From GPU Shortages to a Liquidity Crisis
The era when scaling neural networks hit a wall solely due to the physical shortage of compute accelerators is coming to an end. The new critical barrier for tech giants is simply the cost of money. The first warning signs have emerged in the financial markets: investors are no longer willing to blindly finance the infrastructure race on credit.
A striking indicator of this shifting sentiment is the situation surrounding Project Jupiter in New Mexico. The $18 billion in bonds and credit lines raised to construct a massive data center for Oracle began trading on the secondary market at a discount—at 89 to 91 cents on the dollar. Arranging banks were met with clear reluctance from institutional investors to buy up this debt.
The Oracle Paradox: Record Revenue with Negative Cash Flow
This decline in creditor confidence unfolded against the backdrop of seemingly flawless operational metrics from the corporation itself. For the first quarter of fiscal year 2027, Oracle posted compelling momentum:
- Total company revenue jumped 30%, reaching $19.3 billion;
- The cloud division surged by 121%;
- The remaining performance obligations (RPO) backlog surpassed an astronomical $664 billion.
However, the financial reality turned out to be far more complex than the glowing headlines. To turn those reported $664 billion in commitments into actual revenue, the compute clusters must first be built. As a result, Oracle's capital expenditures for the quarter reached $28.5 billion, dragging free cash flow into the red by $5.4 billion despite a solid operating cash flow of $23.1 billion.
A Closed-Loop Economy
Creditor caution is reinforced by a broader cooling in the debt market. According to analysts, the volume of new debt financing for AI projects fell by more than half in September compared to August—dropping to just $23 billion. Broadcom CEO Hock Tan openly pointed out that access to capital has become the key bottleneck to the sector's expansion.
According to estimates by Columbia University researchers, cumulative investments in AI infrastructure could reach $10.3 trillion by 2032. For such expenditures to pay off, the industry would need to generate around $3.55 trillion in net revenue every year.
Today, a significant portion of the sector's financial activity operates in a circular pattern: venture capital flows to algorithm developers, who rent compute capacity from cloud providers, who in turn order hardware from equipment makers. But as long as this chain is sustained by external investments rather than actual enterprise profits generated by end-user AI adoption, the system's financial stability remains vulnerable to interest rates.
The Technological Paradox and Depreciation Risks
Technological progress itself introduces additional uncertainty. Data centers are designed with years of operation in mind, yet:
- Algorithms are becoming more efficient. The emergence of compact, less resource-intensive models could lower the market price per unit of compute, undermining the business models of ultra-expensive data centers.
- Hardware becomes obsolete rapidly. The lifespan of server components and specialized accelerators is significantly shorter than the payback horizon of a physical facility.
The unfolding situation is reminiscent of the dot-com era: the underlying technology is undoubtedly viable and transforming the world, but expectations regarding the timelines for returns on trillion-dollar investments risk colliding with the fundamental laws of corporate finance.