuation implicitly assumes that this shift succeeds. That AI demand grows fast enough to service the expanding stock of long-dated commitments now being layered onto hyperscaler balance sheets in response to tightening financial constraints. Given Nvidia’s customer concentration, hyperscaler funding constraints are not peripheral; they are central to the firm’s current valuation.

The analysis above suggests that while the AI capex boom may be sustainable in the near term, it ultimately represents a bet on whether demand for AI compute grows into the capital structure now being built to support it.

If you enjoyed this mildly efficient and occasionally rational look at the sustainability of the AI infrastructure buildout, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

That model is beginning to change.

The rise of large-scale AI has introduced a very different production process. Training and deploying frontier models requires massive upfront investment in physical infrastructure: data centers with specialized chips, networking equipment, and in some cases dedicated power generation. What were once primarily software businesses are increasingly becoming infrastructure businesses. Unlike earlier cloud build-outs, these investments are both larger in scale and more technologically specific, tying future returns more tightly to sustained demand for AI compute.

This shift is already visible on the balance sheets of Big Tech. The goal of this piece is not to make the case for whether AI investment will ultimately succeed or fail, nor is it to make a near-term valuation call. It is simply to document the transition of Big Tech from asset-light to asset-heavy firms and to explore the implications of this transition for valuing these businesses going forward.

The Surge in Fixed Capital Investment

Figure 1 plots Property, Plant, and Equipment (PPE) for Alphabet, Microsoft, Meta, and Oracle since 2011. For all four firms, PPE has risen steadily over time. What stands out is the acceleration beginning around 2019, which intensified further over the past several years.

This is not incremental investment. It is a step change in how these businesses operate.

Alphabet and Microsoft now report well over $200 billion in PPE. Meta, which historically required relatively little physical capital, has seen PPE rise sharply as it builds out AI and data center capacity. Oracle also shows a meaningful increase as it expands its cloud infrastructure and positions itself as a provider of AI compute capacity.

Rising levels of PPE alone, however, do not tell us whether this investment reflects a deeper shift in business models or simply physical capital scaling alongside revenue.

Figure 2 compares the average annual growth rates of PPE and total revenue for Alphabet, Microsoft, Meta, and Oracle between 2011 and 2025.

Growth in Big Tech’s physical capital footprint outstripped revenue growth by a wide margin over the last 15 years.

As a result, these firms are now among the largest owners of productive physical capital in the private sector. Their continued commitments to expand the infrastructure needed to deliver cutting-edge AI models suggest that this trend is unlikely to reverse quickly.

A Balance Sheet Regime Shift

This drastic change in capital allocation is reshaping the balance sheets of Big Tech. Figure 3 shows PPE as a share of total assets for Alphabet, Microsoft, Meta, and Oracle since 2011.

This is the key figure. It shows that not only is investment in PPE rising, but that fixed capital is crowding out other assets that once gave these firms flexibility. Across all four firms, fixed capital now accounts for a much larger fraction of the balance sheet than it did a decade ago.

Assets that once sat largely in cash, marketable securities, and other liquid instruments have been transformed into long-lived, capital-intensive equipment. Intangible capital still matters enormously for these businesses, but relative to fixed capital it now represents a much smaller share of total assets, altering both the risk profile and the durability of returns.

For example, in 2011 PPE on Alphabet’s balance sheet was roughly comparable in size to goodwill and intangibles. As of the most recent quarter, PPE represents nearly six times the value of Alphabet’s goodwill and intangibles.

This represents a fundamental shift in capital allocation.

Importantly, much of this investment has been funded internally. With the exception of Oracle, these firms are not levering up in the traditional sense. Instead, they are redeploying accumulated cash and short-term investments into physical capital. For all but Oracle, the risk is not balance-sheet fragility. It is reduced flexibility in a rapidly changing technological environment.

Cash preserves optionality. Fixed capital does not.

Once deployed, data centers and specialized chips cannot be easily repurposed. Their economic value depends critically on future utilization and the pace of technological change. Consequently, depreciation, both accounting and economic, now plays a much larger role in understanding these firms. In effect, depreciation has become a first-order economic variable rather than an accounting footnote.

Implications for Valuation: FCF Yield

This balance-sheet transition has direct implications for valuation, particularly for metrics based on free cash flow (FCF), which is defined as operating cash flow minus capital expenditures. For asset-light firms, or for asset-heavy firms that are not rapidly expanding, FCF yield (equal to FCF divided by market capitalization) typically provides a clean measure of how much free cash firms are generating for each dollar an investor invests in the firm. But when firms enter a period of heavy capital accumulation, that measure becomes far less informative about valuation.

Capital expenditures rise immediately, while the revenue they generate arrives with a lag. Free cash flow is therefore mechanically compressed in the short run, even if long-run cash-generating capacity is expanding. This tends to drive down FCF yields.

In the case of AI infrastructure, investments are front-loaded, technologically specific, and subject to rapid economic obsolescence. Maintaining capacity is likely to require continuous reinvestment in rapidly depreciating chips rather than one-time spending. If a growing share of revenues are allocated toward maintenance rather than expansion, the implications for long-run free cash flow are very different.

This dynamic is already visible in recent valuation metrics. Figure 4 shows the FCF yields of Alphabet, Microsoft, Meta, and Oracle have declined meaningfully in recent years as their market caps have increased while rising investment in PPE has slowed growth in their FCFs.

At first glance, this looks like a familiar story. Declining FCF yields are often taken as a sign that valuations are becoming stretched. In periods of heavy capital accumulation, however, they can instead reflect the timing mismatch between upfront investment and downstream cash generation.

If markets believe that this spending will eventually generate cash flows that exceed the cost of capital, and the recent rise in market caps suggests that this is indeed the case, equity prices can rise even as free cash flow falls. FCF yields decline not because investors are ignoring cash flows, but because they are looking through a transitory investment phase.

Where the Real Valuation Risk Lies

Taken together, these trends suggest that valuation risk has shifted in an important way. The key question is not whether today’s free cash flow looks expensive relative to market capitalization. It is whether these rapidly growing investments in fixed capital will ultimately earn returns that justify their scale.1

Answering that question requires shifting attention away from short-run cash-flow metrics and toward the economics of returns on incremental capital—an issue that merits separate treatment.

If AI-driven revenues scale smoothly, today’s relatively low FCF yields may prove temporary. As utilization rises and capital spending normalizes, free cash flow could rebound sharply, driving up FCF yields.

If revenues fail to materialize, however, the downside looks different than in the past. Depreciation will continue to rise independent of revenue growth. Maintenance capex will be required even if growth slows, creating a persistent drag on free cash flow. And the cash that once provided a buffer against uncertainty has already been spent.

A Different Kind of Big Tech

A defining feature of the AI era is likely to be the transformation of Big Tech into some of the most capital-intensive firms in the economy.

This is a familiar dynamic in capital-intensive industries. Railroads, utilities, and telecommunications firms have all experienced periods where heavy investment distorted traditional cash-flow-based valuation metrics. What is new is seeing this dynamic play out at firms that investors have long treated as archetypal, asset-light software businesses.

As the figures above clearly demonstrate, this transformation is already well underway. Balance sheets are tilting toward fixed assets. Cash and short-term investments are being converted into long-lived, rapidly depreciating capital. Free cash flow, once remarkably stable, is coming under pressure as investment cycles lengthen and the need for ongoing physical capital investment rises.

In this new regime, falling FCF yields are not, by themselves, evidence of excess or irrational exuberance. They reflect the fact that capital is being deployed today to secure compute capacity that may only be fully monetized years from now.

The real valuation question is no longer simply whether these firms look expensive relative to current free cash flow. It is whether the returns on this expanding base of productive capital will ultimately justify its scale.

Big Tech is no longer just writing code. It is building infrastructure. And that shift changes how we should think about valuing these businesses.

If you enjoyed this mildly efficient and occasionally rational look into Big Tech’s balance sheet regime shift, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

To put today’s market in context, Figure 1 shows that the current CAPE ratio of about 40—Shiller’s cyclically adjusted price-to-earnings ratio, which smooths earnings over the business cycle—has only been matched once in the last century and a half. And that occurred during the height of the dot-com bubble.

Many market observers are quick to dismiss any discussion of the market being overvalued based on multiples like the CAPE ratio. They point to the fact that while a near-record high CAPE ratio may be a sign that valuations are indeed currently stretched, timing a market exit based on this metric has caused many investors to miss out on major run-ups in stock prices.

Once this point has been made, the next argument usually goes something like this:

Firms today are far more profitable than they were in the past, which justifies today’s near-record-high multiples.

Unfortunately, this argument is fundamentally flawed. A firm with higher profit margins produces more earnings for every dollar of sales. Holding prices (P) constant, higher profit margins raise earnings (E), which mechanically leads to a lower P/E ratio, not a higher one. Investors may then bid up the stock price if they believe the margin gains (and therefore higher earnings) will persist. As the price rises, the P/E ratio moves back toward its previous level. In this simple equilibrium, prices rise just enough to reflect higher earnings, leaving the P/E ratio essentially unchanged.

High profit margins explain why earnings are high. They do not explain why investors should pay more per dollar of earnings.

To justify a higher P/E ratio, higher margins must signal higher expected earnings growth or lower perceived riskiness of the firm’s cash flows, or both. Margins by themselves are not enough.

For example, if higher margins reflect a new, durable competitive advantage that will persist for many years, then the expected growth rate of future earnings rises. A higher growth rate increases the present value of the firm, which supports a higher price today and, thus a higher P/E ratio.

Alternatively, if higher margins suggest that future cash flows are less risky, investors may require a lower return to hold the stock. A lower discount rate increases the present value of a firm, which leads to a higher price today and, therefore a higher valuation multiple.

In both cases the justification for a higher P/E ratio lies in long run assumptions about growth or risk, not in the firm’s current margins.

This distinction becomes especially important in the AI era. The firms at the center of the current boom have remarkable profit margins (e.g., Nvidia), and some have plausible paths to even greater profitability as AI becomes embedded in their ecosystems. Investors may reasonably believe that these competitive advantages will create a durable moat into the future. They may also believe that AI will lead to higher earnings growth for many years. If so, today’s high valuation multiples could be justified. But the justification hinges on very strong assumptions about the future. It does not follow from record-high margins alone.

Once the margin narrative is removed, the valuation question becomes much sharper. The stock market is clearly expensive by historical standards. To defend these valuations, one must rely on bold claims about the long run impact of AI or the persistence of dominant platforms in the digital economy. Those arguments may turn out to be correct. But they should be stated plainly, not hidden behind the argument that firms are more profitable than they ever were in the past.

The bottom line is simple. If today’s near-record valuations prove justified, it will be because investors are right about the future path of earnings and risk, not because today’s dominant firms happen to earn unusually high margins.

If you enjoyed this mildly efficient and occasionally rational debunking of the margin myth, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths along the way.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

To answer that question, it is useful to step back and recall what the unemployment rate actually measures. The unemployment rate is the ratio of unemployed workers to the total size of the labor force, where the labor force is defined as the sum of employed and unemployed workers:

Unemployment rate = Unemployed / (Employed + Unemployed)

This matters because changes in the unemployment rate can arise for very different reasons. An increase could reflect widespread job losses, a surge in new job seekers, or some combination of the two.

Figures 2 and 3 show how the number of employed and unemployed people have changed since the start of the year.

Between January and November, the number of unemployed people increased by roughly 980,000. Over the same period, the number of employed people fell by only about 150,000. Taken together, these two facts make clear that the rise in unemployment cannot be explained simply by workers moving from employment into unemployment.

So what is going on? How can the number of unemployed people rise by far more than the number of employed people falls?

The answer is that the size of the labor force itself must have increased. In other words, a large number of people who were previously neither working nor actively looking for work must have decided to enter the labor market.

As shown in Figure 4, the civilian labor force grew by approximately 830,000 people between January and November. This figure is not an estimate or a model-based inference. It follows directly from the accounting identity that the labor force equals employment plus unemployment: 980,000 more unemployed workers minus 150,000 fewer employed workers implies 830,000 new labor force entrants.

Viewed through this lens, the recent rise in the unemployment rate takes on a very different interpretation. It was driven almost entirely by an influx of new job seekers, not by a collapse in employment.

As a general rule, people do not enter the labor market and begin actively searching for work unless they believe their chances of finding a job are reasonably good. Job search is time-consuming, mentally exhausting, and often financially costly. While some individuals may be pushed into the labor market by necessity, changes in labor force participation at the margin tend to be pro-cyclical: people are more likely to look for work when jobs are plentiful and hiring conditions are favorable.

The increase in the labor force over the past year therefore suggests that the labor market remains relatively strong. Workers appear confident enough in job prospects to step off the sidelines, even if not all of them are immediately successful in securing employment.

To sum up, breaking the unemployment rate into its underlying components tells a very different story than the headline number alone. The rise in the unemployment rate since the start of the year is not a sign that the labor market is rolling over. Paradoxically, it reflects continued labor market strength, as more people are willing to enter the job market and actively compete for available positions.

For investors, the key takeaway is that a rising unemployment rate driven by labor force expansion carries very different implications than one driven by job losses. Historically, recessions and major earnings downturns are associated with sharp declines in employment, not with large inflows of new job seekers. When unemployment rises because more workers are confident enough to look for jobs, it typically signals ongoing economic momentum rather than an imminent contraction. Investors who react mechanically to higher unemployment readings risk mistaking a healthy, expanding labor market for a weakening one and mispricing recession risk.

A sharper policy implication follows directly from this distinction. From the Fed’s perspective, an unemployment rate rising for “good” reasons, all else equal, provides less justification for rapid or aggressive easing.

If you enjoyed this mildly efficient and occasionally rational look at the recent rise in the unemployment rate, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

A nearly century-long pattern linking firm size and long-run returns broke down in the mid-to-late 2000s. This piece explains what changed and why it matters.

When I teach Finance Theory, I often begin with a simple thought experiment that illustrates a timeless lesson about risk and return. I ask students to imagine they have $100 to invest today on behalf of their great‑grandchildren 80 years from now. The question is straightforward: What asset would you choose?

To anchor the discussion, I show students how $100 invested in 1928 would have grown by 2006 across several major asset classes (Figure 1). Small cap stocks dominate the comparison.1 They turn $100 into nearly $2 million, while large caps, corporate bonds, and Treasury bills generate far smaller sums. The implication seems obvious: over long horizons, small caps outperform, so long‑term investors should favor them.

Next, we compute the average annual return and volatility for each asset class. When plotted in risk and return space (Figure 2), the pattern becomes even clearer. From 1928 to 2006, higher risk was consistently associated with higher return.

The historical relationship between risk and return depicted in Figure 2 is strikingly linear, with each asset class offering investors roughly the same amount of reward per unit of risk.2 If your investment horizon is multiple decades, the lesson is clear. You should hold a diversified portfolio of the riskiest asset class you can tolerate, because in the long run this approach produces the highest average return. Historically, that meant investing in small cap stocks.

When I taught this material in earlier years, this is where the analysis ended, and so the class proceeded under what seemed like a reasonable assumption: the beautifully linear relationship between risk and return across diversified portfolios of various asset classes would continue indefinitely.

Why question it? For nearly eight decades, it looked like a law of nature.

Only now, with nearly two decades of additional data, is the flaw in that argument obvious. The tidy historical relationship between risk and return did not simply weaken in recent years. It broke.

The Modern Era: A Broken Relationship

To understand this break, it is helpful to look more closely at how small cap stocks behaved in earlier decades. Figures 3 and 4 plot the annual returns for small caps and the S&P 500 across two nearly equal length periods.

Figure 3 shows annual returns from 1928 through 1976. Small caps frequently delivered extraordinarily large gains during this era, returning more than 50 percent in fourteen different years, and exceeding 100 percent in three of those years. These explosive upside years help explain why small caps compounded so dramatically over the long run. Despite substantial volatility, the right tail of annual returns was exceptionally long.

Figure 4 covers 1977 through 2024 using the exact same vertical scale. The pattern is strikingly different. Across almost fifty years, there are only two instances of small caps returning 50 percent or more. This suggests that in recent decades small caps have benefited far less from the extreme right tail events that historically compensated investors for bearing their higher degree of risk.

What is perhaps most surprising given this data is that the average return on small caps from 1977 through 2024, at 11.5 percent, is nearly identical to the average return from 1928 through 1976. The disappearance of huge “home run” years is not the primary reason for the more recent underperformance.

In fact, between 1977 and 2006, a period with just a single year in which returns exceeded 50 percent, small caps compounded at an astonishing rate of nearly 15.5 percent per year. At this clip, investors’ money was doubling roughly every 4.5 years!

In contrast, between 2007 and 2024, small caps returned just 5.1 percent per year versus 10.3 percent for the S&P 500.3 The long‑standing historical pattern of small cap outperformance did not fade gradually. It broke sharply in the post‑2006 environment.

Is This Nearly Two Decade-Long Period Unique?

Given the dramatic deterioration in small cap performance since 2007, a natural question is whether this nearly two-decade period is unusual in a historical context. After all, markets are cyclical. Perhaps small caps are simply in a prolonged dry spell and will eventually mean revert.

Figure 5 offers useful perspective by plotting the 20-year rolling average small cap premium relative to the S&P 500 from 1928 through 2024.

Importantly, since Figure 5 plots 20-year rolling averages, short-run fluctuations are muted. This makes the magnitude of the post-2006 decline (and the decline of the late-1990s) even more striking.

Several features of the historical record stand out:

1. Extended periods of small cap underperformance are not unprecedented. The premium fell sharply in the mid-1960s, again in the late 1980s, and moved sharply negative in the late-1990s.

2. Although the small cap premium has fluctuated considerably over time, the depth and persistence of the post-2006 decline is matched only by the decline during the late-1990s.

3. The small cap premium has been drifting downward for decades — the recent episode is part of a longer-run trend.

Taken together, Figure 5 suggests that the past eighteen years may not simply be another market cycle. Instead, the magnitude and persistence of the decline, combined with the long-run downward drift, point to a structural shift rather than a temporary blip. This raises an important question: what changed?

The Risk-Return Trade-off in the Modern Era

If the small cap premium disappeared after 2006, did their volatility decline as well?

Figure 6 again depicts the average annual return and volatility for each asset class in risk and return space. The volatility of small cap stocks did indeed decline meaningfully during the post-2006 period, which in most cases would be considered an improvement, but their returns fell proportionately more. In fact, the average annual return on small caps from 2007 to 2024 was not only significantly lower than the average return on the S&P 500 over this same period, it was lower than the average return on corporate bonds, despite exposing investors to nearly triple the risk.

As a result, the nearly linear relationship between risk and return that characterized the 1928-2006 period simply vanished. Small caps now sit well below even a naive mean-variance frontier constructed from major U.S. asset classes and thus no longer appear to offer a compelling risk-return trade-off.4

Why This Time Is Different

Figures 5 and 6 paint a dismal picture for small caps stocks. The small cap premium has turned negative and shows no signs of recovering and, as a result, small cap stocks as an asset class look far less attractive to the long term investor than the S&P 500.

To understand why this likely isn't just a temporary blip, let's consider how the structure of small firms and the broader economy has changed. Two developments in particular stand out, and together they provide a coherent explanation for the post‑2006 regime shift.

(1) A Lower Quality Small Cap Universe

The quality of small caps has deteriorated over time. The bottom decile of stocks when ordered by market cap now contains a growing share of unprofitable firms, low‑return businesses, and so‑called “zombie” companies.

Academic research supports this view. Fama and French (2015) show that profitability and investment factors explain a large portion of the cross section of returns, and firms with weak fundamentals tend to earn persistently low returns.

Asness, Frazzini, and Pedersen (2019) further document that low‑quality firms underperform systematically, with the effect especially pronounced among smaller companies.

Reinforcing the importance of firm quality, methodology papers from S&P Dow Jones show that the S&P 600’s profitability screen produces substantially better long‑term results than simply investing in the bottom decile by market cap.5 Indeed, between 2007 and 2024, the average annual return for the S&P 600 was a respectable 8.7 percent versus just 5.1 percent for the bottom decile by market cap, a striking difference given that both indexes target the same segment of the market.6

A key factor driving the deterioration in small cap quality is the well-documented decline in IPO activity. Gao, Ritter, and Zhu (2013) show that this decline reflects a structural shift in the economics of going public. High‑growth firms increasingly choose to remain private longer because scale advantages, network effects, and abundant private capital make it attractive to delay an IPO.

As a result, many promising young firms now enter the public markets only after reaching mid‑cap or large‑cap size. For example, Rivian, Coinbase, Snowflake, Airbnb, DoorDash, Robinhood, Arm, Uber, Lyft, Zoom, Spotify, Datadog, and Palantir all entered the public markets at valuations ranging from roughly ten billion dollars to well over fifty billion dollars. The result of companies waiting longer to make an IPO is a weaker pipeline of newly public small cap firms.

(2) Large Cap Dominance in a Winner-take-most Economy

The modern economy increasingly rewards scale, data, and intangible capital. Industries such as cloud computing, search, digital advertising, semiconductors, and artificial intelligence exhibit powerful network effects and large fixed‑cost advantages. These industries reward being big, sometimes astonishingly so. As a result, large cap firms benefit disproportionately.

Small caps, on the other hand, operate more heavily in sectors without these advantages, including regional banking, local services, consumer discretionary products, and industrials.

This structural shift was not present for most of the 1928 through 2006 period. Evidence from Autor et al. (2020) shows that a small number of highly productive superstar firms have increasingly captured a disproportionate share of industry sales, profits, and market power.

Related work by De Loecker, Eeckhout, and Unger (2020) documents a broad rise in markups and market concentration, with the largest firms capturing an increasing share of economic surplus.

Together, these findings support the view that the modern economy structurally channels economic gains toward large firms, which helps explain why large caps have strengthened and small caps have weakened.

So Are Small Caps Actually Dead?

The small cap premium we discuss in lecture, and that appears so cleanly in nearly a century of historical data, was never guaranteed to persist unchanged. It emerged in an economic environment where small, profitable, capital intensive firms could grow quickly, go public early, and challenge incumbents. In that world, the small cap universe was replenished each year with new, dynamic firms that helped sustain strong long run returns.

That world has changed. Today’s economy increasingly rewards scale, data, and intangible capital. At the same time, high growth firms now remain private far longer, often reaching mid cap or large cap size before entering public markets. The combination of a weaker IPO pipeline and an economy that disproportionately channels gains to dominant firms has reshaped the opportunity set for the long term investor.

Small caps are not dead, but the world that once powered a broad and persistent small cap premium is gone, and investing as if it still exists means ignoring the reality of the modern economy.

If you enjoyed this mildly efficient and occasionally rational look at why small cap stocks have underperformed in recent years, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

At first glance, this print appears to show that the labor market remains on solid footing.

But the monthly payroll print is not a precise measure of job growth in the economy. It is a statistical estimate with substantial sampling error, as I alluded to in my last piece. Indeed, the BLS reports that the standard deviation of the monthly payroll estimate is roughly 83,000 jobs.1

Using the familiar 95 percent confidence interval, which corresponds to roughly two standard deviations around the estimate, the actual change in nonfarm payrolls during September almost certainly lies somewhere between -43,000 and +281,000.

In other words, a reported gain of +119,000 is consistent with anything from a modest contraction to strong growth in hiring. And this interval applies to the final, fully processed estimate, the one produced after two rounds of revisions, when nearly all survey responses have been received. The initial estimate, which is the version that makes the headlines each month, carries even more uncertainty because it is only based on a subset of the full sample.

What Sampling Error Looks Like

To see this more clearly, imagine the economy actually added a modest +50,000 jobs last month. Given the sampling error, repeated BLS-sized samples of the actual economy can plausibly produce reported changes in monthly nonfarm payrolls that range from strongly negative to strongly positive, purely because of statistical noise (Figure 1).

The takeaway is simple: a single month’s payroll print can look strong or weak regardless of the true state of the labor market. This is why relying on a single monthly print can produce a materially misleading picture of labor market conditions.

Why You Should Ignore the Monthly Payroll Print

Two issues make the monthly payroll print a poor real-time gauge of labor market health:

1. The signal-to-noise ratio is extremely low.
Unless the reported change is very positive or very negative (roughly ±166k, or about two standard deviations from zero), we cannot reliably say whether payrolls actually rose or fell during the month.

2. Payroll levels can move for structural reasons.
Payroll growth reflects demographic and population dynamics (e.g., changes in immigration flows), not just cyclical factors. During periods of rapid labor force growth, payroll gains can look strong even when underlying tightness is easing, and vice versa.

Better Real-Time Indicators: The Four Horsemen

As I argued in my Four Horsemen of the Labor Market piece, there are far more useful (and robust) indicators of the health of the labor market:

• Job Openings-to-Unemployment Ratio — direct measure of labor demand

• Layoff Rate — the probability an employed worker loses a job this month

• Job-finding Rate — the share of unemployed workers who become employed this month

• Unemployment Rate — the share of the labor force that is unemployed

These are the labor market indicators to focus on. They move early in the cycle, contain far less noise, and provide a clearer real-time signal than monthly payroll prints.

Bottom Line

The monthly payroll print from the BLS is a poor real-time guide to labor market conditions. It is statistically noisy, especially in its first release, and mechanically influenced by population and labor force growth.

In practice, this means the monthly payroll print contains more noise than signal and can mislead more than inform. For these reasons, investors should fade the monthly payroll print and focus instead on more robust indicators of labor market health.

If you enjoyed this mildly efficient and occasionally rational look at why market observers should fade the monthly payroll print from the BLS, consider subscribing below. We’ll keep exploring markets and models, uncovering mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

Investors’ intense focus on small month-to-month employment changes is misplaced. The monthly payroll report has a large sampling error, and more importantly, the structure of the U.S. labor market has changed.1 It now takes a much larger shock to push employment into recession than it once did. In practice, this means the next downturn is far less likely to emerge gradually through softening payroll data and far more likely to arrive suddenly, driven by a large, unmistakable shock.

The key reason is structural. As the U.S. economy shifted from goods-producing industries to service-providing ones, total employment became less sensitive to the kinds of disturbances that used to trigger recessions. In short, the structure of employment changed in ways that make typical business cycle downturns less likely, making month-to-month payroll variations less informative than they once were.

In what follows, I walk through how the labor market has changed and why this shift makes the economy far more resilient to shocks that, in the past, would have derailed both employment and financial markets.

For investors, the implication is straightforward: obsessing over monthly payroll numbers is a recipe for misreading recession risk in an economy that now only breaks when it is hit by a truly large adverse shock.

From Frequent Recessions to Rare Ones

Every recession since the late 1930s has been accompanied by a decline in total employment (Figure 1). This is why investors are right to be laser-focused on the labor market.

But as Figure 1 also shows, recessions have become less frequent over time. The mid-1940s through early 1980s saw downturns every few years, often triggered by war, oil shocks, or monetary tightening. Since the mid‑1980s, the U.S. has endured only four recessions, roughly one per decade.2 Two resulted from extraordinary shocks, namely the global financial crisis and the pandemic‑induced shutdown, and one from the bursting of the dot-com bubble. Moreover, the 1990 and 2001 recessions, while damaging to equity markets, produced only mild and relatively short‑lived declines in total employment.

One interpretation of the fact that recessions have become less frequent is that policymakers became better at managing the economy or more willing to aggressively ease monetary policy and provide fiscal stimulus at the first sign of weakness.

Another explanation, and the focus of this piece, is that the economy itself changed in ways that dampen the economy’s response to a typical recessionary shock. A prime example of this is how the economy withstood the most recent bout of monetary tightening following the high inflation of 2021-22. Nearly everyone, including myself, thought the Fed’s actions would tip the economy into recession. Yet the labor market never rolled over. Instead it remains remarkably resilient.

Does Service or Goods Employment Drive Recessions?

To understand why the economy held up so well despite aggressive tightening, and why it has become more resilient to shocks more generally, it helps to look directly at how different parts of the labor market respond in recessions.

Figure 2 shows that in every recession prior to 2008, goods-sector employment (solid blue lines) fell sharply, while service-sector employment (dashed red lines) barely moved, and in many cases continued rising.

The financial crisis and pandemic recessions were different: both directly hit large service industries (finance/real estate in 2008–09, leisure and hospitality in 2020). Outside these two episodes, service employment has generally held up well during recessions.

A Tale of Two Sectors

The easiest way to visualize differences across sectors in the sensitivity of employment to economic shocks is to detrend the data and focus in on their business cycle fluctuations. Figure 3 does precisely this by plotting the percent deviation from trend for goods-producing and service-providing employment between 1939 and today.

From Figure 3, we see that goods-producing employment tends to rise quickly during expansions and fall sharply during downturns. Service-providing employment, on the other hand, tends to move in a similar fashion, but more smoothly and with a notably smaller amplitude. The exception, of course, is the COVID recession in which leisure and hospitality employment fell precipitously, pushing service-producing employment well below trend.

Why Are Services More Resilient?

Service-providing employment is inherently more stable for several reasons:

1. Demand doesn’t fall as sharply during downturns. Demand for services like health care, education, and government tends to be highly inelastic, so spending holds up even when income softens. Households will delay buying a car or upgrading their washing machine during a recession, but they rarely cancel a medical procedure or pull their kids out of college.

2. No inventory cycle to amplify shocks. Goods can be stockpiled, amplifying booms and busts while services cannot.

3. Limited exposure to global supply shocks. Most services don’t depend on imported inputs or energy-intensive production, so when global supply chains seize up, or oil prices surge, the pain is felt mostly in goods-producing industries.

4. Automatic stabilizers. Employment in the government, education, and health care sectors tends to remain stable or even rise during downturns.

Together, these features make service-sector employment inherently less volatile over the business cycle.

The Labor Market Is Structurally More Resilient Than in the Past

These sector-level differences only matter for the aggregate economy in proportion to each sector’s share of total employment. And as Figure 4 makes clear, the composition of employment changed dramatically between the mid-1950s and the late 2000s.

In the mid-1950s, roughly 62 percent of U.S. workers were employed in the service sector. By the late-2000s, this number had risen to about 86 percent. Conversely, goods-producing employment fell over this same period from nearly 40 percent of total employment in the mid-1950s to just 14 percent by the late-2000s.

The changing composition of the labor market toward less volatile service-providing industries means that the overall labor market exhibits smaller fluctuations when hit by a typical recessionary shock. Whereas prior to 1980, it was common for total employment to be pushed 2 percent or more below its long-run trend during recessions, since 1980 the only two times this threshold was breached were the financial crisis and COVID-19 pandemic (Figure 5).

This basic pattern is not new. Economists have documented many elements of this shift already.

Research by Davis and Kahn, for example, shows that part of the “Great Moderation” came from the U.S. quietly shifting away from volatile manufacturing and construction jobs toward steadier service-sector work. Other studies (e.g., Foerster, Sarte, and Watson [2011]) find that shocks to manufacturing used to drive recessions, but their impact has faded as the sector has shrunk. (If you’re curious, I’ve included a list of further reading at the end with more related academic articles.)

My argument pulls these strands of the literature together and pushes them one step further: because the most cyclical parts of the economy now make up a much smaller share of employment, it takes a much bigger, more system-wide shock to produce an actual recession. The mechanisms that used to generate recessions have weakened, making the economy structurally more resilient.

Today, to generate a broad‑based decline in employment, a shock must either:

(a) hit the service sector directly, or

(b) be severe enough to drag services down indirectly.

This is why typical supply shocks—like oil spikes in the 1970s—rarely trigger recessions today. It now takes something on the scale of a financial crisis or a global pandemic. This also helps explain why the recent Fed tightening cycle did not produce a recession.

Implications for Investors and Policy Makers

These structural changes don’t just alter which shocks are capable of generating a recession. They also have important implications for investors and policy makers:

• Manufacturing revival vs. macro stability. Efforts to “re-industrialize” the economy may have political appeal, but a larger manufacturing base would also make the overall economy less resilient.

• AI and sector‑specific bubbles. If today’s AI investment boom unwinds, the direct job losses would be small—information services account for less than 2 percent of total employment. The risk is that a bust could spill over into construction and equipment manufacturing, or if the market for collateralized debt tied to AI-related infrastructure sours.

• Stock market valuations. A more resilient economy justifies higher equity valuations: if recessions are less frequent and require larger shocks, the risk premium investors demand falls, supporting higher stock prices.

Key Takeaway

The U.S. economy has become far less fragile than stock market fluctuations suggest. Investors may hang on every payroll report, treating routine month-to-month variations as signs that the labor market is cracking, but in a service-dominated economy that’s simply not how recessions form. Routine hiring slowdowns or modest increases in unemployment tell us very little about recession risk because it now takes a truly large, system-wide shock to make the labor market roll over. In short, the rise of service-sector employment has raised the threshold for recession. The economy isn’t immune to downturns, but it is more resilient and far harder to push into recession, than it used to be.

Share

If you enjoyed this mildly efficient and occasionally rational take on how the rise of service-providing employment has made the economy more resilient, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Further Reading

If you’d like to explore the academic research behind this topic, here are some of the most relevant papers and a brief summary of their main findings:

• Davis & Kahn (2008), “Interpreting the Great Moderation.”
Shows that part of the decline in macroeconomic volatility since the mid-1980s reflects a shift away from highly cyclical industries like manufacturing and construction toward steadier service-producing sectors. This is one of the closest academic foundations for the argument in this piece.

• Foerster, Sarte & Watson (2011), “Sectoral Shocks and Aggregate Fluctuations.”
Finds that historically, goods-producing industries accounted for a disproportionately large share of aggregate volatility. As those sectors have shrunk, so has the impact of goods-sector shocks on the overall economy.

• Ariu, Kikkawa, Lange & Leduc (2020), “On the Resilience of Services During the Great Recession.”
Documents that service employment contracted far less than manufacturing during the Great Recession across advanced economies, highlighting the inherent stability of the service sector.

• Moro (2015), “Structural Change, Growth and Volatility.”
Shows that as economies evolve from agriculture → manufacturing → services, both output and employment volatility naturally decline.

• Stock & Watson (2002, 2003), “Has the Business Cycle Changed?”
While emphasizing improved monetary policy as part of the Great Moderation, they also note the role of structural factors, such as shifts in industry composition, in reducing volatility.

• Baily & Bosworth (2014), “The United States Economy: Why Such Slow Growth?”
Provides evidence that employment volatility has fallen significantly and that goods-producing industries were historically the main drivers of recessions.

• Ngai & Pissarides (2007), “Structural Change in a Multisector Model of Growth.”
A foundational model explaining why the service sector rises over time due to different productivity trends and income elasticities. While not directly about recessions, it underpins the long-run shifts discussed here.

• Fang & Rogerson (2011), “Productivity Shocks, Employment, and Hours in the U.S. Postwar Economy.”
Shows that employment volatility fell as economic activity shifted toward less-cyclical industries, consistent with the theme of this post.

These works collectively show that the composition of the economy matters enormously for how recessions unfold, an idea central to the argument here.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

Many have speculated that the challenging job market for recent graduates reflects some combination of pandemic-era overhiring, the rise of AI reducing demand for entry-level workers, or broader macro uncertainty, such as shifting tariff policy.

There’s no doubt that for those new grads who are still searching for their first job, this period can feel especially discouraging. Starting a career in a cooling labor market is never easy, and the mix of headlines about layoffs and automation only amplifies that anxiety. That’s exactly why it’s worth stepping back and looking at what the numbers actually show.

As I’ll argue below, based on the data available today, the picture for new grads looks far less dire than the headlines suggest.1 The data reveal remarkably consistent patterns, both across the seasons and over the long run.

Specifically, I will examine unemployment rates using data from the Bureau of Labor Statistics for 20- to 24-year-olds, the group that captures most new entrants to the labor market with a bachelor’s degree. To separate seasonal noise from real weakness, I’ll mainly be using data that is not seasonally adjusted. As it turns out, this approach yields important insights into the job market for new grads.

Young Workers Always Have a Higher Unemployment Rate

The first observation is that young workers generally have a higher unemployment rate than the economy-wide average (Figure 1). Indeed, the unemployment rate for 20- to 24-year-olds typically runs between 1.5–2 times the overall unemployment rate, a relationship that has held remarkably steady for decades.

Young workers have had a higher unemployment rate in every cycle, from the dot-com bust to the post-COVID expansion. This reflects normal labor market churn: young people are typically entering the workforce for the first time and switch jobs frequently, both of which tend to raise their unemployment rate relative to older, more established workers.

So while the unemployment rate for 20- to 24-year-olds has ticked up recently, increasing the gap between it and the overall unemployment rate, it’s still well within its normal range.

Summer Spikes are Normal and Entirely Predictable

Every summer, the unemployment rate for workers aged 20-24 with a bachelor’s degree jumps sharply (Figure 2). The reason is simple: it is graduation season. Thousands of undergraduate students finish school and enter the labor force during the summer, and it takes time to find their first full-time job.

This seasonal pattern is so strong that the unemployment rate for recent college graduates sometimes rises above that of similarly aged workers without a college education, albeit briefly. This reversal happened during the summers of 2022, 2023, 2024, and it came close to happening again this past summer. But come fall, the unemployment rate for recent college graduates tends to settle back below that of their less educated peers.

It follows that this year’s “spike” in unemployment is not necessarily a warning sign; it’s expected and entirely normal.

Summer Spikes are Mostly Driven by Men

The unemployment rate for males who recently graduated from college tends to rise more during the summer than for their female peers, a pattern that held true again this past summer (Figure 3). The reasons could range from industry mix to graduate school enrollment or simple timing differences in job search. Whatever the cause, the gender gap reinforces a simple point: while joblessness imposes a real hardship on individuals struggling to find work, as of today the data point to a familiar seasonal pattern, not a white-collar recession.

The Bottom Line

Despite headlines about corporate layoffs and “AI eating white-collar jobs,” there’s little evidence that new graduates face a uniquely hostile job market. The summer uptick in unemployment of recent grads happens every year, and by fall, most find work. Still, for those searching, the wait can be excruciating, especially when every news alert seems to confirm their worst fears. The transition from college to career nearly always takes patience, and perhaps this year, a bit more of it.

Of course, none of this means AI and automation won’t reshape the white-collar job market eventually. But for now, the data tell a simpler story. The rhythms of graduation season and the temporary spike in unemployment that comes with it are as steady as ever.

For now, the Class of 2025 can take comfort in knowing that what they’re experiencing is both temporary and entirely normal.

If you enjoyed this mildly efficient and occasionally rational take on the job market for recent college grads, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Notes and Sources

AI tools were used to edit prose; all figures are straightforward to reproduce from the cited sources.

At the center of this investment revolution sits the S&P 500 index fund. It is the go-to vehicle for any investor seeking broad exposure to the U.S. stock market. Indeed, the top 25 S&P 500 index mutual funds had a total of nearly $2.5 trillion in assets under management (AUM) as of October 8, 2025. If we include the three most popular S&P 500 index ETFs (VOO, SPY, IVV), the total AUM of S&P 500 index funds (both mutual funds and ETFs) approaches $5.3 trillion. This staggering sum represents about 8% of the total U.S. equity market capitalization, roughly $67.8 trillion as of October 1, 2025.

In theory, these funds should all be extremely low cost. After all, they invest in the same 500 companies, weighted the same way as the underlying S&P 500 index. There is little for the fund manager to do once the fund is up and running aside from rebalancing quarterly as firms enter and leave the index. In practice, a simple automated system could manage this process at close to zero marginal cost.

However, a quick screen of ETFs and mutual funds on the Charles Schwab platform revealed significant variation in the fees charged by S&P 500 index funds available to investors today. Given that all these funds hold the same stocks, this variation is hard to justify. It is plausible that fund companies differ in their fixed costs of operation, but the range of fees charged by essentially identical index funds is astonishing. Gross expense ratios—the preferred measure when comparing investments to be held over long horizons—range from a low of 0.02% to an egregiously high 1.65% (looking at you, Rydex Funds).

For the financially savvy investor, this should be an easy decision: choose the cheapest fund and move on. Yet the data suggest many investors are doing exactly the opposite.

When I looked at AUM, large sums of money were sitting in funds with far higher costs than the cheapest available option. In particular, the 22 funds with the highest gross expense ratios hold a combined $815 billion in AUM and charge investors an average gross expense ratio of 0.49%, a full 47 basis points above the lowest-cost competition.

These basis points add up. When weighted by AUM, the 22 highest-cost funds are quietly siphoning off nearly $1 billion per year in excess fees, not in exchange for top-tier fund managers working hard to beat the market, but by simply following Standard and Poor’s instructions.

A difference of 40 or 50 basis points (0.40% to 0.50%) in gross expense ratios between otherwise identical S&P 500 index funds may sound trivial. But over decades, those small annual fee differences compound dramatically.

Let’s walk through an example to make this point more salient:

Suppose you invested $100,000 in the Invesco S&P 500 mutual fund (Class A, ticker SPIAX) with a 0.54% gross expense ratio. If the S&P 500 returns 10% annually for 20 years, you’d end up with $609,690. The same investment in Fidelity’s FXAIX or Schwab’s SWPPX, both charging just 0.02%, would grow to $670,308. That’s a $60,000 difference for the same underlying investment!

The math is simple but powerful. The table below shows how small fee differences snowball over time.

To make matters worse, some S&P 500 index fund options also charge what are called “loads.” These fees, which range from 0% up to as high as 5.5%, are levied when you invest in the fund.

In addition to charging investors a full 52 basis points more than the lowest-cost options, the Invesco S&P 500 mutual fund (Class A, ticker SPIAX) also levies a 5.50% front-end load. Returning to our example, if you invest $100,000 in the Invesco fund and the S&P 500 returns 10% annually over the next 20 years, your investment will have grown to just $576,157. The front-end load alone reduces your account value by another $34,000, leaving you with about $94,000 less than if you had simply invested in the lowest-cost alternative.

So what explains this persistence of high-cost funds in a supposedly efficient market?

A few possibilities:

• Low financial literacy. Research by Annamaria Lusardi and Olivia Mitchell shows that many Americans struggle with basic financial concepts such as compound interest and inflation. If investors do not grasp how small fee differences accumulate, they may not see a problem worth fixing.

• Advisor incentives. Some advisors still receive compensation, directly or indirectly, from the funds they recommend. That can steer clients toward higher-cost products even when cheaper, identical ones exist.

• Inertia and confusion. In my own work with Casey Rothschild, “Financial Sophistication and Portfolio Choice Over the Life Cycle” published in the Review of Economic Dynamics (October 2017), we show how limited financial knowledge reinforces this inertia, leading investors to hold suboptimal portfolios and pay unnecessary fees. Mutual fund menus cluttered with share classes and fine print only make matters worse.

The irony is that the S&P 500 index fund, pioneered by John Bogle of Vanguard, was supposed to democratize investing by letting everyone capture market returns without the friction of high fees or opaque strategies. Somewhere along the way, simplicity got rebranded—and repriced. Indexing was meant to make investing easier for individuals. But as the data show, even simplicity demands vigilance.

The takeaway is simple: if you are buying the same index, you should never pay more for it. Check the fund’s expense ratio, avoid load fees, and make sure your advisor’s incentives align with your own. If your financial advisor put you in a high-cost S&P 500 index fund, it’s not too late to find a new one—but it’s long past time to ask why.

Notes and Sources

AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational take on the hidden costs of passive investing, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Given these opposing views, how can we tell whether monetary policy is accommodative, restrictive, or just right? The answer lies in comparing the current real Federal Funds Rate with the neutral rate of interest, the unobservable, inflation-adjusted Federal Funds Rate that neither stimulates nor restrains the economy, often referred to as r-star.1

Identifying the neutral rate of interest is crucial because it serves as a benchmark for judging whether monetary policy is stimulating or restraining the economy. But estimating r-star is easier said than done, and much harder to get right in real time.

The Elusive r-star

Estimating r-star has long challenged economists and policymakers alike. By definition, the neutral rate cannot be directly observed; it must be inferred from the data.

Over the years, economists have constructed increasingly sophisticated models in an attempt to estimate r-star. The best-known model, developed by Laubach and Williams (2003) and later extended following the Covid-19 pandemic by Holston, Laubach, and Williams (2023), attempts to pin down the neutral rate empirically using data on interest rates, output, and inflation. Other models add factors like demographics, productivity, or risk preferences.

Estimates of the neutral rate vary widely across models and shift significantly over time (Figure 1). One of the main sources of time variation in estimates of the neutral rate, as pointed out by Laubach and Williams (2003), is the growth rate of potential output appears to vary over time, and potential output itself is unobservable and can only be inferred from the data.

The disagreement between models and their variance over time makes it difficult to employ them with any confidence in practice. Indeed, even John C. Williams, President of the New York Fed and the “Williams” in Laubach and Williams (2003) and Holston, Laubach, and Williams (2023), cautions against taking the estimates derived from these models too seriously:

“Although these estimates are useful for analysis and transparency, policymakers are well advised to avoid placing too great confidence in precise estimates of the stars in making real-world assessments and decisions. Given the wide range of uncertainties, acting as if one knows the star variables when making policy can lead to persistent deviations of inflation from the target that risk unmooring inflation expectations.” John C. Williams, August 25, 2025

From Theory to Practice

Instead of estimating the neutral rate with a statistical model, we can just ask whether the economy is running hot, cold, or just right based on publicly-available economic data. This is not dissimilar to what the Fed appears to do in practice based on its public-facing communication.

The Fed’s dual mandate of stable prices and maximum employment define what “just right” looks like. If unemployment is near its natural rate (around 5 percent) and inflation near 2 percent, then the inflation-adjusted Federal Funds Rate is likely close to the neutral rate.2

The challenge, of course, is that the economy is inherently dynamic—it’s constantly changing over time. It is only when we encounter a period of relative economic stability that determining whether the neutral rate is above, below, or very close to the current inflation-adjusted policy rate is feasible.

To this end, the neutral rate is simply the inflation-adjusted policy rate that keeps inflation and unemployment near their targets. If the economy is running hot, with inflation above 2 percent and unemployment below 5 percent, then the neutral rate must be above the current rate. Conversely, if the economy is running cold, with inflation below 2 percent and unemployment above 5 percent, then the neutral rate must be below the current rate.

The Bullseye Analogy

To visualize this concept, consider the classic Phillips Curve, which plots the relationship between unemployment and inflation. Historically, as one falls, the other tends to rise, producing a downward-sloping curve. Now layer concentric circles around the Fed’s 2 percent inflation and 5 percent unemployment targets (Figure 2). The inflation-adjusted Federal Funds Rate that is able to maintain the economy within the innermost of the concentric circles (i.e., the bullseye) is the neutral rate.

This practical approach to estimating the neutral rate doesn’t rely on a model, or set of restrictive assumptions. It simply identifies where the economy stands relative to the Fed’s targets. The implications for the neutral rate then follow directly.

Implications for the Current Policy Debate

Figure 3 plots the economy’s path in inflation and unemployment space since late 2021. In the wake of the pandemic, inflation rose well above the Federal Reserve’s 2 percent target while unemployment fell well below 5 percent. The rate hikes of 2022 and 2023 gradually brought inflation down toward the Fed’s goal. During this period, unemployment edged higher but has remained comfortably under 5 percent.

Thinking about policy through this lens makes it easier to interpret where we stand. When inflation and unemployment both sit near their targets, the economy is effectively at the bullseye in the diagram. Deviations in either direction signal that policy may be too loose or too tight. The value of this framework is that it lets us judge policy not by unreliable estimates of r-star, but by outcomes we can actually observe in the data.

As Figure 3 shows, inflation and unemployment have both remained within a relatively tight band since late 2023. After nearly two years of stability, the economy looks remarkably balanced. This is precisely the environment in which to ask whether the neutral rate is above, below, or roughly equal to the current inflation-adjusted policy rate.

At present, core PCE inflation is running near 2.9 percent year-over-year, unemployment is 4.3 percent, and the Atlanta Fed’s GDPNow estimate puts third-quarter real GDP growth around 3.8 percent. By most measures, the economy remains strong and well above the Federal Reserve’s long-run targets.

These numbers suggest that the current stance of monetary policy is not especially restrictive; the neutral rate is likely equal to, or perhaps even higher than, the current inflation-adjusted Federal Funds Rate. In other words, what the Fed considers tight might be neutral, or even loose, when judged through the lens of this framework.

Despite the economy’s strength, the Federal Reserve has begun easing. At its most recent meeting, policymakers trimmed the Federal Funds Rate, perhaps out of concern that the labor market is nearing a tipping point where unemployment could begin to rise sharply. That fear is understandable, and the risk of overshooting is real.

But this concern is not currently supported by the data. In my previous post on the “Four Horsemen” of the labor market, I showed that key rate-based indicators of labor market health—job finding, quits, layoffs, and unemployment duration—continue to signal labor market strength, not weakness.

Economists and Fed officials may continue to debate where r-star lies, but the economy has already given its answer. For nearly two years, inflation has remained just north of the Fed’s target, unemployment has stayed low, and economic growth has been robust. By these measures, policy is not too tight. If anything, it is slightly too loose, suggesting that the neutral rate is likely close to, or perhaps even higher than, the current inflation-adjusted Federal Funds Rate.

Notes and Sources

AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational take on the neutral rate of interest, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

But the data tell a different story.

Since 1950, labor’s share of Gross Domestic Income (GDI) has declined steadily, falling from about 66 percent to 58 percent today, a drop of 8 percentage points.

This decline helps explain a familiar sentiment on Main Street: the economy keeps growing, but the typical worker isn’t feeling much better off.

There is perhaps no clearer way to see this divergence than to compare growth in real GDP per capita with growth in real labor income. Between 1980 and 2025, real GDP per capita more than doubled, while median real weekly earnings for workers aged 16 and over barely increased. This growing gap between output and earnings mirrors the long-run decline in labor’s share.

If output is rising yet workers’ share of the pie is shrinking, it is easy to conclude that corporate profits must be growing at labor’s expense.

It’s an intuitive narrative, but it is not quite right.

When we break GDI into its components—labor compensation, proprietors’ income, corporate profits, interest, taxes, and consumption of fixed capital—we find that corporate profits have indeed risen, but only modestly. Their share of total income has increased from about 10 percent in the 1950s to roughly 13 percent today, capturing just 3 of the 8 percentage points lost by labor.2

So where did the rest of labor’s share of income go?

The answer lies in consumption of fixed capital, what economists call depreciation. Depreciation’s share of GDI climbed from roughly 11 percent in the 1950s to about 16 percent today, accounting for the remaining 5 percentage points of labor’s loss.

Importantly, depreciation is not income to capital holders. It’s an expense that represents the implicit investment needed to replace or repair broken, worn-out, or obsolete capital.

Rising depreciation reflects an economy whose capital stock is both larger and wears out faster than before. This is especially true in an era of high-tech investment, from servers and GPUs to data centers that become obsolete in years rather than decades. The rise of intellectual property capital—software, databases, and research and development—reinforces this trend. These high-tech and intangible assets depreciate far faster than traditional structures or machinery, so a growing share of income now goes toward maintaining and replacing capital rather than enriching capital owners or lifting real wages.

This observation changes the narrative. Capital owners are not capturing all of labor’s lost share. This nuance matters. It suggests that the tension between Wall Street and Main Street is not simply about profits versus wages. It is about how production has shifted toward capital that is shorter lived and, therefore, depreciates more quickly.

If labor’s share were constant, workers would benefit proportionally from growth. But today, more of what we produce each year goes toward repairing or replacing yesterday’s capital rather than lifting real wages.

A large body of research has sought to explain why labor’s share of income has fallen since the 1950’s. Two of the most influential explanations come from Thomas Piketty’s Capital in the Twenty-First Century and Loukas Karabarbounis and Brent Neiman’s The Global Decline of the Labor Share (2014).

Piketty argues that the share of income accruing to capital rises when its return exceeds the rate of economic growth. In that environment, wealth accumulates faster than output and income naturally shifts toward capital owners. Karabarbounis and Neiman take a different approach. They show that the relative price of investment goods has fallen dramatically since the 1980’s, making capital cheaper and encouraging firms to substitute capital for labor. If capital and labor are easily substituted, this shift can lower labor’s share even as productivity rises.

Both explanations fit the broad narrative that technology and capital deepening have tilted the balance of income away from workers. But they treat capital’s share as a single block that moves in opposition to labor’s. The decomposition here adds an important piece of the puzzle. When we examine GDI in detail, we find that most of the decline in labor’s share reflects the growing cost of maintaining the capital stock rather than an increase in income flowing to owners of capital.

Much of what looks like redistribution from labor to capital may instead reflect a structural shift in the nature of capital itself. Modern production relies on assets that are more technologically sophisticated but also shorter-lived. Servers, GPUs, and intellectual property wear out or become obsolete in just a few years. Their rapid depreciation shows up as a rising expense share in the national accounts, reducing the income available to raise wages or increase returns to capital owners.

In this sense, what looks like redistribution is really a story about the changing composition of productive assets and the costs of sustaining a more technologically advanced economy. It helps explain why the economy can expand while much of the population feels left behind.

In the end, the mystery of the missing labor share is less about rising profits and more about the hidden cost of keeping our ever-expanding, high-tech economy running.

Notes and Sources

AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational take on the falling labor share of income, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

Putting Credit Card Debt in Context

$1.3 trillion in credit card debt sounds staggering. But in the broader balance sheet of U.S. households, credit cards remain a relatively small slice. Mortgages account for about 70% of total household liabilities, auto loans and student loans each account for roughly 9%, while credit cards represent just 7% (Figure 1).

Credit card debt is, however, a useful barometer of household health because of its sensitivity to changes in economic conditions (as I have documented in my own work on the subject). When financial stress shows up, it usually shows up here first. Why? Because consumers are more likely to default on credit cards before walking away from their homes, their cars, or their student loans (which risk wage garnishment). That makes credit card debt an early-warning indicator of household financial strain.

The Post-2021 Run-Up

Since early 2021, revolving credit card debt is up 38% (see Figure 2). Defaults and charge-offs have also increased from their pandemic-era lows. Taken at face value, this might seem ominous.

But this framing misses a crucial point: debt levels can only be understood relative to income. This is because income determines a household’s ability to service their debt.

Debt Looks Different Relative to Income

Disposable personal income also rose sharply since 2021, keeping pace with the rise in credit card debt. As a result, the ratio of credit card debt to disposable income has barely budged since 2021. Today it sits at 0.057, a level not seen since the early 1990s outside of the Covid recession and immediate aftermath.

Households are indeed carrying a higher debt load, but they also have more income to service that debt. From this perspective, households are in pretty good shape.

Defaults and Charge-Offs

What about the stress signals? It’s also true that both default and charge-off rates have risen sharply since 2021 and are now above pre-Covid levels. But they remain far below past periods of distress. During the 2001 recession, both measures were materially higher. During the financial crisis of 2008–09, both measures spiked dramatically. By comparison, today’s levels look relatively tame. Moreover, the trend over the past six months has been generally downward.

No Reason to Panic

Put together, the data tell a far less worrying story than the headlines suggest:

• Credit card balances have risen sharply since 2021, but so has disposable income.

• Credit card debt relative to income is historically low.

• Default and charge-off rates are elevated compared to the ultra-low pandemic era, but well below past recessionary peaks.

The Bottom Line

Near-record credit card debt in dollar terms doesn’t mean households are financially strained. By the measures that matter, like debt relative to income and the trajectory of defaults and charge-off rates, households look to be in solid shape.

Notes and Sources

AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational take on credit card debt and the financial health of U.S. households, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths.

No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

In intro macro we teach the quantity theory of money as a way to think about the impact of changes in the money supply on the average price level over the long run:

M v = P Y,

where

• M: money supply, typically measured using M2

• v: velocity of money, or the average number of times each dollar is spent per year

• P: average price level in the economy, typically measured using the CPI

• Y: real GDP

The quantity theory of money is less a grand theory than an accounting identity. On the left, we have the total amount of dollars circulating in the economy (M) multiplied by the number of times each dollar is spent per year (v), which is equal to nominal GDP. On the right, we have real GDP (Y) multiplied by the average price level (P), which is also equal to nominal GDP. Since there is only one value of nominal GDP, both the left and right hand sides must yield the same result.

We can measure the money supply (M), average price level (P), and real GDP (Y), but we cannot observe the velocity of money (v) directly. The quantity theory of money is thus used to impute the value of v such that both sides of the equation match. In other words, v is chosen to make the identity hold.

The key insight comes from rewriting the equation in growth rates (by taking logs and differentiating with respect to time):

g_M + g_v = g_P + g_Y,

where g_x is the growth rate of variable x. Note that g_P is the growth rate of the price level, or inflation.

In the short run, money creation can stimulate demand. But in the long run, printing money cannot generate sustained growth in real GDP. So the first assumption we make is that money growth does not cause real GDP growth in the long run.

The second assumption is that the velocity of money is stable over time (i.e., g_v= 0). This assumption is not trivial. After the 2008–09 financial crisis, many expected rapid money growth to generate inflation, but velocity collapsed instead. The same thing happened during the first few months of the Covid-19 recession. But in this case velocity recovered relatively quickly back to its pre-Covid level, making the assumption of stability a reasonable simplification for this period.

Under these two assumptions, the theory implies that in the long run, money growth equals inflation plus real GDP growth:

g_M ≈ g_P + g_Y.

That’s the scaffolding we’ll use to evaluate what happened since 2020.

Implications for Monetary Policy

To fix ideas, suppose the money supply is constant (g_M = 0). Then if real GDP grows, more goods are being produced but there are no new dollars to buy them, so the average price level must fall (g_P < 0). In other words, we get deflation.

Most modern central banks actively try to avoid deflation by growing the money supply (g_M > 0). If the Fed sets g_M equal to g_Y, then real GDP can grow while the price level stays flat (g_P = 0).

In practice, the Fed targets 2% inflation. The theory tells us how to hit it: set money growth equal to real GDP growth plus 2%.

g_M = g_Y + 2%

Real GDP growth has averaged about 3% per year in the postwar period. So money growth of about 5% per year is consistent with hitting the 2% target. It really is that straightforward.

It seems like every day there are new data releases about the labor market, inflation, or real GDP growth. These headlines distract central bankers and market watchers from what matters most for achieving stable prices: the growth rate of the money supply. As Milton Friedman preached, “Inflation is always and everywhere a monetary phenomenon.”

Application: Covid-19 Recession and Recovery

When the pandemic hit, the Fed slashed rates to zero and bought Treasuries and MBS in unprecedented quantities. The result: the money supply exploded. The quantity theory tells us those dollars had to show up somewhere.

As shown in Figure 1, between January 2020 and July 2025 the money supply increased by 43.5% (M2SL). Over this same period, real GDP rose by just 14.5%. According to the theory, the price level should rise by:

g_P = g_M – g_Y = 43.5% – 14.5% = 29%

I call this potential inflation: how much inflation is in the pipeline if money and real growth stop here.

How does that compare to actual inflation? Between January 2020 and July 2025, the CPI increased by 24.3%. Not the full 29%, but close. Importantly, velocity today is about where it was in early 2020, meaning money growth must show up eventually in real GDP growth or inflation.

So what do we learn?

1. The quantity theory, on its own, explains the inflation we experienced since the Covid-19 pandemic. No need to invoke supply bottlenecks or “greedflation.” Inflation was never going to be transitory; it was baked in by money growth, waiting to be realized.

2. There is still some potential inflation left out there. The theory predicts 29%. We’ve seen 24.3%. That leaves about 4.7% unaccounted for. Faster growth could soak it up (see Figure 2). But unless velocity slows again, that wedge will eventually show up somewhere, either in higher prices or more output.

Implications for the Fed Now

We enter late 2025 with solid growth and low unemployment, but new tariffs are starting to push up import prices, which may kick off the potential inflation still remaining in the system.

Over the last 12 months:

• Money growth (M2): ~4.8%

• Inflation (CPI): ~2.7%

• Real GDP growth: ~2.1%

Through the lens of the identity, this mix is close to target. With a 2% inflation goal and plausible real growth of 2–3% ahead, money growth of about 5% is consistent with keeping inflation near target absent new easing. If real growth accelerates inflation should moderate toward 2%.

If money growth is already stable, easing now risks another round of persistent inflation. The Quantity Theory of Money suggests the Fed should proceed with caution.

What does this mean in practice for monetary policy going forward?

1. Anchor nominal growth. The Fed should keep money growth consistent with 2% inflation plus real growth.

2. Watch broad money. If M2 growth runs well above 5% while tariffs are raising import prices, expect inflation pressure to rebuild. As I argued in The Fed Should Ignore Tariff Induced Inflation, the Fed should not respond to a one-off tariff shock by either tightening or easing financial conditions. If the Fed does ease, it won’t be a one-off shock; it is a setup for persistent inflation.

3. Let QT keep running quietly. Passive balance sheet runoff is an important tool the Fed should use to restrain money growth without constant headlines.

Conclusion

The past five years show the enduring power of a first principles framework. The quantity theory of money was enough to predict that the surge of money printing would lead to inflation. It reminds us that over the long run, money growth and real GDP growth together determine the prevailing amount of inflation. For the Fed, the task now is simple in principle but not in practice: keep money growth aligned with real capacity, and inflation takes care of itself.

Notes and Sources

AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational take on the Quantity Theory of Money, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths. No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

To put numbers on it, the U.S. economy has grown about 3 percent per year in real terms over the past century, while real stock returns have been roughly 6 to 7 percent. If stocks are claims on the economy, how can they consistently return twice as much as the economy grows?

This puzzle prompted me to investigate the link between financial market returns and macro growth. What I found is that the solution to this puzzle is straightforward: profits grow with the economy, on average, and firms also return cash to shareholders through dividends and net buybacks. Layer this shareholder yield on top of real profit growth and it becomes clear how an economy growing at 3 percent can coexist with 6 to 7 percent equity returns over the long run.

Let’s walk through the details to see how it all fits together.

Start with the income approach to computing GDP

According to the income approach, GDP is the sum of wages, profits, rent, and interest. For investors, the profit slice matters most. If corporate profits are a stable share of GDP, then profits scale one-for-one with the economy. As the economy grows, so do the profits earned by firms.

From profits to earnings per share

Not all profits belong to the public. Only a fraction shows up in publicly listed firms. And even within listed firms, profits are divided across millions of shares.

So earnings per share (EPS) for the overall stock market depend on the size of the economy, the profit share of the economy, the public share of those profits, and how many shares are outstanding.

In theory, if the profit share of the economy, the public share of profits, and shares outstanding are stable, real EPS growth will follow real GDP growth.

In practice, if shares outstanding decline via net share repurchases, then real EPS growth will follow real GDP growth plus the average annual market-cap-weighted reduction in share count.

From earnings to returns

Returns on investment in the stock market have three main components:

1) Earnings per share (EPS) growth: real EPS growth tracks real GDP growth, plus any net reduction in share count (net buybacks).

2) Dividend yield: cash returned to shareholders divided by market value.

3) Valuation changes: the average amount investors are willing to pay for a dollar of corporate earnings.

Real return ≈ EPS growth + Dividend yield + Valuation changes

In the long run, EPS growth is equal to real GDP growth plus the net buyback yield:

Real return ≈ Real GDP growth + Buyback yield + Dividend yield + Valuation changes

If valuations remain roughly stable, then:

Real return ≈ Real GDP growth + Shareholder yield ≥ Real GDP growth

Thus, over the long run, shareholder yield (dividends + net buybacks) is what raises the real return on investment in the stock market above the real growth rate of the economy.

A simple numerical example

Suppose the public share of profits is constant and valuations are steady. Then the theory tells us that long-run real returns on investment in the stock market are equal to real GDP growth plus the shareholder yield.

If real GDP grows at 3 percent, the dividend yield is 3 percent, and the net buyback yield is 0.5 percent, then the theory tells us that the implied real return on investment in the stock market should be about 6.5 percent (Figure 1).

How does this theory stack up against the data?

From 1950 through 2024, real GDP in the U.S. grew at about 3.1 percent per year. Dividends averaged roughly 3.0 percent per year, and net buybacks about 0.5 percent per year, which together imply a shareholder yield of approximately 3.5 percent. Over this period, the public share of profits rose modestly by about 0.3 percent per year, while increasing valuations added another roughly 0.5 percent per year on average.

Add it together, and the theory tells us that the real return on investment in the stock market during this period should have been about 7.4 percent per year (Figure 2). The realized real return on investment in the stock market over the same window was 6.9 percent (S&P 500 with dividends reinvested, CPI deflated).

The half-percentage-point gap between theory and reality reflects measurement choices rather than economics: buyback “yield” is a dollar flow while returns depend on the net change in share count, valuation effects depend on the P/E definition and endpoints, and macro profits and public listings do not line up perfectly. Within a reasonable margin of error, the theory appears to match the data quite well.

The takeaway

The stock market, both in theory and in practice, tends to grow faster than the economy. Over the long run, real equity returns are largely determined by the sum of real GDP growth and shareholder yield, with minor adjustments for changes in the profit share of GDP and the public share of profits. While dividends used to carry most of the load of the shareholder yield, today buybacks are becoming increasingly important. Regardless, the sum of the two is what matters — shareholder yield is the wedge that lets real returns exceed real GDP growth.

For those who want the full accounting, the next section walks through all of the details.

Technical Appendix

Definitions:

• GDP: real gross domestic product.

• ψ (psi): profit share of GDP earned as corporate profits.

• φ (phi): public share of corporate profits earned by listed firms.

• Shares: total shares outstanding.

• Dividend yield: cash dividends over market value.

• Net buyback yield: net share repurchases over market value, net of issuance for compensation and M&A (negative when share count rises).

• Shareholder yield: dividend yield + net buyback yield.

• Valuation multiple: price-to-earnings ratio (TTM).

Aggregate earnings captured by publicly listed firms:
(1) E = ψ × φ × GDP

Taking logs and differentiating:
(2) g_E = g_GDP + g_ψ + g_φ

where g_X denotes the real growth rate of X.

Returns decomposition:
(3) Real return ≈ g_EPS + Dividend yield + g_Valuation

where g_Valuation is the contribution from changes in the valuation multiple (i.e. the log change in P/E).

Earnings per share (EPS):
(4) EPS = E / Shares

Taking logs and differentiating:
(5) g_EPS = g_E – g_Shares

where g_Shares is the growth rate of share count (positive when shares increase, negative when they shrink).

Substitution into returns:
(6) Real return ≈ g_E – g_Shares + Dividend yield + g_Valuation

Since –g_Shares is equivalent to the net buyback yield, combining with dividends gives total shareholder yield:
(7) Real return ≈ g_E + Shareholder yield + g_Valuation

Substituting aggregate earnings growth from (2):
(8) Real return ≈ g_GDP + g_ψ + g_φ + Shareholder yield + g_Valuation

Long-run steady state:
If profit shares (ψ) and public shares (φ) are constant, and valuations are stable, then:
(9) Real return ≈ g_GDP + Shareholder yield

Conclusion:
Over long horizons, the real return on equities will exceed real GDP growth by the shareholder yield — which has averaged about 3.5 percent during the post‑WWII period.

Notes and Sources

Real equity returns and real GDP growth are measured in real terms using CPI as the primary deflator; results are similar with PCE. Realized equity returns refer to the S&P 500 total return with dividends reinvested, deflated by CPI, using Robert Shiller’s publicly available dataset (monthly prices, dividends, earnings, and CPI). The 1950–2024 realized real return cited here is the compound annual rate over that window. Shareholder yield is defined as dividends plus net buybacks; buybacks are measured net of issuance for compensation and M&A so that the metric aligns with changes in share count. EPS growth is based on aggregate earnings captured by publicly listed firms divided by shares outstanding. The valuation term reflects the log change in the trailing twelve-month price-to-earnings multiple. Minor differences can arise from start and end dates, deflator choice, and treatment of corporate actions; none of these alter the main conclusions. AI tools were used to help transform raw data into figures and to edit prose; all calculations are straightforward to reproduce from the cited sources.

If you enjoyed this mildly efficient and occasionally rational explanation of how the real return on investment in the stock market can exceed the growth rate of the real economy over the long run, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths. No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is an Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.

What Families Actually Pay

In 2024–2025, the sticker price at this institution was about $84,000. Yet the average family paid closer to $48,000, a 40% markdown.

Why? Roughly 55% of students received need‑based financial aid, and the average grant was about $67,000, leaving aided families with an average bill of just $17,000, an 80% discount off the sticker price.

The remaining 45% of families paid full price (~$84,000).

But not only do aided families pay significantly less than the sticker price, over the last decade they saw their average bill fall ~1.4% per year, even as the sticker price rose ~4% per year (Figure 1).

This trend runs counter to the popular narrative that higher education is becoming increasingly unaffordable. For low income families at this elite institution (and many of its peers), the cost of college has been falling, not rising.

The average bill for aided families is being driven down by a secular increase in the average discount rate (% off the sticker price), which climbed from roughly 68% in 2015 to more than 80% by 2024 at this particular institution (Figure 2).

Why has the discount rate been rising? The publicly available data make it difficult to tell for sure. Most likely, though, it is due to a mix of (1) more generous financial aid packages and (2) shifts in ability to pay among aided families.

Regardless of the cause, the result is a widening gap between what families of students who qualify for financial aid and those who don’t are asked to pay. In 2015, aided families paid about $40,000 less than unaided families. By 2024, this gap was roughly $67,000, a 68% increase in only a decade. Like many of its peers, at this elite institution there is an increasing level of price dispersion between what aided and unaided families pay for their child’s college education.

Subsidies, for Everyone!

But that’s not the whole story. What is perhaps more surprising is that the sticker price, even at an eye-watering $84,000, is not even close to covering the full cost the college incurs to educate, house, and feed its students. Indeed, during the 2024–2025 academic year, per-student operating expenses topped $115,000. This means that even full-pay families received an implicit subsidy of ~27% off the true cost (Figure 3).

How can this institution afford to subsidize everyone? Revenue from the comprehensive fee covers about 42% of operating expenses. Income from the college’s multi-billion dollar endowment covers the vast majority of the remainder. By 2024, endowment income alone covered nearly half of per-student costs (Figure 4).

Gifts and other sources of revenue, while modest, are just enough to fill the rest of the gap not covered by endowment income and revenue from the comprehensive fee (“Family Bill” in Figure 5).

The funding gap between operating costs and what families, on average, actually pay is driven mainly by two factors: (1) the increasing discount rate (see Figure 2) and (2) competitive limits on sticker price increases.

Importantly, the funding gap is not the result of expenses spiraling out of control. At the institution in question, operating expenses have grown by about 2.5% per year over the last decade, slightly below economy‑wide inflation.

Key Takeaway

Two truths can coexist. For full-pay families, the price has risen steadily and is nearing six figures. For aided families, bills are much lower and have fallen over time, widening access to a top-tier liberal-arts education. Yet even full-pay families do not cover the true cost. The question is whether this model is sustainable as competition for college-bound students pushes discount rates higher and federal policy changes threaten endowment income and research funding.

Methods and Sources

All data reported in this post are based on publicly available information. Dollar amounts are nominal. Expense and revenue items are per-student unless noted. All inflation references refer to PCE. AI tools were leveraged by the author to construct some of the figures from raw data and to fine-tune the narrative.

If you enjoyed this mildly efficient and occasionally rational discussion of the real cost of a top-tier liberal arts college education, consider subscribing below. We’ll continue exploring markets and models, revealing mildly surprising truths. No hot takes; just thoughtful ones.

Subscribe now

About the Author: Seth Neumuller is Associate Professor of Economics at Wellesley College where he teaches and conducts research in macroeconomics and finance. He holds a Ph.D. in economics from UCLA. His Substack is Mildly Efficient (and Occasionally Rational) where he explores topics in finance and macro from first principles, cutting through complexity with clear, grounded analysis.