Nobody knows what a used GPU cluster is worth (ciphertalk.substack.com)

The ongoing uncertainty surrounding the valuation of used GPU clusters is raising alarm bells in the AI/ML community. A recent contract involving xAI and its $5 billion debt facility could lead Apollo Global Management to take over Colossus, a massive GPU cluster, if xAI defaults. The catch lies in accurately valuing these GPU clusters, which are not akin to traditional assets; their worth fluctuates based on performance, provisioning, and operational expertise, which is often undocumented and known only to the onsite team. This situation spotlights a significant challenge in the burgeoning AI infrastructure market, where billions in debt are collateralized by GPU technology that has unpredictable depreciation rates and failure modes. As operational issues like silent data corruption or cascading failures plague GPU reliability, lenders face immense risk without a clear means to assess collateral value accurately. The lack of a robust price discovery mechanism for GPUs—unlike established asset classes—compounds this issue, making it difficult to determine both liquidation and going-concern values. As the market matures, the potential for creating standardization and hedging tools could lower financing costs and reshape who can exploit the benefits of AI infrastructure at scale.