Nobody knows what a used GPU cluster is worth
17 hours ago
- #GPU collateral
- #AI infrastructure
- #debt financing
- If xAI defaults on its $5B debt, lenders like Apollo can seize its 200,000 GPU cluster, Colossus, and rent it out, but the cluster's value depends heavily on operational factors that lenders cannot see.
- GPU clusters require constant maintenance; at scale, failures are routine (e.g., ~50 GPUs fail daily in a 200,000-GPU cluster), and operational knowledge resides with the team, making the asset's value as collateral uncertain.
- AI infrastructure is increasingly financed by GPU-collateralized debt (e.g., via SPVs), with lenders charging high premiums (e.g., 8.5% above benchmark) due to unmeasurable risks, as there is no mature price discovery or hedging market for GPUs.
- GPUs have three values: face value (based on depreciation schedules), liquidation value (lower in distress), and going-concern value (dependent on operational handoff), but only face value is currently priced, creating a significant risk gap.
- Key players like KKR avoid GPU-backed debt, investing instead in physical infrastructure (e.g., data centers), signaling that chips may not be durable assets, while the lack of standardized valuation tools (e.g., futures, secondary markets) keeps financing costs high.