The Price of Understanding

Intelligence is becoming free. That is the miracle. It is also the problem. The six companies currently valued at $2.2 trillion are building something that, by its nature, cannot be charged for at the scale their prices require.

There is a particular kind of mistake that intelligent people make. It is the mistake of confusing enthusiasm with analysis, of allowing the magnitude of what is possible to blind us to the arithmetic of what is probable. I have made this mistake myself. Enthusiasm is easy. Arithmetic is hard.

In my previous essay, I argued that intelligence had become effectively free. I remain convinced this is true. The technology is real, the capability is extraordinary, the implications profound.

But the technology and the business are not the same thing. They are related. They are interdependent. But they are not identical. The fact that something is revolutionary does not mean that every company built upon it is valuable. The fact that electricity transformed the world does not mean that every company with "Electric" in its name was a good investment. The fact that the internet changed everything does not mean that Pets.com was worth billions. Price and value are different things, and the distance between them is bridged only by patience, not by faith.

The Valuations

Let us begin with what we know. As of early 2026, the following valuations have been assigned to the leading foundational AI laboratories—the companies building the large language models and AI systems that I described in my previous essay. I have used conservative attributions for the AI components of Google and Meta, because I want this to be difficult to contest. This is the floor, not the ceiling.

OpenAI, the creator of ChatGPT, raised $110 billion in February 2026 at a $730 billion valuation. Anthropic, the maker of Claude, raised $30 billion at a $380 billion post-money valuation. Google Gemini, the AI division of Alphabet, carries roughly $800 billion in attributed value. Meta's Llama initiative, despite its open-source nature, is valued at approximately $250-300 billion within Meta's enterprise. DeepSeek, the Chinese laboratory that shocked the industry with its efficient models, is valued at approximately $3.4 billion. MiniMax, which went public in Hong Kong in January 2026, has a market capitalization of approximately $20 billion.

Adding these together, we arrive at a combined enterprise value of approximately $2.2 trillion.

Two point two trillion dollars. A number so large it ceases to register. And yet these are not abstract numbers—they are the specific prices that smart, well-advised people have agreed to pay for shares in companies that, in many cases, have never earned a dollar of profit.

The Mathematics

Now for the arithmetic. To understand whether these valuations are justified, we need a framework. The standard approach—used by every serious investor, every private equity firm, every investment bank—is the discounted cash flow model. You estimate future cash flows, you discount them back to the present, and you see if the present value exceeds what you are paying.

The logic is straightforward: a dollar earned in the future is worth less than a dollar earned today. The discount rate captures this time value of money and risk. A 25% discount rate is standard for high-risk, high-return investments—it reflects the reality that most startups fail.

For companies like these—unprofitable, high-growth, venture-backed—I will use assumptions favorable to them: a high discount rate (25%) for the first five years, then a lower rate (12%) after year five, once we assume these companies have matured into stable, profitable businesses. This two-stage approach is standard for valuing high-growth companies.

With these assumptions, what level of EBIT must these companies achieve in year five to justify a $2.2 trillion valuation? The formula is straightforward:

Required EBIT = Enterprise Value × (1 + r)ⁿ × (r - g)

Where r = 0.25 for the first five years (the high-risk venture phase), g = 0.03 (terminal growth rate), and r_perpetual = 0.12 (the more modest discount rate that applies after year five, once these are mature, profitable businesses). Let us calculate:

$2.2T × 1.25⁵ × (0.12 - 0.03) / (1 + 0.03) = $2.2T × 3.05 × 0.09 / 1.03 ≈ $586 billion

Five hundred eighty-six billion dollars. In year five alone. The mathematics do not rage or thunder. They simply state: not this much, not this fast, not from this.

A note on terminal value sensitivity: the $586 billion figure is heavily influenced by the terminal value calculation, which assumes the companies grow at 3% forever after year five with a 12% discount rate. If we instead assume zero terminal growth—a more conservative approach that treats these as temporary monopolies that eventually become mature businesses—the required year-five EBIT rises to over $900 billion. If we assume the perpetual discount rate remains at 25% (no maturation), the required EBIT exceeds $1.2 trillion. The numbers do not improve with more conservative assumptions. They worsen.

What Does $586 Billion Mean?

For perspective: the entire U.S. defense budget—across all branches, all agencies, all operations—amounts to roughly $900 billion annually. The AI labs would need to generate nearly two-thirds of what the United States spends on its military. In year five.

Apple generates approximately $130 billion in annual EBIT. Microsoft approximately $110 billion. Saudi Aramco, the world's most profitable company, earns roughly $120 billion. Amazon generates approximately $70 billion. The combined EBIT of these four companies is approximately $430 billion.

The AI labs would need to generate roughly 1.4 times the combined operating profit of these four companies. In year five. To justify valuations that already exist today.

Even in more optimistic scenarios—a 15% discount rate, profitability in three years, 35% margins—the required EBIT is still over $150 billion, comparable to Google's entire operating profit. Only in a scenario where everything goes perfectly—profitability faster than any technology company in history, margins that would make Warren Buffett weep—do we arrive at a number ($35 billion) that is merely improbable rather than impossible.

From a revenue perspective, to achieve the required EBIT at a 25% margin, these companies would need approximately $2.3 trillion in annual revenue—roughly the entire GDP of France or the United Kingdom.

Current revenue gives a sense of the gap. OpenAI generates approximately $10 billion in annualized revenue. Anthropic approximately $9 billion. Combined, perhaps $20 billion. To reach $2.3 trillion in five years would require annual revenue growth of approximately 210% per year—every year, for five years.

Let us work through the math. Year one: $20 billion becomes $60 billion—doable. Year two: $60 billion becomes $180 billion, the entire revenue of a major corporation, added in a single year. Year three: $180 billion becomes $540 billion, exceeding the annual revenue of most countries. Year four: $540 billion becomes $1.6 trillion, more than the GDP of Canada or Australia. Year five: $1.6 trillion becomes $4.8 trillion, larger than the GDP of Germany.

The human mind does not understand exponential growth. We think linearly: if we can grow 100% this year, we can probably grow 100% next year. But each year's base is larger than the entire company was the year before. At $20 billion in revenue, adding $20 billion feels like a big year. At $180 billion, adding $180 billion feels like building an entirely new Fortune 500 company. At $1.6 trillion, you are adding more than the entire output of most nations. The math is not difficult. It is simply impossible to visualize.

Objections Considered

I can hear the objections forming. Let me address the most common ones, because they deserve to be taken seriously.

"The revenue growth will accelerate as adoption scales." Perhaps. But OpenAI grew from roughly $1 billion to $10 billion in revenue in one year—one of the fastest ramps in technology history. Even if they could maintain 100% growth per year, it would take seven years to reach $1 trillion in revenue—and over ten years to reach $2.3 trillion.

"The models will become cheaper to run, increasing margins." This is the argument I made in my previous essay—that intelligence is becoming free. And it is true. But here is the tension: if intelligence becomes free, then the revenue potential also collapses. You cannot charge premium prices for a commodity. The very efficiency that makes these models powerful is the same efficiency that destroys their pricing power.

"They will expand into new markets—agents, robotics, infrastructure." They may. But each of these markets has its own competitors, its own economics, its own limits. Robotics alone is a decades-old industry with razor-thin margins. Infrastructure is a commodity business. The idea that AI companies will simply expand into adjacent markets and capture infinite value is an article of faith, not analysis.

"What about Microsoft or Amazon—won't they win?" This is a reasonable point. Microsoft has invested heavily in OpenAI and integrated it across its product suite. Amazon has done the same with Anthropic. These companies have existing revenue streams, existing customer relationships, existing infrastructure. It is entirely possible that one or both of them will capture significant value from the AI revolution.

But here is the problem: Microsoft and Amazon are already valued at $2.8 trillion and $1.8 trillion respectively. Any value they capture from AI is already reflected in these valuations—they are not cheap. And the AI lab valuations I am analyzing are separate from these platform companies. We are assigning $2.2 trillion to the AI labs themselves, on top of the valuations already held by Microsoft, Amazon, Google, and Meta. The mathematics require not just that one of the existing platforms wins, but that the AI labs themselves become profitable at scales that have never been achieved.

"The infrastructure plays will win—NVIDIA, data centers, the enablers." This is perhaps the most compelling counterargument. If AI becomes a commodity—if intelligence is free—then the money may flow to those who provide the raw materials rather than those who build the finished product. NVIDIA has captured enormous value. The data center operators are doing well. The electricity providers are seeing increased demand.

But even here, there are limits. NVIDIA's valuation already reflects enormous expectations. Data centers are capital-intensive businesses with constrained returns. And the historical pattern with infrastructure plays is that they capture value in the early stages of a technological shift, but the economics tend to compress over time as competition intensifies and the technology matures. The infrastructure may be necessary, but it is rarely sufficient for capturing the lion's share of value.

"These are strategic investments—the buyers expect to lose money." This is the most interesting objection. It is true that many of the buyers—Microsoft, Amazon, NVIDIA, the sovereign wealth funds—are strategic investors who may have reasons beyond pure financial return. They may be buying compute capacity, partnership access, or strategic positioning.

But this argument, taken seriously, suggests the valuations are not based on financial returns at all. They are based on something else—status, strategic positioning, fear of missing out. That is not investing. It is a public goods dilemma, where each participant hopes to exit before the others do.

The Paradox of Intelligence

There is a deeper problem with these valuations—one that goes to the heart of what it means to build a business on intelligence. It is the problem of mutual exclusivity.

Consider what we are actually buying when we buy a share in an AI company. We are buying the right to participate in the production of understanding. And here is the paradox: if these companies succeed—if they truly commodify understanding—then understanding loses precisely the quality that made it valuable. The moment you can buy intelligence, you have already destroyed its scarcity, and with its scarcity, its exchange value. This is, by definition, impossible.

If OpenAI were to generate the kind of profits required to justify its valuation—$500 billion or $1 trillion in annual EBIT—it would have to capture the vast majority of the market for AI intelligence. Not a large share. Not a dominant share. The share. It would have to be the default, the standard, the only game in town.

And here is the problem: if OpenAI achieves this, then Anthropic cannot. If Anthropic achieves this, then Google Gemini cannot. The mathematics require that one company—one company out of six—capture essentially all the value in a market that is already generating trillions of dollars in enterprise value. The valuations assume that every horse in the race will win. That is not how markets work.

We have seen this pattern before. In the early days of the automobile, there were hundreds of manufacturers. In the early days of the internet, everyone thought they needed their own search engine. What happened? The market consolidated. One or two players captured the lion's share of value. The rest became footnotes.

The AI market will likely follow the same pattern. And the problem is this: we are assigning $2.2 trillion in enterprise value to six companies, but the mathematics only work if one of them captures almost all the value. At most, one or two will.

The Collapse of Economics

There is another force at work here: the collapse of economics that is already underway, driven by two interconnected forces—the collapse of marginal costs and the intensity of competition.

Consider the cost structure of running a large language model. The fixed costs are enormous: data centers, chips, engineering talent, training runs that cost hundreds of millions of dollars. But the marginal costs—the cost of serving one additional token to one additional user—are trivially small. Once the model is built, running it costs little more than electricity. And electricity, at scale, is cheap.

This is a classic structure that economists have understood for centuries: when fixed costs are high and marginal costs are low, prices inevitably trend toward marginal cost. Not toward what the market will bear. Not toward what investors expect. Toward marginal cost. Because any price above marginal cost is an invitation for a competitor to undercut you.

And the competitors are coming. MiniMax offers models that are approximately 95% cheaper than Anthropic's leading models—for performance that is only slightly below the state of the art. Token consumption is growing at approximately 700% per year, a staggering rate. But prices are falling at approximately 70% per year. The volume of intelligence being consumed is exploding, but the revenue generated per unit is collapsing. The two curves are moving in opposite directions.

This is the collapse of economics in action. The industry is producing more value than ever before, but capturing less of it. The gap between the value created and the value captured is widening with each passing quarter.

And there is another problem. If AI truly does begin replacing white-collar workers en masse, then who will be left to pay for the tokens? The very customers who would fund the AI companies' valuations are the ones being automated away. We are building machines that make intelligence free, and then wondering why we cannot charge for intelligence. The answer is already embedded in the technology itself.

A Personal Confession

I want to be clear about where I stand, because what follows may seem to contradict the argument I have been making. I use these tools. I depend on them. I have integrated them into my thinking, and I cannot imagine working without them now—any more than I could imagine going back to handwritten letters or horse-drawn carriages. This is the position of someone who sees the cliff and walks toward it anyway.

I believe, genuinely and deeply, that AI will transform everything. Not in the science-fiction sense—the robots do not need to rise up. But in the practical sense: AI will embed itself into every facet of human life. Our phones, our cars, our homes, our workplaces, our governments, our schools. We will interact with it constantly, often without knowing it. It will be the invisible infrastructure of civilization.

And yet.

And yet I look at the valuations and I see a problem. Because the belief that AI will transform the world is not the same as the belief that every company building AI will capture that transformation's value. In fact, the opposite may be true: the more powerful and pervasive AI becomes, the cheaper it will be, and the less any individual company will be able to charge for it.

There is something almost philosophical about this paradox. Like all commodities, intelligence is subject to the same brutal economics: when everyone can have it, no one can charge for it. The value does not disappear—it simply diffuses outward, to the people who use the intelligence, to the companies that build the infrastructure, to the society that benefits from having a smarter world.

This is what makes the current moment so strange. We are building something that will change everything—and yet the business models that would capture that change's value are dissolving even as we build them. The technology transforms society. The economics transform the pockets of those who build it. These are not the same things.

I could be wrong. It is entirely possible that these AI labs will find business models I cannot imagine, will capture value in ways I have not considered. I have been wrong before. I will be wrong again.

But the numbers are what they are. To justify these valuations, the math demands profits that exceed the combined profits of the four most profitable companies on Earth. This is not a matter of opinion. It is a matter of arithmetic. And arithmetic, unlike enthusiasm, does not forgive.

I use these companies every day. I am building on top of their platforms. I believe the technology they are building will change everything. But I also believe the valuations placed on them are disconnected from any plausible model of future cash flows. These are not contradictory positions. They are simply honest ones.

What, Then, Is the Bubble?

I am not arguing that AI is a fad—the technology works, the capabilities are real, the impact on society will be profound. I am not even arguing that these specific companies will fail. Some may succeed spectacularly. Some may be acquired.

What I am arguing is simpler: the aggregate valuations—$2.2 trillion at minimum—cannot be justified by any plausible model of future cash flows. The numbers do not work. The mathematics are not ambiguous. They are wrong by an order of magnitude. This is what a bubble looks like: not when everyone is enthusiastic, but when the enthusiasm has detached from the underlying economics.

Historical Parallels

We have seen this pattern before. Each time, the same three movements: a genuine transformation that changed what was possible; a mania that conflated possibility with profit; a reckoning that separated the wheat from the chaff. And in every case, the technology survived. It was the speculation that died.

In the late 1990s, the internet was genuinely transformative—and the valuations placed on internet companies were genuinely insane. Cisco traded at a P/E ratio of 100+. Yahoo was worth more than IBM. Pets.com had a higher market cap than the entire retail sector in some measures. The internet did not go away after the dot-com crash. It became more important, not less. Amazon, Google, eBay—some of the most valuable companies in human history today. But the bubble burst anyway, because the valuations had decoupled from the economics. The technology was real. The prices were not.

Consider the railroads. In the 1840s and 1850s, railroad stocks were the Bitcoin of their day—everyone wanted in, and prices bore no relationship to earnings. Thousands of miles of track were laid, many to nowhere in particular, funded by borrowed money and optimism. When the crash came in 1857, it was brutal. Yet the railroads themselves persisted. They reorganized, consolidated, and became the backbone of American commerce. The technology transformed the nation. The speculation bankrupted some investors and made others rich. The two outcomes had almost nothing to do with each other.

What This Means

If the valuations cannot be sustained, what follows? They will come down, not because the technology fails, but because the market will eventually insist on some relationship between price and earnings. This may take the form of a sharp correction, or it may take the form of years of stagnation while revenue catches up. Either way, the gap between perception and reality will close. It always does.

If the AI labs cannot capture the value of their technology—if their margins are compressed by competition and the commoditization of intelligence—then where does the value go? Possibly to the enablers: chip makers, data centers, infrastructure providers. Possibly to the customers who benefit from falling prices. The history of technology suggests that the most valuable companies are rarely the ones that built the foundational technology—they are the ones who found ways to apply it at scale. Of the six AI labs currently valued at $2.2 trillion, perhaps one or two will emerge as genuine winners; the rest will likely be acquired, merge, or fade. This is the pattern of every technological revolution.

The crucial point—the one most likely to be missed—is that the AI labs may not earn their valuations while the technology itself continues to improve, continues to spread, continues to transform everything it touches. The bubble is in the companies, not in the technology. The technology is real. The transformation is real. It is only the financial expectations that are detached from reality.

Conclusion

This is not a bearish essay. The technology is extraordinary—the most important development in human history, perhaps. The capabilities are real. The transformation is genuine. But the investment thesis—the idea that you can buy into these companies at current prices and earn a reasonable return—is divorced from arithmetic.

We are living through the strangest economic paradox in modern history: we are building machines that make intelligence free, and then trying to charge for it.

Perhaps we will look back at these valuations and laugh, the way we laugh at Pets.com now. Or perhaps we will look back and wonder why we ever doubted. Either way, the arithmetic does not lie. The bubble is real. The technology is real. These are not the same thing.