In 1999, a small Las Vegas car dealership called Uniprime Capital Acceptance Inc more than doubled its stock price by announcing—without any evidence—that one of its subsidiaries had found a cure for AIDS. Investors didn’t blink. They bought the stock. This episode captures a core feature of bubbles. It is not just about high valuations, but a collective abandonment of basic underwriting discipline.
In financial theory, bubbles form when three forces reinforce each other: extrapolative expectations (investors naively project recent returns indefinitely), abundant liquidity (new money continuously enters to validate those expectations) and narrative-driven coordination (investors act not on fundamentals but on what others believe).
This is akin to the Keynesian ‘beauty contest’, where investors buy not what they think is valuable, but what they think others will soon find beautiful. Behavioural finance calls this ‘positive feedback trading’. Macro-finance models call it a ‘liquidity-driven asset boom’. In practice, it is the same phenomenon—a system in which capital inflows, sentiment and balance-sheet leverage co-evolve faster than fundamental cash flows can justify.
Seen through that lens, the artificial intelligence cycle of 2023-25 resembles a bubble’s early stage. The irony is that the fundamentals look spectacular. Nvidia’s latest quarter reported $57 billion in revenue (up 62 percent from last year), $51.2 billion of it from data-centre AI chips alone, and a forward outlook of $65 billion with roughly half a trillion dollars in future chip orders booked. AMD has almost quadrupled in value in three years, signed multibillion-dollar chip deals with OpenAI and Oracle, and now publicly projects that AI data-centre spending will reach $1 trillion per year by 2030. Yet, the market’s reaction has been ambivalent in recent months.
The most immediate stress point is the pace and scale of AI capital expenditure. In 2025, Microsoft, Google, Amazon and Meta together spent over $350 billion on capex, almost entirely on AI infrastructure. Market analysts now project more than $400 billion in 2026 from these four alone. Morgan Stanley estimates that the global AI infrastructure build-out for 2025-28 will require $2.9 trillion, financed in a strikingly mixed structure—$1.4 trillion from Big Tech cash flow, $800 billion from private credit, $200 billion via corporate bonds, $150 billion in asset-backed securities, and $350 billion through private equity and auxiliary channels. This composition may be a classic signal of a maturing bubble. The marginal dollar is no longer funded from generated cash, but borrowed liquidity.
For instance, Meta’s Hyperion complex, a 500-megawatt data centre, was financed with a $30-billion package. The key financier for the Oracle-OpenAI Stargate facility, which required a $10-billion loan, backed out last month. A Vantage Jacquard deal for two gigawatt-scale centres in the US was underwritten by a $38-billion lending syndicate involving more than 30 banks.
Bubbles expand until liquidity becomes scarce. When companies that could once self-fund begin tapping global debt markets to maintain AI momentum, the liquidity engine sustaining expectations becomes vulnerable to exhaustion.
Complicating this is the rise of circular financing. Oracle leases a 1.2-GW facility whose viability depends on OpenAI generating roughly $300 billion in longterm revenues. Nvidia invests up to $100 billion in OpenAI, which in turn buys Nvidia chips and deploys Nvidia-powered computers in data centres financed by investors whose returns depend on these same hyperscalers meeting ambitious AI adoption curves. AMD offers OpenAI warrants to buy up to 10 percent of AMD stock at deep discounts in exchange for volume commitments.
These loops are legal, sophisticated and sometimes efficient. But they intertwine the ecosystem’s balance sheets in ways that obscure counterparty exposure. It means a small revenue disappointment could propagate through equity markets, private-credit vehicles, cloud-infrastructure leasing, municipal tax bases and even household wealth accounts.
The macro-concentration amplifies this. AI-linked stocks now comprise 48.9 percent of S&P500’s market cap, surpassing the 43.8 percent peak of the dot-com era. Equity constitutes 32 percent of American household wealth, and nearly half of the past year’s wealth gains came from AI-adjacent stocks. A correction on the scale of the early-2000s’ episode would erase roughly 8 percent of US household net worth, which historical consumption elasticities suggest would reduce GDP by 1.6 percent, enough to tip a fragile economy into recession.
Meanwhile, the behavioural signatures are unmistakably bubble-like. Corporate boards no longer ask, what is the return on your AI investments? They ask, what is your AI plan? OpenAI speaks of needing 26 GW of compute power (about $1.5 trillion in capex) while still running core products at a loss even on $200-a-month subscriptions. Meta forecasts ‘notably larger’ capex without concrete monetisation. OpenAI CEO Sam Altman admits “many parts of AI are bubble-y right now”. And yet, spending accelerates because the strategic cost of under-investing now exceeds financial risk of over-investing.
While AI is grounded in genuine technological progress, real demand for computing power and transformative longterm potential. But the financial architecture being built atop it is increasingly fragile. And as history repeatedly reminds us, fragility is the prick that bursts bubbles.
Aditya Sinha | Public policy professional
(Views are personal)
(On X @adityasinha004)