Nobody Knows What a Used GPU Cluster Is Worth

If xAI defaults on its debt, Apollo Global Management ends up in the GPU rental business. That's the clause in the $5 billion debt facility arranged by Morgan Stanley in June 2025. Lenders can seize Colossus—xAI's 200,000 GPU cluster outside Memphis—and rent it out until the loan is repaid. The question nobody is answering: what is that cluster actually worth?

A GPU cluster isn't a building. Its value depends on how it's been provisioned, how it's currently performing, and whether the operations team that knows its quirks is still around. All of that sits off the lender's balance sheet, invisible to the people pricing the debt.

The cluster does not run itself

At scale, hardware fails constantly. Meta's Llama 3 training run on 16,384 H100s saw 419 disruptions over 54 days—148 GPU failures, 72 HBM3 memory failures. At 200,000 GPUs, that's about 50 failures per day. At xAI's million-GPU target, Epoch AI projects a failure roughly every three minutes.

The expensive failures don't look like failures. Silent data corruption (SDC) produces wrong answers without crashing anything—a multi-day training run completes normally, but the model weights are quietly poisoned. Cascading failures crash training jobs across thousands of GPUs, costing days of compute. Then there are thermal throttles, ECC memory errors, NVLink flaps, GPUs falling off the bus.

NVIDIA built NVSentinel because traditional monitoring detects these problems but rarely fixes them. Crusoe built AutoClusters because queue wait time is the largest controllable variable in cluster goodput. Without these tools, remediation timelines run hours to days.

The operations team knows which racks run hot in summer, which cooling loops have been flaky since the last firmware update, which jobs to reroute when a node degrades but hasn't failed yet. None of that knowledge is written down. It lives in the team. That's the asset serving as collateral for tens of billions in debt.

How the chips became the collateral

In the last eighteen months, AI infrastructure financing shifted from corporate debt to chip-backed debt. The xAI Colossus 2 SPV is the cleanest example: roughly $7.5 billion in equity (up to $2 billion from NVIDIA itself) and $12.5 billion in debt. Apollo and Diameter Capital Partners sit on the debt tranche. The debt is collateralized by the chips, not by xAI's balance sheet.

Pricing reflects the uncertainty. xAI's $5 billion round was priced at up to 12.5%. CoreWeave's GPU-backed deals priced at roughly 8.5% above benchmark. For comparison, a typical aircraft loan prices at 1–2 points above benchmark. The extra 6–7 points is what lenders charge to bear a risk they cannot measure.

CoreWeave alone holds $18.8 billion in GPU-collateralized debt across multiple SPVs. FluidStack's $50 billion deal with Anthropic uses a different wrapper but the same logic. Every neocloud and most major AI labs are now financed this way.

What real collateral looks like

Every other major asset class used as collateral at this scale has decades of price discovery infrastructure. Aircraft have ISTAT-certified appraisers, a global registry, standardized maintenance logs, and an active secondary market. Ships have BICA. Cars have NADA. Oil has had a forward curve since the early 1980s.

GPUs have Silicon Data's H100 Rental Index on Bloomberg terminals (launched 2024) and Ornn AI ($5.7 million raised in October 2025 to build the first regulated exchange for GPU compute derivatives). That's the entire price discovery infrastructure for an asset class now backing tens of billions in debt.

H100 hourly rental rates went from roughly $8 in early 2024 to $1.70 by October 2025, then surged 40% to $2.35 by March 2026 on inference demand nobody had priced in. No aircraft lender would underwrite five-year debt against an asset whose price swings like that without a way to hedge it. They would not be allowed to.

The spread should compress as the market matures. Hedging instruments will appear. Residual value curves will standardize. Secondary markets will deepen. When that happens, the cost of capital for AI infrastructure drops meaningfully—and the companies that benefit are the ones positioned to access cheap debt once the financing infrastructure catches up to the asset class.