My client David Wolf built the cost model of a $1 billion AI data center. Corrected and checked on Bloomberg, it says a GPU-hour must earn $3.83 if the chips last three years. Credit markets have started to price it, and French banks are exposed.
A huge thank you for taking the time to dig into my model, challenge it, correct it, and turn it into something genuinely actionable.
For me, one of the most valuable things to watch from here will be the H100, H200 and Blackwell rental indices, and how quickly each new generation puts pressure on the previous one.
Combined with the framework in the article, this should give readers a very good window into what is happening beneath the surface of the AI data center economics, well before everything necessarily shows up in the financial statements.
Thank you for this article, as rigorous as it is insightful. You make a complex subject accessible and show how to test a thesis before defending it, which is very helpful for my own analysis. A nod as well to David's work, which served for your demonstration. Thank you for sharing.
A ten-year loan backed by a rapidly depreciating GPU fleet creates a problem that accounting earnings alone won’t reveal.
The critical variable is cash flow available for debt service after replacement capex.
A data center can report positive EBITDA while its GPUs lose economic competitiveness, forcing another investment cycle before the original debt has matured.
That creates three simultaneous pressures: declining collateral value, rising maintenance and replacement capital requirements, and deteriorating debt-service coverage.
The real stress test is not whether today’s GPU-hour pricing covers operating costs.
It’s whether the project can fund its next hardware generation, service existing debt and survive lower utilization without fresh equity.
Once refinancing depends on continuously rising compute demand, technological obsolescence becomes a credit transmission mechanism.
The GPU depreciation problem is ultimately a duration mismatch between technological assets and financial liabilities.
Fantastic article!
A huge thank you for taking the time to dig into my model, challenge it, correct it, and turn it into something genuinely actionable.
For me, one of the most valuable things to watch from here will be the H100, H200 and Blackwell rental indices, and how quickly each new generation puts pressure on the previous one.
Combined with the framework in the article, this should give readers a very good window into what is happening beneath the surface of the AI data center economics, well before everything necessarily shows up in the financial statements.
Thanks again for the great work.
David
Thank you for this article, as rigorous as it is insightful. You make a complex subject accessible and show how to test a thesis before defending it, which is very helpful for my own analysis. A nod as well to David's work, which served for your demonstration. Thank you for sharing.
A ten-year loan backed by a rapidly depreciating GPU fleet creates a problem that accounting earnings alone won’t reveal.
The critical variable is cash flow available for debt service after replacement capex.
A data center can report positive EBITDA while its GPUs lose economic competitiveness, forcing another investment cycle before the original debt has matured.
That creates three simultaneous pressures: declining collateral value, rising maintenance and replacement capital requirements, and deteriorating debt-service coverage.
The real stress test is not whether today’s GPU-hour pricing covers operating costs.
It’s whether the project can fund its next hardware generation, service existing debt and survive lower utilization without fresh equity.
Once refinancing depends on continuously rising compute demand, technological obsolescence becomes a credit transmission mechanism.
The GPU depreciation problem is ultimately a duration mismatch between technological assets and financial liabilities.