Calling AI a bubble leaves two questions unanswered. Which asset is overpriced, and what event would force its price back toward the cash flows it can produce?
At the time of writing, private model companies carry large valuations, Nvidia can gain or lose hundreds of billions of dollars around an earnings report, and infrastructure spending recalls the late-1990s fiber buildout. Those facts establish high expectations. They do not establish that model use, accelerator demand, private-company valuations, and every data-center project share one fate.
A Company Can Fail While the Market Grows
OpenAI’s failure is one proposed trigger. Nokia and BlackBerry show why that claim is incomplete.
At its peak, Nokia reported about $50 billion in revenue and served roughly 40% of the mobile-phone market. It had built smartphones before the iPhone, but Symbian, its organization, and its eventual Windows Phone commitment left it unable to compete with iOS and Android. BlackBerry reached about $20 billion in revenue and 100 million BBM users, then lost its enterprise advantage as secure communications became available on competing devices.
Both companies declined while the smartphone market grew. Their histories show that a leading company can lose its position without invalidating the technology category. OpenAI could likewise lose share because of price competition, product decisions, enterprise requirements, or a stronger rival. That would correct OpenAI’s valuation; it would not by itself show that demand for models had collapsed.
The same distinction applied during the dot-com crash. Many internet companies failed, public valuations fell, and the internet continued to spread. A technology can be useful and its financing can still form a bubble.
Plausible Correction Mechanisms
A serious bubble claim should identify a gap between expected and realizable cash flow, then explain what exposes it. Several mechanisms qualify.
Revenue and margins disappoint. Model use may grow while competition lowers prices faster than inference costs fall. If providers cannot earn the margins assumed in their valuations, financing terms and equity prices will adjust even with rising usage.
Built capacity exceeds economic demand. Data centers can be technically useful yet fail to earn their cost of capital. Long power commitments, debt, and specialized equipment would magnify losses if utilization or prices came in below plan.
Technical returns diminish. If additional training and inference spending produces smaller improvements than customers will pay for, the expected return on frontier investment falls. Model progress need not stop; it only has to underperform the assumptions embedded in current spending.
Capital becomes more expensive. A rise in interest rates, tighter credit, or losses at a major borrower could force projects to refinance or stop. This is the ordinary route by which distant expected cash flows lose present value.
A different technical approach devalues current assets. A new architecture could reduce the value of existing clusters, model weights, or proprietary methods. The relevant evidence would be a material cost or capability advantage, not the announcement of another model.
Law changes deployment economics. Privacy, copyright, safety, energy, or trade rules could raise costs or restrict particular uses. A global prohibition is unnecessary; regulation in a large market could change expected returns.
These mechanisms do not require the models to be useless. They require realized revenue, margins, or asset utilization to fall short of the expectations that financed them.
What to Ask
The useful disagreement is therefore quantitative. What revenue growth, gross margin, utilization, and cost of capital does a valuation assume? Which projects remain economic if token prices fall, power costs rise, or utilization arrives two years late? How much debt or contractual exposure connects one failed project to others?
“OpenAI will fail” does not answer those questions. Neither does “the models work.” A company can fail in a growing market, and a useful technology can attract more capital than its near-term cash flows justify.
Ask anyone calling AI a bubble to name the overpriced asset, the unsupported assumption, and the event that closes the gap. That turns a mood into a claim that evidence can test.