The Dangerous Architecture of the AI Capex Bubble

In 2006, credit markets priced U.S. housing risk as diversified. The subprime market was born. The mathematics looked safe. Reality delivered collapse — and the 2008 Global Financial Crisis.

In 2025, a strikingly similar illusion has taken hold — this time in artificial intelligence.

For now, AI has been largely a capital-expenditure phenomenon, dominated by a very small number of companies such as Nvidia and Oracle — the modern “shovel makers” — building chips and data centers to absorb nearly $1.5 trillion of cumulative investment by a handful of hyperscalers such as Microsoft, Meta, and Amazon. This unprecedented spending wave has driven stock prices to unimaginable heights.

At the same time, the AI sector has constructed a circular financing architecture in which the same capital flows repeatedly among a small group of dominant players. Each books revenue from the others. Each justifies valuation expansion through the others’ spending. And each is treated by financial markets as if its risks were independent.

They are not.

At the core of this structure lies a single, unprecedented contract: a $300 billion, five-year agreement between OpenAI and Oracle.

This contract has quietly become the load-bearing pillar of the global AI investment thesis. If it holds, the AI capex boom survives. If it fractures, the consequences extend far beyond the two signatories.

OpenAI raises capital at ever-higher valuations — capital that is then spent on Nvidia chips and Oracle infrastructure. Oracle, in turn, uses OpenAI’s commitment to borrow massively in order to fund data-center expansion, much of which flows straight back to Nvidia.

Nvidia’s equity appreciation — which briefly carried the company to a mind-boggling $5 trillion market capitalization last October — then feeds capital back into OpenAI through direct investment and ecosystem reinforcement.

The same dollars appear as revenue multiple times across corporate balance sheets. Growth is booked. Valuations rise. More capital is raised. The wheel spins faster.

This is circular finance, not organic end-user demand.

What appears to be explosive adoption increasingly amounts to companies paying each other with investor capital — all predicated on future enterprise AI monetization that has yet to materialize at scale.

The mathematics of the Oracle–OpenAI relationship are stark. OpenAI is on track to lose approximately $16 billion in 2025 and projects cumulative losses exceeding $100 billion before decade-end. To honor its Oracle commitment, it must grow revenue from roughly $20 billion to $100 billion within three years — a feat no software company in history has ever achieved.

Oracle, meanwhile, must finance up to $180 billion of AI-specific infrastructure to serve essentially one anchor tenant: OpenAI. Its free cash flow has turned negative. Its debt load is rising.

This is a bilateral dependency. If OpenAI cannot pay, Oracle is left with stranded assets. If Oracle cannot finance the build-out, OpenAI faces a compute crisis at precisely the wrong moment. Neither party can exit. Neither can survive the other’s failure.

Engineers call this a single point of failure. Markets usually discover them too late.

Every systemic episode has a financial correlation node. In the 2008 subprime ecosystem, it was AIG. In today’s AI ecosystem, that role is increasingly occupied by Japan’s SoftBank.

SoftBank holds massive, overlapping exposure to OpenAI, the Stargate infrastructure project, and the broader AI complex — all while operating with limited incremental funding capacity. Its balance sheet is leveraged to a single outcome: that AI adoption accelerates rapidly and smoothly within a narrow time window.

If OpenAI stumbles, SoftBank’s equity collapses. If Stargate underperforms, infrastructure capital is impaired. If both occur simultaneously, forced deleveraging becomes unavoidable — with knock-on effects across the entire AI equity universe.

The entire AI capex thesis hinges on one critical assumption: that AI adoption and monetization will explode between 2027 and 2028.

On the adoption side, competition is intensifying. ChatGPT’s growth rates are already peaking as new entrants — including Chinese open-source AI models — gain traction. On the monetization side, academic and industry data offer little comfort: 95% of enterprises deploying generative AI report no measurable return on investment. Many projects are already being scaled back or quietly abandoned.

As with the internet, the AI revolution is real, and its long-term impact will be profound. But technological revolutions are rarely linear — and they almost never reward every participant equally. Oracle itself was a darling of the internet boom in 1999, before losing nearly 80% of its market value in the 2000 collapse.

What we are witnessing today is a self-referential financial architecture, where capital circulates internally while equity markets extrapolate perfection through extreme valuations.

For now, equity markets remain optimistic. Credit markets are not. In recent days, Oracle’s credit-default-swap spreads have climbed to levels last seen during the global financial crisis.

Historically, credit markets price reality before equity investors do.

With Oracle shares falling 11% in a single session despite spectacular earnings growth, and SoftBank dropping 7.7% on the same news, equity investors may finally be starting to question whether the AI investment narrative is as solid as advertised.

This is not 2008.

But the mathematics rhyme.