Unlike past bubbles driven by single factors like credit, tech, or real estate, the current AI moment uniquely sits at the intersection of all major historical bubble ingredients simultaneously: loose credit, a great technology story, a real estate component (data centers), and a policy angle.
AI tokens are a hyper-deflationary commodity, with prices falling 70-80% annually. To simply maintain flat revenue, frontier AI companies must achieve a staggering 400% growth in unit volume year-over-year, creating an incredibly difficult economic environment before even considering profit or pleasing Wall Street.
Early AI adoption in coding was an 'expansive' use case (prompt to massive code). Most future white-collar uses are 'compressive' (large documents to summaries). Naively extrapolating from the unique characteristics of coders has led to flawed projections for data center capacity and overall demand.
Lenders financing data centers are increasingly divorced from the underlying economics. They don't care about GPU utilization; they look through the facility to the prime credit of the hyperscaler tenant with a long-term lease. This financialization fuels overbuilding, regardless of actual AI demand.
Despite narratives of scarcity, GPU utilization in some data centers is as low as 35-40%. This is due to colossal hoarding and double-ordering by companies terrified of being caught short in a future parabolic moment. This behavior creates a phantom demand and points to a future supply glut.
A massive wave of IPOs from companies like SpaceX and Anthropic will require large institutional funds to free up trillions in capital. Since funds don't hold much cash, they will be forced to sell their most liquid, overlapping, and best-performing stocks, creating predictable downward pressure on today's market leaders.
Contrary to the hype, year-over-year performance gains for LLMs have dramatically slowed and nearly flatlined. Furthermore, the performance variance between competing models has collapsed. It's now nearly impossible to distinguish between them in a blind test, indicating they are becoming commoditized.
AI is fundamentally transforming semiconductor design, reducing verification stages from months to days. This will enable a flood of new, specialized chip designs from startups, collapsing the tribal knowledge moats of incumbents and bursting the narrative of a perpetual semiconductor super-cycle.
For a company at a very high valuation, failure is 'overdetermined.' There are so many different, low-probability ways things can go wrong that the combined probability of one of them occurring becomes very high. This makes a negative outcome statistically likely, even if each individual risk seems small.
Americans are anomalously negative towards AI compared to other developed nations. This is because any threat to employment is also a direct threat to healthcare and personal solvency, a link that doesn't exist in most other Western countries. This makes the job-loss narrative of AI particularly frightening in the US.