There is a number that clarifies the current moment better than any analysis: NVIDIA, a company that does not manufacture a single physical product, is now worth more than the entire economy of India.
India has 1.4 billion people. It has steel mills, pharmaceutical factories, software firms, agricultural systems, ports, railways, banks, and a middle class that is larger than the entire population of the United States. NVIDIA has a design office and some very good engineers.
This is not a criticism of NVIDIA. It makes genuinely extraordinary products and has executed one of the most remarkable runs in corporate history. It is, however, a description of a market that has drifted so far from productive reality that the valuations being assigned to AI-adjacent companies no longer reflect any coherent theory of how money is actually made.
What a Bubble Looks Like From Inside It
Bubbles are not characterized by fraud, at least not primarily. They are characterized by a collective agreement to believe in a future that has been borrowed from so aggressively that the present can no longer support the debt.
The AI bubble has a specific character. Unlike the dot-com bubble, which involved thousands of companies with essentially no revenue, the current concentration is in a smaller number of companies with very large revenues — but valuations that price in decades of growth at rates that no large company has ever sustained. The distance between what these companies earn and what the market is paying for them is being filled by a story. The story is AI.
The story may even be partially true. That is what makes it dangerous. The most destructive financial bubbles are always built on something real. The railroad boom was built on real railroads. The dot-com boom was built on a real internet. The housing bubble was built on real houses. The underlying technology or asset was genuine. What was not genuine was the price.
SoftBank and the Credit Card Logic
Nothing illustrates the structural fragility of the current moment more precisely than the behavior now attributed to SoftBank: attempting to borrow money from banks using its own stock as collateral.
The logic is worth unpacking, because it is the same logic that underpins most of the AI investment boom, just rendered in unusually naked form.
SoftBank holds large positions in AI and technology companies. Those positions have risen dramatically in value on paper. SoftBank is using that paper value as collateral to borrow cash, which it is presumably deploying into further AI investments, which rise in value, which supports further borrowing.
This is not leverage in the productive sense — borrowing to build something that generates returns that exceed the cost of the debt. This is leverage in the circular sense: using the appearance of value to generate the cash to sustain the appearance of value. It is, as Professor Jiang puts it, maxing one credit card to pay the interest on another. The mechanism works until it does not, and when it stops working, the unwind is not gradual.
The amount of AI investment that is currently structured this way — not identically, but with the same circular dependency between asset prices and the capital flows that sustain them — is not publicly disclosed and not easily quantifiable. But the SoftBank dynamic is not unique. It is illustrative.
NVIDIA and the Design Company Problem
NVIDIA’s valuation requires confronting an awkward structural fact: it is a design company.
This is not a trivial distinction. NVIDIA does not own fabs. It does not manufacture the chips it designs. That manufacturing is done by TSMC, primarily in Taiwan. NVIDIA’s competitive position rests on its chip architectures, its CUDA software ecosystem, and the relationships and market position it has built in AI computing.
These are real and significant advantages. NVIDIA’s H100 and subsequent chips are genuinely difficult to replicate, and the CUDA ecosystem creates switching costs that give the company meaningful pricing power.
But a design company’s assets are, at their core, intellectual. They live in the minds of engineers, in codebases, and in the accumulated design knowledge embedded in products. Unlike a factory, they cannot be insured against fire. Unlike a mine, they do not sit in a fixed geographic location that competitors must physically displace. Unlike a rail network, they do not create permanent geographic barriers to competition.
The competitors trying to displace NVIDIA — AMD, Intel, and an array of custom chip efforts at Google, Amazon, Microsoft, and Meta — are not facing the barriers that typically protect a company valued at several times the GDP of a major nation. They are facing a software ecosystem and a design lead that is real but also the kind of advantage that, in the history of technology, has been overcome before.
Elon Musk, Trillionaire
Elon Musk becoming the world’s first trillionaire is a data point, not an accomplishment. The distinction matters.
His wealth is not a trillion dollars in any form he can spend. It is the aggregated market capitalization of equity stakes in companies — Tesla, SpaceX, and others — multiplied through the specific dynamics of markets that are rewarding narrative and proximity to AI with a generosity that has no clear analytical foundation.
SpaceX is valued at roughly two trillion dollars. This is more than Meta, a company that generates tens of billions in annual profit and owns the social media platforms used by three billion people. SpaceX’s revenue is real and growing. Its technological achievements — reusable rockets, Starlink — are genuine and significant. But two trillion dollars implies a scale of future cash generation that requires SpaceX to dominate industries that do not fully exist yet, at margins that no capital-intensive hardware business in history has sustained at scale.
The question is not whether space is interesting. The question is what price is appropriate to pay today for a future that is, at best, speculative and, at worst, dependent on a single individual whose attention is demonstrably divided and whose relationship with markets is, to put it diplomatically, unusual.
The Moon and the Problem of Promises
Both the United States and China have announced intentions to land humans on the Moon by 2030. This is worth examining not to be cynical about space exploration, which carries genuine scientific and strategic value, but to illustrate the larger problem with the valuations being attached to the companies involved.
The Moon program is hard. It has been hard before — the original Apollo program required roughly 400,000 people and consumed four percent of the federal budget at its peak. The current programs are operating on different budgets, with different contractors, under political pressures and timelines that do not bend to engineering realities.
Whether the 2030 deadline is met is, in a narrow sense, a technical question. In a broader sense, it is a question about whether the gap between announced ambition and deliverable reality is as wide in rocketry as it has historically been in most capital-intensive technology programs. The track record on self-imposed deadline in this domain is not encouraging.
The market is pricing SpaceX as if the Moon and Mars are foregone conclusions. Engineering history suggests they are not.
What Happens When the Story Stops Working
Every financial bubble ends not when the underlying technology fails, but when the gap between the story and the cash flows becomes too wide to sustain the price.
The AI story is powerful, and in important respects it is true. AI is a genuinely transformative technology. The companies building it are doing real and difficult things. The infrastructure being constructed to support it is unprecedented in scale. None of this is fabrication.
What is under-examined is the profit question. The capital being deployed into AI infrastructure — the data centers, the chips, the energy buildout — is vast. The revenue being generated by AI products, for all its growth, has not yet demonstrated a capacity to service the investment at the rate the investment is being made. The gap is being bridged by a belief that the revenue will eventually arrive at the required scale.
It may. But “it may” is not the same as “it will,” and the valuations being assigned to AI companies are not priced for “it may.” They are priced for certainty that the capital markets are not entitled to assume.
The bubble will not announce itself. It never does.