For the last several years, the largest technology companies have insisted that the extraordinary sums they are spending on artificial intelligence are justified by extraordinary demand. Perhaps they are right. AI may well transform search, advertising, software development, cloud computing, medicine, education, and much else.

But that is not the question investors should be asking.

The relevant question is whether the companies financing the transformation will earn an adequate return on the capital they are investing. Those are two very different propositions. Railroads transformed the nineteenth-century economy, airlines transformed the twentieth, and the internet transformed the twenty-first. In each case, society captured enormous benefits while many of the companies funding the transformation earned disappointing returns.

Artificial intelligence may be following the same path. Google’s latest announcement suggests that the process is accelerating.

Google raises the stakes

On July 22, Alphabet raised its estimate of 2026 capital expenditures to between $195 billion and $205 billion, $15 billion above its previous guidance. Management also indicated that spending could rise further in 2027.

The operating results were strong. Alphabet’s revenue increased 24 percent, while Google Cloud revenue surged 82 percent to $24.8 billion. This is not a story about a company investing into a collapsing business. Demand is clearly growing.

Nonetheless, the financial consequences of the investment program were equally striking. Alphabet spent $44.9 billion on capital expenditures during the quarter, compared with $39.1 billion of operating cash flow. The result was negative free cash flow of $5.9 billion—the first negative quarterly free-cash-flow figure in Alphabet’s history as a public company.

Investors reacted by marking down the stock. The important point, however, is not the one-day market reaction. It is the strategic response that Google’s commitment is likely to provoke.

Amazon, Microsoft and Meta cannot simply watch Google expand its computing capacity, accelerate Google Cloud and improve Gemini while they conserve cash. If they do, they risk losing cloud customers, developers, AI talent and, ultimately, their positions in markets that support trillion-dollar valuations. Each company therefore has a powerful incentive to match or exceed the spending of the others.

The battle has been joined.

The prisoner’s dilemma

This is a classic prisoner’s dilemma.

Collectively, the major technology companies might be better off if all of them moderated their capital spending. They could develop AI at a more measured pace, preserve free cash flow, avoid redundant infrastructure and allow demand to catch up with capacity. Industry profits would probably be higher.

But no company can safely adopt that strategy on its own.

If Google spends aggressively while Microsoft restrains itself, Google may gain an important advantage. If Microsoft spends while Google restrains itself, Microsoft may gain the advantage. The same logic applies to Amazon and Meta. As a result, aggressive investment becomes the rational strategy for every individual company even if the ultimate outcome is lower returns for the group as a whole.

That is the defining feature of the prisoner’s dilemma: individually rational behavior produces a collectively inferior result.

The numbers are becoming staggering. Microsoft has indicated that its 2026 capital expenditures may approach $190 billion. Amazon is expected to spend about $200 billion. Meta has raised its guidance to between $125 billion and $145 billion. With Alphabet now guiding to approximately $200 billion, the four companies together are on course to spend more than $700 billion in a single year.

That is not merely a technology investment cycle. It is an industrial mobilization.

And it is unlikely to end with the current budgets. Once one company expands capacity, trains a larger model or cuts the price of an AI service, the others must respond. Today’s investment creates tomorrow’s competitive requirement. The race feeds on itself.

Good news can make the problem worse

The obvious response is that Google’s Cloud growth demonstrates that the investment is paying off. An 82 percent revenue increase is hardly evidence of weak demand.

That is true, but it misses the economic issue.

Rapid growth can justify investment by one company. It does not prove that simultaneous investment by every major competitor will produce attractive returns for all of them. Indeed, strong demand may intensify the prisoner’s dilemma by persuading each firm that it must build even faster.

The resulting capacity is likely to place continuing pressure on prices. AI models are already becoming more capable and less differentiated. Open-source systems are improving. Customers can shift workloads among cloud providers, use smaller models or decide that a cheaper product is good enough. Meanwhile, the owners of the infrastructure must absorb depreciation, energy costs, maintenance, chip obsolescence and the cost of replacing equipment whose economic life may be far shorter than its accounting life.

The result could be spectacular growth in AI usage accompanied by mediocre returns on AI capital.

That distinction is critical. Revenue is not return. Growth is not value creation. An investment creates value only if the resulting cash flows exceed the cost of the capital committed to produce them.

From asset-light to capital hungry

For most of the last two decades, the great attraction of the dominant technology companies was that they combined rapid growth with unusually high free-cash-flow conversion. Their most valuable assets—software, networks, brands, data and intellectual property—did not require the continuous construction of factories on the scale associated with traditional industrial businesses.

AI is changing that model.

The new competitive unit is increasingly the data center: chips, networking equipment, cooling systems, power contracts and enormous supporting facilities. The Magnificent Seven are beginning to look less like the asset-light platforms investors learned to love and more like capital-intensive utilities locked in a technological arms race.

Accounting earnings will be slow to reveal the full transition because the cost of the infrastructure is depreciated over time. Free cash flow reveals it immediately. Alphabet’s negative $5.9 billion quarter is therefore more than an accounting curiosity. It is a warning that the economics of Big Tech are changing.

The broader trend is equally concerning. A Reuters analysis of consensus estimates found that Microsoft, Alphabet, Amazon, Meta and Oracle could collectively spend more on capital expenditures than they generate in free cash flow by 2027. The increase in capital spending between 2025 and 2027 is projected to equal about $1.57 for every additional dollar of operating cash flow.

That is not the free-cash-flow machine on which current technology valuations were built.

Who wins?

The most likely answer is that consumers win. Businesses that use AI may also win as models improve and prices fall. Suppliers of scarce inputs—advanced chips, memory, networking equipment, power and specialized construction—can profit while the buildout continues.

It is much less clear that the companies funding the race will all win.

One or two may establish durable advantages. Google has formidable assets: proprietary chips, a leading cloud platform, unmatched data, a global distribution system and the cash-generating power of search. Microsoft has its enterprise franchise and partnership ecosystem. Amazon has AWS. Meta has billions of users and an advertising engine capable of monetizing small improvements at enormous scale.

But those strengths do not eliminate competition. They are the reason the competition is so intense.

Nor is it necessary for the AI boom to collapse for investors to be disappointed. The technology can succeed, demand can grow rapidly and revenues can rise—and shareholders can still earn substandard returns if too much capital chases the opportunity. In competitive markets, the benefits of innovation are often competed away and transferred to customers.

The investment conclusion

Google has now made clear that it will not conserve cash and hope the AI race develops gradually. It intends to build. Its competitors will almost certainly do the same.

Each company can defend its decision. Given the strategic stakes, each decision may be entirely rational. But investors own the collective outcome, not the logic of an individual boardroom.

The central question is no longer whether artificial intelligence will work. It is whether several companies can simultaneously spend hundreds of billions of dollars, construct overlapping infrastructure, compete aggressively on price and still earn returns above their cost of capital.

The prisoner’s dilemma has come home to roost.

Google has made its move. Now everyone else has to answer.

Not long ago, investors were anticipating three potential trillion-dollar IPOs: OpenAI, Anthropic, and SpaceX.

SpaceX—by far the largest of the three—made it out the door on June 12, 2026, at a market capitalization of $1.75 trillion. At the time, I argued in my June 14, 2006 Substack that the valuation was rich. The stock is now trading below its IPO price.

OpenAI and Anthropic face a more fundamental problem. In my view, neither company can go public at anything approaching a $1 trillion valuation. Even half that amount may be a bridge too far.

The reasons are straightforward.

The margins are not there

A trillion-dollar valuation requires the prospect of very large and durable profits. That, in turn, requires customers willing to pay prices that cover the enormous cost of developing models and building data centers—with a healthy margin left over for the producer.

It is far from clear that enough such customers exist.

Citrini Research recently described a shift from “tokenmaxxing” to “tokenpanic.” Once every prompt carried a visible price, corporate procurement departments began asking the question that eventually undermines premium pricing in every maturing market:

Is the cheaper alternative good enough?

Increasingly, the answer appears to be yes.

Most customers do not need the world’s most capable model for every task. They need a model that performs the job reliably and inexpensively. As the performance gap narrows, the premium commanded by frontier models becomes progressively harder to defend.

Competition will erode whatever margins emerge

Even if OpenAI and Anthropic eventually produce attractive margins, those margins must remain stable for many years to justify valuations approaching $1 trillion.

That seems increasingly unlikely.

In the early days of generative AI, it was reasonable to believe that the immense capital required to train frontier models would create a formidable barrier to entry. The surge of capable, low-cost, open-weight models has steadily demolished that assumption.

As I wrote in the latest Cornell Capital Group quarterly memo:

“The past eighteen months have steadily undermined that idea. Chinese open-weight models—DeepSeek and Qwen among them—now handle the majority of commercial workloads at prices 7 to 25 times below the U.S. frontier labs. A capability that commands a premium today is matched by a model costing a fraction as much within a year.”

Today’s release of Kimi K3 suggests that I may have understated the competitive threat.

The problem is not simply that OpenAI and Anthropic have competitors. Every valuable business does. The problem is that model capabilities are diffusing rapidly while prices are falling. Each new advance by a frontier laboratory is soon replicated, approximated, or made available more cheaply elsewhere.

That is not the traditional recipe for durable monopoly profits.

A great product is not necessarily a trillion-dollar company

OpenAI and Anthropic may continue to build extraordinary products. They may become large and successful public companies. But extraordinary technology does not automatically produce extraordinary shareholder returns—particularly when competition is intense, capital requirements are enormous, and pricing power is uncertain.

A trillion-dollar IPO requires investors to believe not merely that a company will grow rapidly, but that it will eventually convert that growth into immense, defensible profits.

For OpenAI and Anthropic, that belief is becoming increasingly difficult to sustain.

The trillion-dollar AI IPO, once treated as a reasonable possibility, now looks more like a relic of peak enthusiasm. SpaceX made it through the window.

For everyone else, the window has closed.

Perhaps. But those are still very large numbers. For Anthropic to meet them, customers have to be willing and able to pay for the services. And when people say AI will “transform the economy,” what they usually mean is that AI will increase output and productivity by enough to justify those payments.

But there is a catch. Where, exactly, will customers capture those gains in their own businesses?

The most obvious answer is labor substitution: replacing expensive human employees with faster, cheaper, and more efficient AI agents. If that is the path, however, it raises a deeper problem. In that scenario, the people ultimately financing the AI boom are the workers whose jobs are being displaced by it.

If displacement occurs on a scale large enough to support the projections in Claude’s model, the social consequences are hard to ignore. A portion of the population would, in effect, be paying for the AI boom through the loss of their employment.

That may be profitable for the AI companies and their customers, but it is not a stable foundation for a broad-based economic transformation. If AI-driven labor displacement occurs on a large scale, society will need to develop a new method for distributing wealth fairly. Otherwise, the gains from AI may accrue primarily to the owners of the technology and the capital behind it, while the costs are borne by workers who no longer have a clear role in the production process.

To get a better sense of the challenge, I asked Claude to build a five-year discounted cash flow model for its parent company, Anthropic. The goal was simple: identify the operating assumptions required to justify Anthropic’s stated valuation of $950 billion.

The results presented below are dramatic.

In Claude’s model, revenues have to increase by a factor of ten over five years. They also have to do so while gross margins rise. The implied compound growth rate in both revenue and earnings is roughly 80% per year. Even more striking, the model assumes that this growth can be achieved with essentially no meaningful increase in capital expenditures. That is an extremely optimistic assumption.

The deeper problem is that someone has to pay for all of this. Anthropic’s revenues are its customers’ costs. As more open-source alternatives become available, how many customers will be willing to pay premium prices for Anthropic’s products? And for how long?

The sensitivity of the valuation to modest changes in the assumptions is sobering. If the model is adjusted simply to hold gross margins constant and reduce the revenue growth rate to 50%, Anthropic’s valuation falls dramatically, to roughly $200 billion.

These are only back-of-the-envelope calculations and should not be taken as investment advice. But they do provide useful insight into the scale of the assumptions required to justify current valuations for companies like Anthropic. The numbers do not merely require strong performance. They require extraordinary performance, sustained over many years, in a market that is likely to become increasingly competitive.

But extraordinary companies can still become poor investments when the price requires too much of the future.

In that regard, I thought Professor Aswath Damodaran’s valuation of SpaceX was generous but within the bounds of reason. His estimate produced an equity value of approximately $1.3 trillion, or about $98 per share. That conclusion already required favorable assumptions about the launch business, Starlink, future growth opportunities, margins, and reinvestment efficiency.

In other words, $98 per share was not a skeptical bear case. It was a constructive valuation of a remarkable company.

Where the Valuation Gets Stretched

My greatest concern was the value assigned to xAI. Artificial intelligence is a vast potential market, but it is also one of the most competitive markets in the world. The leading firms already possess enormous advantages in data, distribution, infrastructure, talent, capital, and customer relationships.

Against that backdrop, xAI appears to be playing catch-up rather than leading the field. To assign very large value to that business requires not only belief in the size of the AI market, but also belief that xAI will capture a meaningful and profitable share of it despite formidable competitors. That may happen. But it is far from assured.

Even accepting Damodaran’s optimistic valuation, the IPO price was difficult to justify. At $135 per share, SpaceX was valued at approximately $1.75 trillion, roughly 35 percent above the estimated intrinsic value. When the shares began trading and closed the first day at $161, the implied equity value rose to approximately $2.1 trillion. That was roughly 70 percent above Damodaran’s estimate and more than 25 percent above the IPO price.

At that level, SpaceX is appears to be uninvestable.

This Is Not a Criticism of SpaceX

The reason is not that SpaceX is a weak company. It is because the market price appears far in excess of what can reasonably be supported by expected cash flows.

To justify a $2.1 trillion valuation, investors must assume not only that SpaceX will dominate launch, scale Starlink successfully, earn attractive margins, and reinvest capital efficiently, but also that its newer and less-proven ventures will generate enormous incremental value. That is a narrow path. The margin for error is small. The valuation leaves little room for disappointment.

The missing piece in a conventional discounted cash flow analysis is what might be called the Elon effect.

For more than a decade, Tesla has traded at a premium to other automobile manufacturers because investors were not merely valuing car production. They were valuing Elon Musk’s promises about autonomy, robotics, energy, artificial intelligence, and a range of future businesses that were difficult to model but easy to imagine. In effect, the market capitalized Musk’s ability to expand the story.

The same phenomenon now appears to be operating at SpaceX. The market is not simply discounting launch revenues, Starlink subscriptions, operating margins, and reinvestment needs. It is discounting the possibility that Musk will do something remarkable that is not yet fully visible.

That is a powerful narrative. It is also a dangerous one.

The Value of Optionality Is Not Infinite

The problem is not that such optionality is worthless. Elon Musk has repeatedly shown an ability to take seemingly impossible projects and turn them into commercial reality. SpaceX itself is evidence of that. A valuation model that gives no credit to entrepreneurial option value would almost certainly understate the company.

But option value is not infinite. At some price, the market moves from paying for identifiable cash flows plus reasonable optionality to paying for imagination itself. We believe the first-day closing price crossed that line.

At that price, the valuation no longer appears anchored in what SpaceX can plausibly earn. It appears anchored in what investors hope Elon Musk might eventually create.

This is the hallmark of a frothy market. In such markets, investors become increasingly comfortable paying today for businesses that are expected to emerge tomorrow. The more successful the entrepreneur, the more expansive the story becomes. Valuation discipline gives way to narrative momentum. Instead of asking, “What cash flows are we buying?” investors begin asking, “What if this becomes the next trillion-dollar opportunity?”

That reasoning can persist for some time. Frothy markets do not collapse simply because valuations are high. They collapse when the story stops expanding, when capital becomes more expensive, when growth disappoints, or when investors realize that the cash flows required to justify the price are too far in the future and too uncertain.

The risk is not that SpaceX suddenly becomes a bad company. The risk is that it remains a very good company but fails to satisfy the extraordinary expectations embedded in the stock price.

A Great Company Can Still Be a Terrible Investment

That distinction is crucial. Great companies can be terrible stocks if purchased at prices that already assume greatness and then some.

At $98 per share, one could argue that SpaceX was richly but reasonably valued. At $135, the IPO price already stretched the investment case. At $161, the valuation seemed to detach from fundamental support.

My conclusion is therefore straightforward. SpaceX may be one of the most impressive companies of the modern era. It may continue to reshape launch, communications, and possibly other markets. But at a $2.1 trillion equity value, the stock requires investors to underwrite a future that is not merely optimistic, but heroic.

I do not believe that is an attractive risk-reward proposition. The company is remarkable. The price is more remarkable still.

By comparison, an Apple Watch costs roughly $400 and does just about everything. I have owned one for some time and still do not know all its features. It monitors health, receives messages, makes calls, gives directions, tracks workouts, plays music, and, yes, keeps essentially perfect time.

The contrast is striking. Apple Watch Series 11 sales to date are roughly 12–16 million units worldwide. Patek Philippe’s annual sales of all Calatrava models combined are probably closer to 10,000. The interesting question is not why Apple sells so many more watches. That is obvious. The interesting question is how Patek can sell 10,000 watches a year at nearly 100 times the price of an Apple Watch and still maintain a waiting list.

Part of the answer is materials, scarcity, and brand prestige. But there is a deeper element to the story. The answer is that the Calatrava embodies a centuries-long tradition of human mechanical artistry. It is not valuable because it is the most efficient way to tell time. It is valuable because it represents a level of craft, history, and human skill that cannot easily be replicated. Some people are willing to pay a great deal for objects that carry that meaning.

That distinction may become increasingly important as we move deeper into the world of AI. AI will be faster, cheaper, more knowledgeable, and more available than human beings in a growing number of domains. In many cases, that will be enough. But pockets of value are likely to remain — and perhaps even grow — around distinctively human capabilities.

A teacher with a special ability to reach students may still matter, even when AI can explain every concept instantly and patiently. A doctor who develops unusual rapport with patients may still be valued, even when AI has broader diagnostic knowledge. An advisor, coach, writer, therapist, or leader may succeed not merely because of technical competence, but because of trust, judgment, empathy, taste, and presence.

Those relationships work for deeply personal reasons. That cannot be reduced to information processing. That does not mean AI will be unimportant in such settings. Quite the opposite. AI will reshape them. But the adoption and acceptance of AI will involve an interesting and complicated feedback between efficiency and meaning, between capability and trust, between what machines can do and what people still want from other people.

No matter how far electronic watches advance, the Calatrava is unlikely to disappear. It may become less useful in any practical sense, but that is beside the point. Its value lies elsewhere. The same may prove true for certain forms of human skill in an AI world.

The comparison is striking. Only four markets have a total capitalization greater than one NVDA — or only three if Hong Kong is treated as part of China. In Europe, the United Kingdom has the largest equity market, but even it comes in at only about 0.75 NVDAs. Germany barely clears the 0.50 NVDA threshold.

It is not entirely clear what one should conclude from the chart, other than that it underscores the extraordinary valuation of the largest U.S. technology companies. Can a world in which a single company is worth more than most national stock markets be sustained over the long term? My instinct is to be skeptical. But perhaps that skepticism reflects a failure to fully appreciate the enormous earning power — and perceived future earning power — of the great technology platforms.

This post focuses on that gap using the Magnificent Seven — Apple Inc., Microsoft Corporation, Amazon.com, Inc., Alphabet Inc., Meta Platforms, Inc., NVIDIA Corporation, and Tesla — as a case study.

Conceptually, free cash flow is the cash a business generates after covering the expenditures required to sustain and grow the enterprise. In practice, the standard definition is:

FCF = Cash Flow from Operations − Capital Expenditures

Net income, by contrast, is an accrual measure. It reflects economic activity as it is recognized under accounting rules, not as cash actually moves.

There is, however, an important wrinkle. Under the standard cash flow framework, stock-based compensation (SBC) is treated as a non-cash expense and is therefore added back in calculating cash flow from operations. That treatment is mechanically correct from an accounting standpoint — but questionable from an economic one.

From the perspective of shareholders, stock-based compensation is very much a real cost. Paying employees with equity rather than cash does not eliminate the expense; it simply shifts the burden to shareholders through dilution. A firm could, in principle, issue shares in the market and use the proceeds to pay employees in cash. The economic effect is the same.

For that reason, I introduce a hybrid measure:

FCF ex-SBC = Reported FCF − Stock-Based Compensation

This leaves us with three closely related—but economically distinct—measures of performance:

• Net Income (accrual)

• Reported Free Cash Flow

• Free Cash Flow ex-SBC

The differences between them are not just accounting curiosities. As the table below using annual data for the past decade shows, when aggregated across the Magnificent Seven, they tell different stories about the underlying cash generation of the market’s most important companies.

I leave it the reader to draw his or here own conclusions from the data but there are several interesting observations.

· Most basically all the measures of income along with the market capitalization grew dramatically in the past decade.

· Net income grew faster than FCF because investment related to emerging AI grew sharply toward the end of the decade.

· FCC ex SBC grew the slowest of the three because stock-based compensation ramped up throughout the decade.

Taken together, the table and the bullet points highlight a central tension in today’s market: while the headline measures of profitability and valuation for the Magnificent Seven have surged, the portion of that growth that ultimately accrues to shareholders in true cash terms is less robust than it first appears. The divergence between net income, reported free cash flow, and free cash flow ex-SBC is not merely technical — it reflects real economic choices around investment intensity and compensation structure. To the extent that rising earnings are accompanied by heavier capital spending and increasing reliance on equity compensation, the link between reported profitability and distributable cash weakens. For valuation, this matters. If the market is implicitly capitalizing earnings while the underlying cash generation, properly adjusted for dilution, grows more slowly, then multiples may be embedding assumptions that are more fragile than they seem.

On the other hand, the rise in investment and stock-based compensation may reflect rational positioning for a potentially transformative AI cycle. If the returns on that investment ultimately exceed the cost of capital, then today’s suppression of free cash flow, particularly on an ex-SBC basis, may prove value enhancing rather than dilutive. From this perspective, the gap between earnings and cash flow is not a warning sign but a signal of aggressive reinvestment into high-return opportunities. Because valuation depends on future, not past cash flows, the weaker growth in FCF ex-SBC could, paradoxically, be consistent with a more bullish outlook. The question, therefore, is not whether current cash generation is lagging, but whether these incremental investments will translate into sufficiently higher future cash flows. That question remains unresolved and it is ultimately the one that matters.

In several earlier posts I’ve written about the growing divide between the macro economy and the stock market. Here I want to take a deeper dive, beginning with Gross Value Added.

Gross Value Added is a core concept in national income accounting. It measures what a producer — a firm, an industry, a sector — contributes to the economy, which is the value it adds to the inputs it buys. Formally:

GVA = Output − Intermediate Consumption

where output is the total value of goods and services produced (sales plus inventory change) and intermediate consumption is the cost of inputs used up in production (raw materials, components, energy, purchased services).

An example of a bakery makes it concrete. The bakery sells $1,000 of bread using $600 of flour, yeast, and electricity. Its GVA is $400. That $400 is distributed three ways: compensation of employees, gross operating surplus (profits plus depreciation), and net taxes on production.

GVA matters because it’s how we compute GDP without double-counting — one firm’s output is another firm’s input. For instance, we’d be adding steel twice if we simply summed revenue across the economy. At the aggregate level, GDP = Σ GVA + net taxes on products.

That framing sets up the real question: how can corporate profits — one slice of GVA — grow persistently faster than GVA itself? There are three mechanisms, and all three have been running hard since 2009.

1. Profit share expansion. GVA splits into labor, capital, and net taxes, so profits can grow faster than GVA only if one of the other two shrinks. Since 2009, the compensation-of-employees share of gross domestic income has fallen from roughly 54% to around 52%, with the offset going almost entirely to gross operating surplus. Wage growth has lagged productivity, margins have expanded as relative labor costs fell, and the corporate tax cuts of 2017 widened the after-tax version of the same story. Many analysts worry AI will accelerate this shift by substituting capital for entry-level labor.

2. Global GVA capture. GDP measures domestic value added, but S&P 500 earnings include global profits. Earnings growth = U.S. GVA share + foreign GVA share, and for two decades the foreign leg grew faster. This is a durable source of divergence between domestic output and listed-company earnings.

3. The Kalecki identity. Government deficit spending injects income into the private sector that supports profits, even when underlying GVA growth is modest. Post-2009 deficits — both recessionary and, unusually, peacetime expansionary — have been a persistent tailwind.

The investor’s question is how long each of these can keep running.

On labor share, AI has the technical potential to accelerate the capital-for-labor substitution, but politics may not allow it. We could instead see protection of jobs, or failing that, higher taxes on capital funding transfers to labor. The resolution is unclear; the stakes are not.

On global profits, geopolitics — including Trump’s tariff experiments — makes continued aggressive overseas expansion harder, not easier. The foreign-earnings tailwind is at minimum weakening.

On deficits, with federal interest payments now exceeding $1 trillion annually, it’s unclear how much further fiscal support can be pushed before bond markets object. The engine is still running but the fuel gauge is visible.

Of the three, AI is the critical uncertainty. If it drives accelerating substitution of labor with capital, that could push equity valuations higher still. But it’s far from obvious Americans will accept that outcome. The more likely path is a significant rewrite of how we tax and redistribute — our income-based system is already creaking, since the wealthiest Americans don’t earn much in the form of taxable income, and AI will make that mismatch starker.

For these reasons, and many others, AI is likely to bring about fundamental changes to way in which our economy operates to create and distribute wealth. It’s hard to believe those changes won’t reshape financial markets in turn. Investors should proceed with caution and not assume business as usual. We may be entering an upheaval on the order of the Industrial Revolution, compressed into a decade. Hold on tight.

But that’s not how risk feels.

What feels like risk — real risk — is something quite different: a market that falls, then falls again, and keeps falling. A market that takes years, not months, to recover. A market that forces you to sit with losses long enough to begin questioning your assumptions, your strategy, and eventually, your resolve.

As shown by the chart below, history offers several stark examples. During the 1973–1974 bear market, the S&P 500 declined in waves, ultimately falling about 50% from peak to trough. The recovery was not swift. Investors waited roughly seven years just to break even.

The bursting of the dot-com bubble followed a similar pattern. The market again fell by roughly 50%, and the recovery took nearly seven years. And just as investors regained their footing, the global financial crisis struck. The S&P 500 dropped another 57%, requiring roughly five more years to recover.

These were not brief interruptions. They were extended episodes of loss, uncertainty, and psychological strain.

This is what prompts the real questions:
When will it end?
Will I ever get my money back?
Is it even worth owning stocks?

That is what risk feels like. And it is precisely why equities have historically offered a substantial risk premium.

What is striking today is how absent this experience has been. Yes, the market fell 34% during the COVID shock. But it recovered in just six months and quickly pushed to new highs. There was barely time to process the decline, let alone internalize it.

The 2022 drawdown — about 25% — was more prolonged, but even that recovery took only about two years. More recent episodes have been even shorter-lived. The “Liberation Day” decline approached 20% but was reversed in under two months. The more recent 10% drop we just experienced was erased by a remarkable 14 straight days of positive returns.

In short, declines have been brief, recoveries swift, and the realization of sustained loss largely absent. And that has consequences.

At the Cornell Capital Group, our concern is that current market prices reflect very little in the way of a risk premium. But why should they? If every decline is met with aggressive dip-buying and rapid recovery, then risk — at least as investors experience it — appears minimal.

The implicit lesson is simple: hold on, buy the dip, and wait for the rebound. There are no prolonged bear markets—only temporary dislocations.

That belief works — until it doesn’t.

And that brings us to the real question: with retail equity holdings at all-time highs, what happens if the market delivers not another quick correction, but a prolonged, grinding bear market? Not a drop that reverses in months — but one that unfolds over years.

It is the willingness to bear that kind of risk that the market has historically priced.

But it’s the type of risk the market does appear to be pricing today. It raises the concern that we are dancing on thin ice.

The table below compares annual U.S. nominal GDP growth with S&P 500 earnings growth from 2000 through 2025.

What stands out immediately is the striking difference in average growth rates. On a simple arithmetic basis, earnings growth (17.84%) appears to vastly outpace GDP growth (4.58%). At first glance, this might suggest a persistent structural advantage for corporate profits.

But that conclusion would be misleading.

Earnings are far more volatile than GDP, and volatility has a well-known statistical effect: it inflates the arithmetic average relative to the compound (geometric) growth rate. In other words, large swings—particularly sharp rebounds following declines—boost the average without delivering comparable long-term compounding.

The table makes the point clearly. While GDP’s arithmetic and compound growth rates are nearly identical, earnings tell a very different story. The compound growth rate for earnings drops sharply — from 17.84% to just 6.47%.

Even so, earnings have still grown roughly 40% faster than GDP over this period. That kind of gap can persist for long stretches, but it cannot widen indefinitely. At some point, the underlying arithmetic reasserts itself.

For now, however, there is little sign of that adjustment. Consensus expectations for 2026 call for nominal GDP growth in the range of 4–5%, while S&P 500 earnings are projected to grow by 14.8%. If anything, the gap is expected to widen further in the near term.

It should be noted that S&P 500 earnings are likely to overstate the true growth of corporate profits in the economy. The index is not static: companies with weak earnings trajectories are removed and replaced by faster-growing firms. This introduces a selection bias that boosts reported earnings growth over time. But while this effect is real, it is not nearly large enough to account for the magnitude of the divergence.

Given this sustained gap, it is no surprise that the post-2000 period has been highly favorable for equities. The more difficult question is what comes next.

Will the rise of AI further extend the divergence between earnings and GDP? Will persistently large government deficits continue to support corporate profits? Or will competitive forces, regulation, and political backlash begin to act as constraints, pulling earnings growth back toward the pace of the broader economy?

The political dimension is especially sensitive. The gains from rising equity markets have accrued disproportionately to wealthier households, increasing the likelihood that policy responses — whether explicit or indirect — could weigh on future profit growth. AI is likely to compound the issue by eliminating entry level jobs and replacing them with capital.

These issues — not the past — are what ultimately matter for investors.

But there is a catch. To use AI effectively, you must understand both the data it is pulling and the calculations it is performing. Without that, you are outsourcing judgment—not just computation. And that can be dangerous.

A simple case study illustrates the point.

At the Cornell Capital Group, one of the valuation measures we rely on frequently is the Shiller CAPE. As developed by Robert Shiller, the CAPE is calculated as the level of the S&P 500 divided by the inflation-adjusted average of the past ten years of earnings across index constituents.

While typically applied at the index level, the same framework can be extended to individual companies. That is the exercise I asked three leading AI models—Claude, Gemini, and ChatGPT—to perform.

The results are presented in the attached spreadsheet (download required to follow along).

Where Things Start to Break

At the top of the spreadsheet is Claude’s initial summary output, including its CAPE calculations. Claude also provided a detailed backup tab showing the ten-year earnings averages used in those calculations.

There was just one problem. The summary results did not match the underlying data.

In other words, the model’s final answer was inconsistent with its own work. Had I accepted the output at face value (which looked a little wacky), I would have been relying on incorrect numbers.

To Claude’s credit, when the discrepancy was pointed out, it corrected the error and produced the revised results shown in the next tab. But the episode highlights the central issue: verification is not optional.

Agreement… Until It Isn’t

The next section of the spreadsheet reports outputs from Gemini and ChatGPT. At first glance, the results are reassuringly similar.

But they are not identical. The most striking example is NVIDIA. Gemini’s calculated CAPE is nearly twice that reported by ChatGPT. The reason is straightforward: Gemini’s estimate of ten-year average earnings is roughly half of ChatGPT’s. Alphabet shows smaller—but still meaningful—differences in the same direction.

These are not rounding errors. They are differences in the underlying data.

Even the “Easy” Inputs Differ

One might expect agreement, at least, on current prices—the numerator of the CAPE calculation. Not so. All three models report different prices. In practice, this is manageable. An analyst can (and should) impose a consistent price vector. But again, the point is broader: even seemingly objective inputs are not reliably standardized across models.

The Real Lesson

The takeaway is not that AI tools are flawed or unusable. Quite the opposite — they are already extremely useful. But they are not yet reliable enough to be used uncritically.

Today’s models can:

• Perform complex calculations quickly

• Aggregate large amounts of data

• Provide structured outputs

But they can also:

• Misalign calculations with underlying data

• Pull inconsistent or incomplete datasets

• Produce internally inconsistent results

Bottom Line

AI is a powerful assistant — but not an autonomous analyst. For now, investors should treat model outputs the same way they would treat a junior analyst’s work: useful, often insightful, but requiring careful review.

Over time, these issues will likely diminish as models improve their handling of data provenance and internal consistency.

But until then, the rule is simple: Trust the productivity gains. Verify the results.

When risk premiums compress to the point that they implicitly assume a benign, near-frictionless future, we view that as a warning signal. In 2026, that signal has not just appeared — it has been flashing.

In February, I wrote about the growing risk of a bear market in the “Magnificent 7” and the potential for meaningful losses. Since then, conditions have deteriorated materially. Using the conventional definition of a bear market — a decline of 20% or more from peak levels — the chart below shows that six of the seven stocks now qualify. The lone exception, Apple, has declined approximately 14.1% from its high. At the other extreme, both Meta and Microsoft have fallen more than 30%, placing them firmly in deep correction territory.

Individually, these drawdowns are significant. Collectively, they are extraordinary. From peak to present, the Magnificent 7 have shed more than $5.8 trillion in market capitalization—an amount that exceeds the total value of all publicly traded equities in the United Kingdom.

The magnitude of this decline has been sufficient to pull down the broader market. As the next chart illustrates, year-to-date returns for the S&P 500 closely track the performance of the Magnificent 7. Strip them out, and the remaining S&P 493 is roughly flat. In other words, the market’s weakness is not broad-based—it is concentrated.

Which brings us to the central question: has the adjustment run its course? Have prices fallen far enough to restore a reasonable relationship between valuation and expected cash flows? That is not a question that can be answered definitively in a short note—but it is the question every investor must confront.

In finance, no parameter is more important than the risk premium—the additional return investors require for holding risky assets such as stocks instead of safe government securities.

Historically, looking back more than a century, a broadly diversified portfolio of equities—such as the S&P 500—has produced a risk premium of roughly 6% over the real risk-free rate. With today’s real Treasury yield around 2.5%, and assuming 2% inflation, that traditional framework implies an expected nominal return on stocks of about 10.5%.

But there is a problem with that estimate: it assumes the risk premium has remained constant. There is considerable evidence that it has not.

Consider a simple valuation exercise. The inverse of the P/E ratio is the earnings yield. With the CAPE ratio around 40, the earnings yield is roughly 2.5%. Add a plausible long-run 2% real growth rate for the economy and you obtain an expected real equity return of about 4.5%. Add 2% inflation, and the implied nominal return is roughly 6.5%. That figure is far below the 10.5% implied by the traditional 6% equity risk premium.

The discrepancy arises because a 6% risk premium is inconsistent with a CAPE ratio near 40. Valuations at that level are more consistent with a risk premium closer to 3% or less.

One reaction is to say: I wouldn’t buy equities if the risk premium were only 3%. But that implicitly assumes that risk premiums in other markets—corporate credit, private lending, real estate—remain unchanged.

There is good reason to doubt that assumption. As Antti Ilmanen emphasized in his book Investing Amid Low Expected Returns expected returns have declined across most asset classes, not just equities. The issue is not that stocks are uniquely expensive. Rather, risky assets in general are priced to deliver lower returns than they did in the past.

That outcome should not be surprising—and it may not be entirely negative.

In a world that has become more global, more financialized, and vastly more sophisticated in its ability to analyze and price risk, the supply of capital willing to bear risk has increased dramatically. Advances in computing, data availability, and financial engineering have made markets deeper and more competitive. With more wealth chasing investable opportunities, the price of bearing risk has fallen.

If that interpretation is correct, then today’s high asset prices are not necessarily evidence of widespread mispricing. Instead, they may simply reflect a new world equilibrium in which the cost of risk capital has declined.

The good news is that businesses can raise funds more cheaply to finance productive investment. The bad news is that investors must adjust to a new reality: future real returns are likely to be lower than the ones we grew accustomed to in the past century.

Of course, I could be wrong, and risk premiums may return to their long-run historical averages. But there are only two likely ways for that to happen: either a prolonged period of stagnant stock prices or a sharp market crash. Faced with that Hobbesian choice, I would be quite content to live with lower expected returns.

At the opposite end of the spectrum, real estate accounts for less than 2% of total market capitalization. It is important to emphasize that this figure refers only to the market value of publicly traded real estate companies. The vast majority of real estate — including virtually all single-family homes — is not held by public firms and therefore lies outside these sector weights.

The focus of this post is not the sector weights themselves, but how they evolved. A sector’s weight can rise for two reasons: its earnings grow faster than those of other sectors, or the valuation multiple the market assigns to it — typically measured by the forward P/E ratio — increases. To explore this distinction, the next chart plots forward P/E ratios for all 11 sectors from January 1, 2010, through February 25, 2026.

Several findings are surprising. Real estate, not information technology, consistently commanded the highest valuation multiple over much of this period. In fact, prior to the COVID-era technology boom, the information technology sector traded near the middle of the pack. In recent years, information technology — along with consumer discretionary — has moved into second place among the 11 sectors. This re-rating helps explain the sector’s current dominance, but only in part. The bulk of information technology’s outsized weight reflects the extraordinary earnings generated by its five dominant firms.

Conclusion

The key lesson is that today’s market concentration is not primarily a story of speculative excess or runaway multiples. Valuations have played a role, but the decisive force has been earnings — massive, persistent, and highly concentrated. Market-cap weights are ultimately the cumulative result of profit generation, and the dominance of information technology reflects the simple arithmetic of exceptional profitability compounded over time.

For investors, this distinction matters. If concentration were driven mainly by inflated multiples, it would be fragile and prone to reversal (see my post on Beyond Meat). But when dominance rests on earnings power, it can persist far longer than skeptics expect — even as it raises difficult questions about diversification, competition, and the future structure of the market.

The company launched its first commercial product, Chicken-Free Strips, at Whole Foods in 2012 and expanded nationally the following year. A pivotal moment came in 2016 with the debut of the Beyond Burger. Unlike earlier meat substitutes, it was placed directly in the meat aisle beside conventional beef — a deliberate strategy aimed at flexitarians rather than committed vegetarians. Designed to “bleed” using beet juice and to sizzle on the grill, the product blurred the line between imitation and replacement.

Between 2013 and 2016, Beyond Meat attracted venture capital from prominent investors including Kleiner Perkins, Bill Gates, Twitter co-founder Biz Stone, and the Humane Society. The company’s narrative — climate solution, animal welfare, and food innovation — proved irresistible.

On May 2, 2019, Beyond Meat became the first plant-based meat company to list on a major U.S. exchange. Trading on the Nasdaq under the ticker BYND, the IPO was priced at $25 per share and raised approximately $241 million. The debut was spectacular: shares surged 163% on the first day, the best-performing U.S. IPO by a major company since the dot-com era. The company closed its first day with a valuation of $3.77 billion.

Celebrity investors such as Leonardo DiCaprio and later endorsements from figures like Kim Kardashian — hired as a brand ambassador and “Chief Taste Consultant” — amplified the cultural momentum behind the brand.

The stock continued to climb through 2019, reaching an all-time high of about $234.90 per share in July — nearly a tenfold return from the IPO price in just two months. Revenue growth appeared to validate the enthusiasm: sales rose from $298 million in 2019 to $407 million in 2020 and $465 million in 2021. By mid-2021, Beyond Meat’s market capitalization approached $9.4 billion.

Yet beneath the hype, the core economic question remained unresolved: could plant-based meat compete on price and repeat purchase behavior?

Cost proved to be a persistent obstacle. Plant-based burgers typically retailed at roughly $4 more per pound than conventional ground beef. Early marketing positioned the products as healthier alternatives, but growing consumer awareness of their highly processed nature triggered backlash. For many consumers, the value proposition — higher price, ambiguous health benefits, and unfamiliar taste — failed to justify a permanent switch.

High-profile fast-food partnerships, including McDonald’s McPlant trials and tests across multiple quick-service chains, generated publicity but rarely secured permanent menu placement. U.S. fast-food customers proved reluctant to pay a premium for plant-based substitutes.

By October 2022, as demand slackened, the company began layoffs. Between 2022 and 2025, headcount fell by roughly 25% as Beyond Meat pivoted from growth-at-all-costs to a focus on survival and cost discipline.

The trajectory of BYND’s stock price plotted below illustrates the speculative excess that can surround companies valued primarily on growth options. From its $25 IPO price in May 2019, the stock soared to nearly $235 within two months — an 840% gain driven far more by narrative than by earnings power. Unfortunately, the decline that followed was prolonged and severe. By the end of 2022, BYND had fallen below $15. The slide continued through 2023 and 2024 as revenues contracted and losses persisted. By March 2025, the company’s market capitalization had shrunk to roughly $250 million — a fraction of its peak. As of this writing, Beyond Meat shares are trading at $0.7, down 99.5% from their high.

Whether Beyond Meat can close the price gap with conventional meat, improve margins, and rebuild consumer trust in the healthfulness of its products remains uncertain. Its story stands as a cautionary tale about the gap between visionary disruption and profitable execution — and about the risks of valuing a food company as if it were a technology platform.

In this analysis, I use Robert Shiller’s CAPE (cyclically adjusted price-earnings) ratio as the valuation metric to estimate future returns for the S&P 500. All data are taken directly from Professor Shiller’s website, without adjustment or enhancement. Many analysts employ more elaborate models that attempt to account for regime shifts or structural breaks. That path is tempting — but dangerous. Searching the data for a better fit invites data mining and the discovery of relationships that are little more than statistical mirages. By relying on Shiller’s long-standing, consistently constructed dataset, we avoid that trap — though at the possible cost of reduced explanatory power.

The valuation logic behind return forecasts is straightforward. When stocks are expensive, that is when the CAPE is high, subsequent returns tend to be lower. When stocks are cheap, future returns tend to be higher.

Below, I present four scatterplots illustrating this relationship. Each chart plots the CAPE level on the horizontal axis against subsequent real (inflation-adjusted) S&P 500 returns over 1-, 3-, 5-, and 10-year horizons. The data span January 1950 to the present. The dotted red lines represent trendlines fitted to the observations.

At the one-year horizon, the observations form little more than a cloud, offering scant predictive value. The dotted red trendline does suggest that higher valuations are associated with lower subsequent returns, but the data are so noisy as to be of limited practical use.

As the return horizon lengthens, predictive power improves. At a three-year horizon, the interval many institutional investors use to evaluate performance, the relationship becomes visible, yet the signal remains largely obscured by noise.

By five years, the pattern is more evident, though the data are still dominated by variability. Only at the ten-year horizon does the relationship emerge with real clarity: higher starting valuations are consistently associated with lower returns over the subsequent decade. It is also worth noting that given the current CAPE of 40 the predicted ten year real returns is negative.

The bottom line is that valuation does possess predictive power for long-run stock market returns. Even at a ten-year horizon, however, a substantial amount of noise remains. At shorter horizons, that noise is so overwhelming that investors can largely ignore the relationship for tactical decisions.

But this does not imply that some alternative works better. Valuation, like democracy, is the worst tool for forecasting returns — except for all the others that have been tried.

“At today’s pricing, Carvana trades at more than six times CarMax’s market cap despite lower revenues. That screams “overvalued.” Yet the same conclusion would have seemed obvious when it was three times — and betting against it then would have been ruinous. Markets don’t just get irrational; they can stay that way — or become even more so.

Final Thought

Carvana’s roller-coaster ride is a reminder that the “obvious trade” — long the steady incumbent, short the flashy upstart — can be a disaster. Whether the story of Carvana ends as a lasting disruption or another speculative bubble, one thing is certain: in used cars, the wildest ride isn’t on the lot, it’s in the stock.”

How would an investor have fared had they followed the implied valuation advice — shorting Carvana and using the proceeds to buy CarMax? We will get to that in a moment but first consider the stock price behavior.

The chart below plots the two companies beginning on March 4, 2024, the last date on which they traded at essentially the same price. From that point, Carvana skyrocketed while CarMax dawdled. By the publication date of the Substack, Carvana’s share price was 6.24 times that of CarMax. The post noted that “Carvana’s sales have grown quickly yet still amount to only about half of CarMax’s. The implication is clear: most of the valuation drama comes from price action, not fundamentals.”

The post suggests — perhaps indirectly — the trade described above: short 100 shares of Carvana and use the proceeds to purchase 624 shares of CarMax. Ignoring transaction costs, the chart below plots the profits from that trade. Things looked promising for the first several weeks, and then the trade collapsed. Carvana surged further, nearly hitting $500 intraday, while CarMax continued to lag. Losses grew to more than $22,000.

It is not possible to calculate a conventional “return” on this trade because the initial investment was zero. One alternative is to compute returns relative to the initial long position. On that basis, losses on the CarMax leg approached 60%. Professional investors, however, typically use margin, which requires a cash commitment far smaller than the long position itself. Measured against the actual cash invested, losses at times exceeded 100%. It is highly likely that an investor who implemented the trade exited with a big loss rather than posting more margin.

On the bright side, on February 18, 2026, Carvana announced earnings that disappointed expectations, and the next day the stock plunged 7.94% to 332.79. Even at that level, however, the trade still showed a loss, as CarMax had fallen to 44.58 from a starting point of 59.67.

The episode highlights how difficult value investing can be, particularly when applied to high-flying stocks like Carvana. The market, it seems, stayed irrational longer than this trade remained solvent.

But recent experience tells a different story. A bear market has been quietly creeping through the group.

The table below compares each company’s 52-week high to its current price (as of 11 a.m. EST, February 12, 2026). Unless you have been watching closely, the magnitude of the declines may surprise you.

If we define a bear market conventionally — as a decline of 20% or more from a recent high — two members of the group are already there. Amazon is down 23%, and Microsoft nearly 28%. Meta and Tesla are approaching the threshold. Even the “resilient” names — Apple, NVIDIA, and Alphabet — are all solidly below their highs.

Individually, these drawdowns are notable. Collectively, they are staggering. From peak to present, the Magnificent 7 have shed approximately $3.8 trillion in market capitalization. That figure is large enough to rival the GDP of major developed economies.

Yet the broader index has not collapsed. That is the interesting part. For several years, investors were told that market strength depended on these companies. Now we are learning something more subtle: the market can hold up even as its former leaders retrench. Leadership concentration works both ways. It amplifies upside during ascent — and magnifies downside when momentum fades.

The more important question is not whether these stocks are “down.” It is whether their extraordinary valuations at which they were trading discounted expectations of years of flawless execution. When expectations are extreme, even very good outcomes can disappoint.

The largest technology companies are locked in an AI arms race, spending hundreds of billions on infrastructure they cannot afford to forgo. History suggests that when every competitor is forced to make the same massive investment, the primary beneficiaries are consumers and smaller companies — not the investors footing the bill.

We begin with a premise that we believe is increasingly difficult to dispute: artificial intelligence will be a transformative technology. It will reshape industries, redefine productivity, and alter the competitive landscape across the global economy. The question for investors is not whether AI matters — it is whether the companies spending the most on it will be rewarded for doing so.

Our answer, informed by both economic theory and the long arc of technological history, is that the odds are stacked against them.

The Arms Race

The numbers are staggering. In 2024, the five largest U.S. technology companies — Amazon, Alphabet, Microsoft, Meta, and Apple — collectively spent approximately $227 billion on capital expenditure. In 2025, that figure nearly doubled to $391 billion. Management guidance for 2026 suggests the total could exceed $635 billion. Goldman Sachs projects cumulative hyperscaler capex of $1.15 to $1.4 trillion between 2025 and 2027 alone — more than double the $477 billion spent from 2022 to 2024.

In 2025, Amazon spent $131.8 billion, predominantly on AWS infrastructure. Alphabet invested $91.4 billion — nearly double its 2024 figure — with 60% going to servers and 40% to data centers. Microsoft’s calendar-year capex reached $83.1 billion. Meta spent $72.2 billion. Looking ahead, the guidance for 2026 is even more aggressive: Amazon is targeting $200 billion, Alphabet has guided to $175–185 billion with CEO Sundar Pichai remarking that even this “still won’t be enough,” and Meta has guided to as much as $135 billion. Only Apple, with its asset-light manufacturing model, remains a relative outlier at roughly $15 billion.

These figures represent a step-change in corporate capital allocation. To put them in context, the entire U.S. telecom industry invested approximately $500 billion — total — during the fiber optic boom of 1996 to 2002. The mega cap tech companies are now on pace to spend nearly triple that amount in a three-year window.

The Prisoner’s Dilemma

Why are these companies spending at such an extraordinary pace? The answer lies in a classic game theory framework: the prisoner’s dilemma.

Consider the strategic calculus facing any individual mega cap company. If a company invests heavily in AI and its competitors do not, it achieves dominance — capturing the market and earning outsized returns. If it fails to invest while competitors do, it risks obsolescence. The rational strategy, regardless of what competitors do, is to invest aggressively.

The problem is that every company faces the same incentive structure simultaneously. When all of them invest hundreds of billions of dollars, the result is not dominance for any single player but a mutual compression of margins. The massive capital outlays become table stakes rather than a source of competitive advantage. Each company is individually rational in its decision to spend, yet the collective outcome — an industry-wide capex arms race — leaves all of them worse off than if none had been compelled to invest at this scale.

This is the Nash equilibrium of the AI era: every major technology company is locked into a spending trajectory it cannot unilaterally abandon.

History Rhymes

This dynamic is not new. Technological revolutions have repeatedly demonstrated a consistent pattern: transformative innovations create enormous social wealth while destroying investor capital. The benefits flow overwhelmingly to consumers, workers, and downstream adopters — not to the companies making the largest investments.

The Jet Airplane and the Airlines

Perhaps no industry better illustrates this phenomenon than commercial aviation. The development of the jet engine in the 1950s was a world-changing innovation. It shrank the globe, enabled mass tourism, and transformed international commerce. For passengers, the benefits have been extraordinary — a flight from New York to Los Angeles costs less in real terms today than it did in 1960.

For investors, the results have been catastrophic. Over roughly 80 years of commercial aviation, the global airline industry has generated cumulative net profits of approximately $86 billion — a rounding error on the trillions of dollars in cumulative revenue. COVID alone wiped out nearly $190 billion in three years, more than erasing the industry’s most profitable decade.

The structural problem was persistent overcapacity and ruinous competition. Airlines were forced to buy the latest aircraft — first jets, then wide-bodies, then fuel-efficient models — not because the investments promised attractive returns, but because failing to invest meant falling behind competitors who did. The capital expenditure was a cost of survival, not a source of profit.

Warren Buffett captured the dynamic perfectly: “Indeed, if a farsighted capitalist had been present at Kitty Hawk, he would have done his successors a huge favor by shooting Orville down.” After purchasing substantial positions in four major U.S. airlines, Buffett ultimately sold the entire portfolio in 2020 at an estimated $5 billion loss — a painful illustration that even the world’s most celebrated investor could not solve the structural economics of the industry.

The Telecom Boom and Bust

The late 1990s telecom buildout offers another cautionary parallel. Following the Telecommunications Act of 1996, companies invested more than $500 billion into fiber optic cable, switches, and wireless networks — most of it financed with debt. The vision was compelling: an explosion in internet traffic would require vast amounts of new bandwidth. The vision was correct. The investments were not.

By 2002, more than 23 major telecom companies had filed for bankruptcy, including WorldCom ($30 billion in debt) and Global Crossing ($12.4 billion). Approximately $2 trillion in telecom market capitalization was destroyed, and 500,000 jobs were eliminated. Less than 5% of the fiber installed during the boom was ever “lit.”

And yet — and this is the crucial point — that overbuilt infrastructure enabled the next generation of technology companies. Netflix, YouTube, and the entire cloud computing industry were built on top of dirt-cheap bandwidth that existed only because telecom investors had massively overbuilt capacity and then gone bankrupt. The social wealth created by the fiber optic boom was immense. The investor wealth destroyed was equally immense. The beneficiaries were not the companies that laid the cable but the consumers and startups that used it.

The Railroads

The pattern extends further back. In the 19th century, U.S. railroad companies attracted more capital than any other industry in history to that point. By 1897, railroad stocks and bonds totaled $10.6 billion — roughly $400 billion in today’s dollars — compared to a national debt of just $1.2 billion. The railroads were magnificent engines of social progress, enabling continental commerce, the settlement of the West, and the industrialization of America.

They were also magnificent engines of investor destruction. Nine out of ten railroad companies failed during the Long Depression following the Panic of 1873. A quarter of all U.S. rail mileage went into receivership during the Panic of 1893. Dozens of duplicate routes competed for the same freight, creating chronic overcapacity that bankrupted wave after wave of investors even as the underlying infrastructure proved transformational for the economy.

Who Actually Benefits

The historical pattern suggests that the primary beneficiaries of massive technology investment are not the investors making the investments. They are the consumers who gain access to better and cheaper services, and the downstream companies that build on top of the new infrastructure.

In the railroad era, it was the manufacturers, agricultural producers, and merchants who benefited from cheap transportation — not the railroad shareholders. In aviation, it was the traveling public and the tourism, hospitality, and global trade industries that reaped the rewards of affordable air travel. In telecom, it was Netflix, Google, and a generation of internet startups that built billion-dollar businesses on the back of overbuilt fiber networks.

The AI era appears likely to follow the same script. As the mega cap companies pour hundreds of billions of dollars into AI infrastructure — competing fiercely with one another to build the largest data centers, train the most powerful models, and deploy the most capable AI services — the cost and quality of AI capabilities available to everyone else will improve dramatically. Smaller companies across every industry will be the beneficiaries of world-class AI tools available at commodity prices, precisely because the hyperscalers are competing so aggressively to win their business.

This is the paradox of transformative technology: the more companies spend, and the more fiercely they compete, the faster the technology becomes commoditized. And commoditization is wonderful for consumers but corrosive for the margins of the companies doing the spending.

Implications for Investors

None of this means that the mega cap technology companies are bad businesses. Many of them generate enormous free cash flow from existing operations, possess powerful network effects, and occupy commanding market positions. The question is whether the incremental returns on AI capital expenditure will exceed the cost of that capital — and whether the competitive dynamics of the AI arms race will allow any single company to sustain a durable advantage.

The historical evidence is sobering. In railroads, airlines, and telecom, the technology proved far more valuable to society than to the investors who funded it. The companies that spent the most on infrastructure often earned the least. The winners were those who built on top of the infrastructure rather than those who built the infrastructure itself.

Investors considering the mega cap technology companies at current valuations should ask a simple question: are they buying the railroad, or are they buying the businesses that the railroad will enable?

The answer matters enormously. History suggests that in a prisoner’s dilemma, the prisoners rarely escape.