The AI data center debt machine is priced for perfection and built for chaos. Ed Zitron's 'Hater's Guide to AI Debt Part 2' argues that somewhere between $100 billion and $150 billion in financing has flowed to data center projects run by companies with no experience building AI infrastructure, securing power grids, or managing the thermal and mechanical complexity of condensing city-scale energy into chip-dense facilities. The debt structures, covered in Part 1, include project financing tied to customer lease payments, convertible notes, and off-balance-sheet obligations totaling trillions. The core problem in Part 2 is not the form of that debt. It is what happens when that debt meets reality.
Zitron's central thesis is a doom loop: delays drive up costs, rising costs require more debt, more debt becomes more expensive, and none of this math works for most projects. Every data center is its own infrastructure monster shaped by geography, weather, power access, and cooling requirements for hardware that did not exist five years ago. The finance industry has treated these projects like commercial real estate. They are not. The entire cloud revenue model for Amazon, Google, and Microsoft is also functionally dependent on compute spend from OpenAI and Anthropic, two companies that are themselves debt-financed.
Read the full piece for Zitron's breakdown of why the economics of building and paying off an AI data center are, by his account, near-impossible for the majority of active projects. The argument is not that AI infrastructure will fail to get built. It is that the people underwriting it have systematically mispriced the risk of building it. Part 1 is linked throughout for context on the debt mechanics. Part 2 is where the structural fragility becomes a financial reckoning.
[READ ORIGINAL →]