← Back

Thesis: The AI Buildout Has a Capital Structure Problem

I build with AI every day, and I think it is the most important technology wave of my lifetime. I also think the way the buildout is being financed carries real risk. Both can be true.

This is a personal macro thesis I have been developing and stress-testing. Some arguments survived scrutiny. Some did not, and I think it is worth being clear about which is which.

What survived

1. Circular vendor financing. Some of the demand for AI infrastructure is funded by the suppliers of that infrastructure: chipmakers and cloud providers investing in AI companies that then spend that money back on their chips and compute. That makes revenue look stronger than end-customer demand alone would justify, and it ties the health of suppliers and customers together.

2. The capex surge is increasingly debt-funded. Early data center spending came mostly from the cash flows of the largest tech companies. More of it now runs through debt, private credit, and off-balance-sheet structures. Equity-funded overbuilding hurts shareholders. Debt-funded overbuilding can spread to lenders and the wider financial system.

3. Open-weight models compress pricing. If capable open-weight models keep closing the gap with frontier models, the price of a unit of intelligence keeps falling. That is great for builders like me. It is a problem for anyone whose return assumptions depend on premium pricing for model access holding up long enough to pay back the infrastructure.

What I corrected

I originally leaned on market concentration figures and simple investment-versus-revenue ratios. When I checked them carefully, some of the numbers I had were wrong or framed in a misleading way. Comparing today's capex to today's AI revenue ignores that infrastructure is built ahead of demand. I dropped those arguments rather than keep the ones that sounded best.

How I hold it

This is not a call that AI is a bubble that pops tomorrow. It is a view that the technology can be transformative while the capital structure financing it is fragile, and that the risk sits in the financing, not the technology.

As a builder, my takeaway is practical: assume intelligence keeps getting cheaper, design products whose value does not depend on any one model provider, and keep AI cost per customer low enough that falling prices are a tailwind.