AI Briefing
KO

Nobody Knows the True Value of Used GPU Clusters

·2026.07.23 09:00

Key point

Amid the AI infrastructure boom, tens of billions of dollars in debt are collateralized by GPU clusters whose value shifts depending on operational condition, raising serious risk concerns.

Details

If xAI defaults, Apollo Global Management has the right under the contract terms to seize the 200,000-GPU Colossus cluster near Memphis and lease it out to other AI companies. The problem is that nobody can say exactly how much that cluster is actually worth.

Unlike a building, the value of a GPU cluster depends entirely on its operational condition. Modern data center GPUs fail at a rate of roughly 9% annually, meaning about 50 units are replaced per day for a 200,000-GPU fleet. Even more serious are undetectable failures like Silent Data Corruption (SDC), which can quietly contaminate multi-node consensus training outputs.

Operations teams hold undocumented operational know-how — which racks overheat in summer, which cooling loops are unstable, and so on. This knowledge is an asset that can 'walk out the door' with the team. Creditors have no external means of assessing this operational condition.

The xAI Colossus 2 SPV structure consists of roughly $7.5 billion in equity and $12.5 billion in debt, with the debt collateralized by the GPUs themselves rather than xAI's overall finances. GPU prices fall rapidly after launch, and demand for older models collapses when new models are released. The used-market price of the H100 has already dropped to less than half of its price when new.

From a creditor's perspective, there is no standard for valuing the collateral. Neither GPU book value, used market price, nor income capitalization value serves as a clear benchmark. As the AI infrastructure debt market grows rapidly, a system for evaluating collateral value does not yet exist.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.