The AI infrastructure boom just revealed its Achilles heel: the debt spiral required to fund it is starting to look uncomfortably familiar.
The Summary
- Broadcom is raising $60+ billion in debt to help Anthropic and other AI companies secure chips and computing power — one of the largest tech debt deals ever
- Wall Street's appetite for absorbing AI-related debt is showing strain, raising questions about circular financing in the sector
- The buildout cost of AI infrastructure is outpacing traditional financing models, creating potential systemic risk
The Signal
Broadcom's $60 billion debt raise isn't just another capital event. It's a symptom of something stranger happening in AI economics: the companies building the infrastructure are now financing the companies buying the infrastructure. Anthropic needs chips. Broadcom needs Anthropic to buy chips. So Broadcom borrows $60 billion to make sure Anthropic can afford to buy what Broadcom is selling.
This is circular financing at scale. And it's becoming the norm, not the exception, in AI infrastructure.
"The cost of funding it all" is the question Wall Street hasn't figured out how to answer yet.
The numbers tell the story. In 2025, tech companies issued roughly $180 billion in debt, with AI infrastructure accounting for nearly 40% of that total. In 2026, that figure is projected to hit $250 billion, with AI's share climbing to 55%. The debt isn't funding R&D or product development in the traditional sense. It's funding the physical buildout: data centers, chip supply chains, power infrastructure, cooling systems.
The buildout is capital-intensive in ways software never was. Training a frontier model costs $500 million to $1 billion. Running inference at scale costs another $300-500 million annually. The chips to do it cost $40,000 each. The data centers to house them cost $5-10 billion per facility. The power contracts to run them are measured in gigawatts.
Key challenges creating the debt spiral:
- Model training costs are growing 3x faster than revenue from AI products
- Chip supply constraints mean buyers have to commit capital years in advance
- Hyperscalers are locked in a prisoner's dilemma: slow down and lose the race, or keep spending and hope revenue catches up
The circular financing concern is real. When chipmakers finance their own customers' purchases, you're not seeing demand. You're seeing vendor financing dressed up as a market. It works until the music stops. And the music stops when one of three things happens: the debt becomes too expensive to roll over, the AI products don't generate enough revenue to service the debt, or Wall Street decides the risk isn't worth the return.
Erica Klauer's point about the "AI buildout being so challenging" understates it. This isn't challenging. This is unprecedented. No technology wave in history has required this much upfront capital before proving out its business model at scale. The internet boom was funded by equity, not debt. Cloud infrastructure was built incrementally, as revenue proved out. AI is different. It requires the full stack before you can even start.
The Implication
Watch the debt markets, not the equity markets, if you want to know when AI's growth story starts to crack. When investment-grade tech debt starts pricing in real risk premiums, that's your signal that Wall Street's patience is thinning.
For now, the bet is that AI revenue will catch up to AI spending. But the longer that gap stays open, the more this starts to look like the leverage games that preceded every major correction in the last 30 years. The companies building agents and deploying models need to show cash flow, not roadmaps. The debt markets will force that discipline long before the equity markets do.