There are now more than $20 billion in outstanding loans collateralized by NVIDIA GPUs. CoreWeave has $14.2 billion in total debt. Fluidstack has secured up to $10 billion in borrowing capacity from Macquarie and other lenders. Lambda Labs securitized $500 million in the first GPU-backed asset-backed security. Crusoe has $425 million in GPU-backed debt. BlackRock, Blackstone, PIMCO, Carlyle, JP Morgan, and Macquarie are all lenders.
Every one of these structures rests on the same assumption: that NVIDIA GPUs will hold enough value over five to six years to justify the loans they’re backing. And right now, some very serious people are betting that assumption is wrong.
The Smart Money Is Nervous
The spot market is the canary. H100 rental prices have collapsed from $7-10/hour in early 2024 to $2-4/hour by late 2025, with some providers offering rates below $2.00/hour. That’s a 50-70% decline in revenue-generating capacity per GPU-hour, and the Blackwell generation is just starting to ship.
The depreciation schedules that underwrite all this debt are already a contested number. Amazon shortened its server useful life estimate from six years to five in February 2025, citing the “increased pace of development in artificial intelligence and machine learning,” taking a $700 million operating income hit and booking $920 million in accelerated depreciation in Q4 2024 alone. Meta, meanwhile, extended its estimate to 5.5 years and booked a $2.9 billion depreciation reduction. CoreWeave maintains a six-year schedule. Microsoft and Alphabet hold at six.
When the largest technology companies in the world can’t agree on how long their own GPUs are worth something, the market is telling you that the core collateral assumption behind $20 billion in lending is genuinely uncertain.
Famed short seller Jim Chanos told Yahoo Finance that GPU-backed business models are “just loss-making enterprises using an asset that’s depreciating as their collateral” and predicted outright debt defaults. Michael Burry has been shorting NVIDIA, arguing the actual useful economic life of GPUs is closer to two to three years. Satya Nadella said in November 2024: “I didn’t want to go get stuck with four or five years of depreciation on one generation.”
These aren’t fringe voices. They’re pattern-matching to a scenario where the five-to-six-year depreciation assumption, which is the load-bearing wall of the entire GPU-backed lending market, turns out to be structurally unsound.
The Development Nobody in Credit Is Pricing
Dean Ball argued this week that the frontier labs have begun automating large fractions of their research and engineering operations. Within a year or two, the effective workforces of each lab will scale from thousands to tens of thousands, and then hundreds of thousands, most of whom will neither sleep nor eat. Their only objective will be to make themselves smarter.
This is not speculative futurism. It is an explicit, public milestone on the roadmap of every frontier lab. OpenAI envisions hundreds of thousands of automated research “interns” within about nine months. Lab staff already say that AI models write most or all of their code. The models currently fall short on sustained multi-day experimental execution and on generating novel research hypotheses, but the direction of travel is obvious.
Ball sketches a range of outcomes. In the bear case, automated researchers merely find the same algorithmic improvements human researchers would have found, just faster. In the bull case, they discover a vast field of low-hanging fruit that we simply didn’t have enough researcher time to pluck. Either outcome accelerates the rate of algorithmic efficiency gains. The question is by how much.
The baseline is a number that Dario Amodei has cited publicly: with human-driven research and engineering, individual labs achieve approximately 400% algorithmic efficiency improvements per year. That means the same amount of compute delivers four times the capability year over year, without new chips. These gains come from everywhere: architecture tweaks, training optimizations, data pipeline improvements, inference tricks. They’re the grinding, incremental work that constitutes the majority of what happens inside frontier labs.
Now multiply the workforce doing that grinding by 10x. Or 100x. Ball’s range spans from modest acceleration (maybe 800% annual efficiency gains) to transformative (4,000%). Even the conservative end of that range has profound implications for anyone holding GPU-backed debt.
The labs themselves are telling you, publicly, on the record, that they are about to massively accelerate the pace at which their software improves. And here’s the part that nobody in credit seems to be working through: when the software improves faster, the hardware it runs on becomes economically obsolete faster. There is a direct, mechanical link between the rate of algorithmic improvement and the collateral value of every GPU sitting in every data center backing every one of those $20 billion in loans.
That link, between recursive AI improvement and GPU collateral impairment, is what the rest of this piece unpacks.
