The Anatomy of the AI Bubble: Circular Money, Fragile Foundations, and What Comes Next

The AI boom of the mid-2020s looks impressive on the surface. Hyperscalers and AI labs are pouring capital into infrastructure at a scale never seen before. Amazon, Microsoft, Google, Meta, and OpenAI together are on track to spend roughly $700 billion a year building data centers, power infrastructure, and networking — betting that demand for AI compute will explode and justify every dollar.

The uncomfortable reality is that the revenue needed to support this spending has not arrived yet. OpenAI alone is burning through approximately $14 billion annually and has already signaled that it may struggle to pay for the massive computing capacity it has already contracted. The money keeping the machine running comes from three sources: the giants’ own cash, heavy borrowing (Oracle’s debt has already crossed $100 billion), and fresh venture capital. This is not yet self-sustaining growth — it is forward-loaded investment hoping future profits will catch up.

How the money actually moves: circular financing

The most striking feature of the current setup is how capital circulates inside a relatively small group of companies. This pattern is often called circular financing.

Here is the typical loop:

Nvidia invests in or provides favorable terms to an AI company such as OpenAI.
OpenAI uses that capital (or related funding) to buy GPUs from Nvidia and cloud capacity from Microsoft or Oracle.
Microsoft and Oracle record those purchases as revenue, which lifts their reported numbers, stock prices, and borrowing capacity.
With higher valuations and easier access to debt or equity, they invest or lend even more into the ecosystem.
The same dollars effectively travel around the circle, appearing as genuine growth at every stop.

Analysts tracking these interconnected transactions estimate their total size at nearly $800 billion. On paper, everyone’s revenue and market cap are rising. In reality, a large share of the activity is the same capital being passed around and counted multiple times. As long as the loop keeps spinning and the promise of future revenue stays credible, the system looks healthy. The moment that promise weakens, the same circular nature turns from amplifier to accelerator of collapse.

The hardware time bomb

There is a second structural problem that receives less attention: the mismatch between asset life and debt duration.

Cutting-edge GPUs and AI accelerators lose significant economic value within three to four years because newer, more efficient chips arrive rapidly. Yet much of the financing used to buy this hardware is long-term debt — often ten years or more. Companies are effectively borrowing decade-long money to purchase equipment that may be economically obsolete in half that time. When utilization rates or pricing power disappoint, the combination of high fixed debt service and rapidly depreciating assets becomes extremely painful.

How the unwind could play out

Bubbles of this kind rarely pop from a single dramatic event. They usually start with one link in the chain weakening and then propagate through the interdependencies.

A plausible sequence:

Demand disappoints or a major customer wobbles. OpenAI, for example, cuts orders or delays payments because its own revenue growth slows.
Cloud providers feel the hit immediately. Oracle, Microsoft, and others see their largest AI customer pull back. Earnings forecasts are missed, stock prices fall sharply (Oracle has already demonstrated how quickly sentiment can turn), and the cost of new borrowing rises.
Nvidia loses orders. With fewer chips being bought, its growth narrative cracks. Because Nvidia represents such a large portion of major indices, the broader market feels the pain.
Capex gets slashed across the board. Hyperscalers slow or cancel data-center construction. Suppliers of chips, servers, power equipment, transformers, and even electricity generation feel the contraction.
Systemic spillovers. Pension funds, index funds, and banks with exposure to these highly correlated names absorb losses. The concentration of AI-related market value makes the impact unusually wide.

The circular structure makes this cascade more dangerous than a conventional slowdown. In a normal industry, one company’s trouble does not automatically destroy its suppliers and customers at the same time. Here the participants are financially and operationally entangled: Nvidia’s revenue depends on OpenAI’s spending power, OpenAI’s solvency depends on Microsoft and Oracle, and Microsoft’s and Oracle’s valuations are heavily supported by the AI story that Nvidia powers. When one link breaks, the feedback runs in reverse and pulls everyone down together.

What happens to the price of AI?

Today, tokens are frequently sold below their true marginal cost. The difference is covered by investor subsidies in the form of cheap capital and high valuations. This keeps usage high and helps companies gain market share, but it is not economically sustainable on its own.

If the funding spigot tightens:

Short term: Providers can no longer afford to subsidize usage. Premium subscriptions and enterprise contracts would likely rise in price. Marginal or poorly differentiated services would shut down or be acquired. Token prices would move upward.
Longer term: Several counter-forces would push prices down. Distressed sales of used GPUs would create a secondary market. Model efficiency keeps improving (less compute per useful output). Open-source models and cheaper Chinese hardware would compete aggressively, especially for on-premises or sovereign deployments. The net effect is likely a period of higher prices followed by renewed downward pressure as supply adjusts and technology advances.

The bottom line

The current AI investment cycle is not a simple story of visionary companies building the future. It is a tightly coupled system of circular capital flows, aggressive leverage against fast-depreciating hardware, and valuations that rest on the assumption that revenue growth will accelerate dramatically and soon.

That assumption may yet prove correct. AI capabilities are advancing quickly, and real enterprise adoption is happening. But the financing structure itself has introduced fragility that did not exist in previous technology buildouts. The same interdependencies that created rapid apparent growth can amplify any disappointment into a much larger correction.

Watch the metrics that actually matter: sustained revenue growth at the AI labs that is not dependent on further rounds of circular investment, rising utilization rates in the new data centers, and the ability of the biggest spenders to service their debt without constantly raising more capital at ever-higher valuations. If those indicators keep improving, the infrastructure being built today will look prescient. If they stall, the circular nature of the financing will make the unwind faster and more painful than most observers currently expect.

The bubble, if it bursts, will not be because AI failed as a technology. It will be because the financial engineering that funded its current scale ran ahead of the economics that can support it.

---