NEW YORK, 23 AUG 2026 — Broadcom, Apollo and Blackstone have built a financing platform called AI XPV whose first transaction raised US$35bn against artificial intelligence compute. The proceeds fund more than a gigawatt of infrastructure for Anthropic.
The mechanism is a special purpose vehicle that raises the debt, buys the accelerators — including Google tensor processing units that Broadcom co-designs — and leases them to Anthropic, whose lease payments service the loan. Broadcom provided a deficiency guarantee on the senior notes and agreed to support the residual value of the chips.
How the structure works
The vehicle owns the hardware; Anthropic rents it. That separation is the point, because it converts a company's future compute spending into an asset a lender can hold, and lease payments from a named counterparty are a more familiar credit exposure than a loan to a company that did not exist five years ago.
The two Broadcom guarantees are doing all the work
The financing works only because of two commitments from Broadcom.
The deficiency guarantee aligns Broadcom's balance sheet with the credit risk of the senior notes, which is what compressed the yield to roughly 5.75 per cent. Without it, debt secured on depreciating specialist hardware leased to a loss-making company would price very differently.
The residual value support is the more remarkable guarantee. If Anthropic defaults and the chips are sold for less than is owed, Broadcom covers the shortfall for senior noteholders — in effect, underwriting the second-hand value of its own hardware.
Read plainly, the risk has not been transferred to private credit so much as routed through it. Investors hold paper that is economically closer to Broadcom credit than to compute credit, priced accordingly, while the deal is described as chip-backed financing.
Residual value is the assumption nobody can check
Every part of this rests on what a used AI accelerator is worth in five years, and that is a genuinely unanswerable question today.
There is no established secondary market for data-centre accelerators at scale, no depreciation history across a full technology cycle, and a product cadence that has been compressing rather than lengthening. A generation of hardware that is competitive today may be uneconomic to operate well before it fails, because the constraint is performance per watt against newer parts rather than reliability.
Aircraft leasing, the closest analogue for this structure, works because a twenty-year-old airframe has a known value, a liquid market and a residual curve derived from decades of transactions. Accelerator leasing has none of those, and the residual assumption is therefore a judgement rather than a calculation.
Broadcom taking that risk is a meaningful statement of confidence in its own products. It also concentrates an unmodelled risk on one balance sheet. That a company with a strong credit rating has absorbed the risk doesn’t make the underlying uncertainty any smaller.
The circularity is real and worth naming precisely
A supplier guaranteeing the financing for purchases of its own products creates an obvious feedback loop. It is useful to be exact about what is and is not circular here.
What is genuinely circular: Broadcom's guarantee enables debt that buys chips Broadcom co-designs, generating revenue that supports the balance sheet backing the guarantee.
What is not: the cash is real, the lender is a third party, and Broadcom's exposure only crystallises if Anthropic defaults and the hardware is worth less than the outstanding balance. This is not revenue recognised on a sale to oneself.
Supplier credit support has moved from a marginal sales tool to a structural feature of how the sector finances itself. Vendor balance sheets are now load-bearing for the industry's debt.
Twenty gigawatts is the number that changes the question
One transaction is a financing. A platform intending more than twenty gigawatts through 2028, serving several model developers, is a market.
At that scale the structure stops being a clever arrangement for one borrower and becomes the mechanism by which a large share of AI infrastructure gets funded. The exposures then aggregate: multiple vehicles, overlapping counterparties, the same handful of suppliers providing credit support, and residual value assumptions that are correlated because they all depend on the same technology cycle.
The pattern to watch is correlated assumptions across many such vehicles. This is how a manageable risk in one deal can become unmanageable in aggregate. This is not a prediction of failure, but a reason to measure the aggregate exposure while it is still being assembled.
Why this reaches data centre projects in this region
Sovereign and commercial AI projects across ASEAN are financed conventionally: government budgets, corporate balance sheets, project finance against contracted revenue. Compute-collateralised structures are a different instrument, and their availability changes what is buildable.
The condition is a creditworthy lessee with contracted demand. A regional operator with a signed offtake agreement from a hyperscaler can plausibly access this market; one building speculative capacity in the hope of filling it cannot, and the gap between those two is what determines who gets funded.
The second implication is supplier concentration. If credit support from a handful of chip suppliers makes these structures financeable, those suppliers gain influence over which projects get built anywhere in the world, including here. That is a form of leverage that does not appear in any procurement negotiation.
One lessee is carrying the whole structure
Beneath the guarantees and the collateral sits a simpler exposure that the structure does not remove: the lease payments come from one company.
Anthropic is growing quickly and the compute is being bought to serve demand it already has. But a lease is a fixed obligation and revenue is not, and the arrangement assumes a growth path continues for as long as the debt is outstanding. That is a bet on one counterparty's trajectory dressed as a bet on hardware.
The guarantees change who absorbs the loss, not whether the scenario can occur. If demand disappoints, Broadcom pays and the chips are sold into a market that will be soft for the same reason the lessee struggled — which is the awkward feature of residual value support generally. The guarantee is least valuable precisely when it is most needed.
None of that makes the deal unsound. It does mean the risk being priced at 5.75 per cent is correlated in a way that a diversified lease portfolio would not be, and the platform's twenty-gigawatt ambition is what would eventually diversify it.
What remains unconfirmed
The full terms are not public. The tranche structure beyond the two senior tranches, the subordinated capital, the lease term and the residual value assumptions underpinning Broadcom's support are not disclosed in the material reviewed.
Open questions include what triggers the deficiency guarantee, how the chips would be valued on a default, whether Anthropic has a purchase option, and how the arrangement is treated on each party's balance sheet. Whether further AI XPV transactions have closed, and with which counterparties, is not stated.
What to watch for
Watch for a second transaction. A platform with one deal is a bespoke arrangement. A second deal with a different lessee would establish the template as repeatable and show that the residual assumptions can survive a fresh negotiation.
The second is whether other suppliers offer comparable guarantees. Broadcom acting alone is a company decision. Several suppliers doing it is an industry norm, and norms are harder to withdraw when conditions change.
The third is any disclosure of aggregate compute-backed debt. The individual deals are visible; the total is not, and the total is the number that matters.
