If an AI cloud company goes bankrupt, its Nvidia GPUs do not simply disappear into a warehouse sale.
What happens depends on who legally owns the hardware, whether it was pledged as collateral, which subsidiary borrowed the money and which creditors sit highest in the capital structure.
That distinction matters because AI infrastructure companies increasingly finance GPU purchases with secured loans tied directly to computing equipment and customer contracts. Coinpaper’s look at who is paying for the AI boom shows how debt has become central to the buildout as CoreWeave, Nscale and other operators raise billions before the infrastructure generates enough cash to fund itself.
CoreWeave offers a useful real-world example. In its latest SEC filing, the company says some debt facilities are secured by substantially all assets of specific subsidiaries, including computing equipment and related collateral.
So if such a borrower defaults, the lender may have a much stronger claim on the GPUs than ordinary creditors or shareholders.
Secured Lenders Usually Stand Closest to the Hardware
A secured loan is different from ordinary corporate debt because the lender has a legal claim against specified assets.
Imagine an AI cloud company creates a subsidiary that owns $1 billion of GPU servers and borrows $700 million against them. If that borrower fails, those GPUs could be sold, refinanced or transferred as part of a restructuring, with proceeds used to satisfy the secured debt attached to them.
CoreWeave has used multiple layers of financing to fund expansion, including convertible debt. A recent CoreWeave financing deal highlighted just how much outside capital is required to buy GPUs, secure power and build data-center capacity before customers fully pay for the compute.
Nebius has also used secured financing backed by deployed GPU infrastructure and contracted customer cash flows. Its earlier AI cloud debt raise showed the same basic tension: huge demand can exist at the same time as significant financing risk.
| Claimholder | Typical position |
|---|---|
| Secured lender | First claim on pledged GPUs or assets |
| Equipment lessor | May reclaim leased hardware |
| Unsecured bondholder | Paid from remaining estate assets |
| Trade/vendor creditors | Compete with other unsecured claims |
| Shareholders | Usually last |
The GPUs May Not Always Belong to the AI Company
Some operators own GPUs outright. Others lease equipment, use vendor financing or place hardware inside special-purpose entities.
That changes the outcome.
If the GPUs are leased, the lessor may retain ownership and seek to recover them after a default. If the hardware sits inside a bankruptcy-remote subsidiary, creditors of that subsidiary may have priority over creditors of the parent company.
This is why AI cloud bankruptcy increasingly resembles infrastructure finance rather than the collapse of a normal software company.
Coinpaper’s broader analysis of AI data-center financing explains how banks, private-credit funds, bondholders and special-purpose vehicles can each sit at different points in the ownership and repayment chain.
The Hard Part Is Recovering the GPUs’ Value
Even when lenders have the right to seize GPUs, that does not guarantee they recover everything they lent.
High-end accelerators can lose value quickly when newer generations arrive. A lender that financed H100 servers during peak demand could recover less if newer Blackwell or Rubin systems become more attractive.
Recovery therefore depends on more than the chips themselves. Creditors also care about:
- hardware age;
- secondary-market prices;
- customer contracts tied to the machines;
- available power;
- data-center location;
- whether the servers can be redeployed quickly.
This is one reason AI debt risk becomes more serious if utilization or customer revenue falls before the debt is repaid.
In many cases, the best outcome may not be selling individual GPUs at all. A buyer could acquire the data center, hardware and customer contracts together, preserving more value than a liquidation.
