Ivan, sharp analysis — the duration mismatch between debt tenure and chip earning power is rarely framed this cleanly.
One hypothesis worth adding: the bet may not be on inference margin at all — it's on controlling energy infrastructure. A 30-year data center is land, power capacity, and grid interconnection. The chips rotate. The building doesn't.
Jevons also hasn't been repealed. If inference drops 10x in cost, demand could grow 100x in volume. Cloud economics already ran that script.
The debt may not be justified by today's model economics — but by a calculated bet that the rules of the game change before the bills come due.
Oleg Kofanov thank you, fair addition, and you’re right on the durable layer.
The question is where the capital sits. In a hyperscale build, silicon is the single biggest line, roughly half to 60% of the spend (McKinsey puts it near 60%). The shell and the interconnection are the smaller share, and much of the debt is collateralized by the GPUs themselves. So the asset carrying the financing is the one with a 2-3 year economic life running against a 5-6 year schedule. The durable layer keeps its value. The part the lenders are exposed to does not.
On Jevons I agree, and I made that case in an earlier piece on inference decentralizing demand as cost falls. The still open question is sequencing. Jevons governs the direction of demand over years. Debt service runs on a fixed timeline. The bet pays if the volume arrives before the refinancing does.
That timing is why I keep Sovereign Compute to a 1-2 year horizon. Where AI economics settle at the end of the decade is complete guesswork. The next refinancing window is something you can underwrite, and the last few sessions suggest the market is starting to weigh the same question.
Oleg Kofanov wrote:
Ivan, sharp analysis — the duration mismatch between debt tenure and chip earning power is rarely framed this cleanly.
One hypothesis worth adding: the bet may not be on inference margin at all — it's on controlling energy infrastructure. A 30-year data center is land, power capacity, and grid interconnection. The chips rotate. The building doesn't.
Jevons also hasn't been repealed. If inference drops 10x in cost, demand could grow 100x in volume. Cloud economics already ran that script.
The debt may not be justified by today's model economics — but by a calculated bet that the rules of the game change before the bills come due.
Oleg Kofanov thank you, fair addition, and you’re right on the durable layer.
The question is where the capital sits. In a hyperscale build, silicon is the single biggest line, roughly half to 60% of the spend (McKinsey puts it near 60%). The shell and the interconnection are the smaller share, and much of the debt is collateralized by the GPUs themselves. So the asset carrying the financing is the one with a 2-3 year economic life running against a 5-6 year schedule. The durable layer keeps its value. The part the lenders are exposed to does not.
On Jevons I agree, and I made that case in an earlier piece on inference decentralizing demand as cost falls. The still open question is sequencing. Jevons governs the direction of demand over years. Debt service runs on a fixed timeline. The bet pays if the volume arrives before the refinancing does.
That timing is why I keep Sovereign Compute to a 1-2 year horizon. Where AI economics settle at the end of the decade is complete guesswork. The next refinancing window is something you can underwrite, and the last few sessions suggest the market is starting to weigh the same question.