What a special-purpose vehicle is, and why nobody is buying the AI chips any more.

A company wants to put an enormous, expensive asset to work. Buying it outright would consume cash it would rather deploy elsewhere and would load its balance sheet with debt, which raises its borrowing costs across everything else it does. Meanwhile there are investors who would happily lend against that asset — but they want a clean claim on the asset itself, not a general claim on the company.

A special-purpose vehicle resolves both wants at once.

What an SPV actually is. A separate legal entity, created to do one job and nothing else. It has its own balance sheet. It borrows money, buys a specific asset, and earns revenue from that asset. It does not run a business in any broader sense — no other operations, no unrelated liabilities. Its narrowness is the entire point: lenders can assess one asset and one revenue stream, rather than an entire corporation.

This is old technology. Aircraft, ships, container fleets, toll roads and power stations have been financed this way for decades. The assets share a profile: expensive, long-lived, hard to move, and producing predictable contracted income.

How the pieces fit together. In the AI version currently taking shape:

- Investors — private credit funds and banks — lend money to the vehicle

- The SPV uses that money to buy chips and networking equipment from a manufacturer

- The SPV owns the hardware

- An AI lab leases the compute and pays to use it

- Those lease payments service the debt

The lab never owns the machines. It buys access to them.

Senior and junior debt, and why the split exists. Lenders to an SPV are not all in the same position. Senior debt is repaid first and is usually secured against the asset — lower risk, lower interest. Junior debt sits behind it and is only repaid once the senior lenders are satisfied — higher risk, higher interest.

Splitting the borrowing this way lets one deal serve two different appetites. Conservative capital takes the senior tranche at a modest return; capital hunting yield takes the junior tranche and is paid for standing further back in the queue. In the Broadcom financing reported in August 2026, the reported structure pairs a senior-secured tranche of roughly $60–70 billion with about $30 billion of junior debt. Broadcom, Anthropic, Apollo and Blackstone all declined to comment on it.

Leasing versus buying, from the lab's side. Leasing converts an enormous upfront purchase into an operating cost spread across the life of the equipment. It preserves cash, avoids balance-sheet debt, and hands someone else the question of what the hardware is worth in five years.

That last point deserves its own name. Residual-value risk is the risk that the asset is worth less at the end than everyone assumed. Whoever owns the machine carries it. For an aircraft with a thirty-year life and a deep second-hand market, that risk is well understood. For AI accelerators, whose useful economic life is compressed by rapid generational turnover and whose resale market is young, it is much less well understood — and in this structure it sits with the SPV and its lenders, not with the lab.

Guarantees: where risk quietly comes back. Here is the part that most often surprises people. If the manufacturer guarantees part of the SPV's debt, it has promised to pay if the vehicle cannot.

A guarantee is a contingent obligation. It is not borrowing, it does not appear as debt, and the accounting treats it differently. But the economic exposure is real. The company has not transferred that portion of the risk; it has changed where the risk is recorded.

Consolidation, at a high level. Accounting rules do contemplate this. The variable interest entity framework asks, in substance, whether a company controls an entity and absorbs its economic risks and rewards — and if so, requires it to consolidate the entity's numbers into its own. The rules are designed to stop companies parking obligations in structures they effectively control.

They do not catch everything. Off-balance-sheet arrangements of this kind do not necessarily trigger consolidation, which is precisely why they are useful. Whether any specific structure must be consolidated depends on its terms, and is a question for the filings and the auditors, not for a reader of press reports.

Who ends up carrying what. Roughly: the AI lab carries the obligation to keep paying the lease. Senior lenders carry a secured claim, first in line. Junior lenders carry the risk that the asset underperforms. The SPV — and through it, its lenders — carries residual-value risk on hardware in a fast-moving market. And the manufacturer carries whatever it has guaranteed, plus the commercial risk that a financing channel supporting its own sales stops working.

Why this matters for AI specifically. Two consequences follow. First, the pace of the AI build-out becomes a function of credit-investor appetite rather than technology-company cash flow — a more cyclical master, and one that can withdraw quickly. Second, aggregate leverage in the sector becomes genuinely hard to measure, because the borrowing sits in vehicles that do not appear in the accounts of the companies most associated with the build-out.

What to watch, if you want to follow this. Whether guarantees are disclosed in quarterly filings and at what size. Whether credit markets price the guarantor as though the debt were its own — credit default swap spreads are the fastest available read. And whether assumed residual values for AI hardware hold up, because that assumption is doing a great deal of quiet work in every deal of this kind.