Ares Management's alternative credit newsletter ran an exhibit titled "Interconnected exposures": 32 financings, eight counterparties, and one line at the bottom — "common dependency: sustained AI capital spending." Michael Burry posted it, and it's being read as a crash signal. The more useful reading is about what each deal actually rests on, and for inference that comes down to what the chips are worth.
A web, not eight companies
The exhibit sorts 32 financings into seven markets: corporate bonds, syndicated project and construction loans, 144A private placements, GPU and asset-backed finance, high-yield, chip SPVs and ABS/CMBS. Read by size, it is mostly investment-grade paper. The corporate bonds alone, from Amazon, Alphabet, Meta, Oracle, Nvidia, SpaceX and AMD, add up to roughly $290 billion on the exhibit's own figures. Ares labels the gold band on the left as transactions designed for insurance buyers. The lines show who has to perform for each deal to work, who is backstopping it, and the circular ones between the eight names: Nvidia investing in OpenAI and Anthropic, OpenAI's compute commitments to Oracle, and so on.
That structure is the real finding. Few of these deals depend on a single company. They depend on the same handful of counterparties continuing to spend, and on each other.
Where chips sit in the web
It would be wrong to say most of this debt is secured by GPUs. The directly GPU-backed deals on the map — CoreWeave's delayed-draw loans and the Lambda, Nebius and NScale facilities — add up to about $17 billion. The larger exposure to chip value is indirect. The chip SPVs (a Broadcom vehicle at $60 billion or more, and a $35 billion Anthropic structure) are built around hardware. Several deals carry residual value guarantees, including a Nvidia-backed one capped near $105 billion on OpenAI's Ohio financing. And the data-center leases behind the project and 144A deals only pay if the tenants keep running paying workloads.
That is why a chip's value still matters. Burry cited an Ares chart showing H100 residual value down 51% over three years; we haven't seen Ares's method, so treat it as a headline number. The mechanism holds either way: a GPU is worth the cash it earns over its remaining life, and that cash comes from inference. Utilization decides whether it earns anything, and price per token decides how much. Cold-start and scheduling technology, like what Nebius bought with Inferize, moves the first one. New chip generations move the second.
Why this isn't a crash call
A web of financing isn't evidence the loans go bad. They go bad if the revenue underneath doesn't arrive in time, and the picture is mixed in a way we've been tracking. OpenAI's annualized revenue is reported near $70 billion by Axios, a figure the company hasn't confirmed to us. a16z's data puts paid consumer adoption at 2.2% of US households. Bain estimates AI needs $6 trillion a year in revenue by 2031, with $4.2 trillion still from products that don't exist. Those facts support Burry's worry about timing, and they also show demand growing quickly from a small base.
Our read is the same one we gave on the household data. The financing is priced for a future running ahead of today's revenue, and the open question is whether the gap closes before the collateral reprices. That is a statement about timing and risk, not a forecast of failure. Burry's view that this resembles 2000 is his opinion, and he has a financial interest in being right.
What we're watching
Burry has said more of his analysis is coming, and we'll read it against the Ares material directly. We want to see how lenders actually value GPU collateral and residual value guarantees, and whether those assumptions get marked down as new chips ship. And we're watching price per token, because it links all of this back to inference: if it keeps falling faster than utilization rises, the collateral weakens no matter how many users sign up.