Nobody knows what a used GPU cluster is worth

There's a strange clause buried in xAI's June 2025 debt facility. If the company defaults on the $5 billion arranged by Morgan Stanley, Apollo Global Management and Diameter Capital can take over Colossus, the 200,000-GPU cluster outside Memphis, and rent it out until the loan gets paid back. On paper that sounds like a tidy backstop. In practice it means Apollo would inherit a machine that fails roughly 50 times a day and depends entirely on an operations team that just walked out the door.

Modern GPU clusters don't behave like buildings or ships or airplanes. Meta's own Llama 3 training report logged 419 unplanned disruptions across 16,384 H100s in 54 days, and extrapolating that failure rate to a million-GPU cluster gets you a failure every three minutes, according to Epoch AI. Most of these aren't dramatic crashes. The worst kind, silent data corruption, lets a bad chip quietly poison a training run for days before anyone notices. Keeping a cluster productive is a craft that lives in people's heads: which racks overheat in July, which cooling loop has been temperamental since the last firmware patch. None of that is written down anywhere a lender could read it.

And yet this is exactly the collateral now backing tens of billions in debt. CoreWeave alone carries $18.8 billion in GPU-backed loans. Anthropic's FluidStack arrangement runs to $50 billion, with Google backstopping lease payments. Compare the infrastructure supporting this to what exists for aircraft, which have ISTAT-certified appraisers and a secondary market going back to the 1970s, or oil, which has had a forward price curve since 1983. GPUs have Silicon Data's rental index, launched in 2024, and a startup called Ornn AI that raised $5.7 million to build a derivatives exchange. That's essentially the entire toolkit.

The pricing gives the game away. CoreWeave's GPU loans carry an 8.5-point premium over benchmark rates, versus 1 to 2 points for aircraft debt. That gap is what lenders charge when they're guessing rather than measuring. And the guesses have been wild — H100 hourly rental rates crashed from $8 to $1.70 between early 2024 and October 2025, then jumped 40% back to $2.35 by March 2026 on unexpected inference demand. Depreciation assumptions are just as scattered: CoreWeave books GPUs over six years, Nebius uses four years for the same hardware, and Michael Burry argues hyperscalers will collectively understate depreciation by $176 billion through 2028.

What's actually missing is any real distinction between three different numbers — what the spreadsheet says the chips are worth, what a buyer would pay in a fire sale (secondary market data suggests 30 to 50% of face value if multiple neoclouds default at once), and what the cluster is worth as a running operation to whoever takes it over next. That third number, going-concern value, is the one nobody prices, because it depends on operational knowledge that isn't transferable. It's telling that KKR, the most aggressive infrastructure buyer in the sector, sticks to buildings, power, and land rather than chip-collateralized debt. Peter Thiel dumped his entire NVIDIA position last year. Neither bet is subtle.