Broadcom's exposure to the AI buildout could reach around $370B by the end of the decade, on Bank of America's estimate of the maximum it would owe under the residual value guarantees sitting behind AI equipment debt. The bank downgraded the chipmaker over it.
The scale of what those guarantees support became clearer this week. Anthropic's IPO prospectus, reported on 29 September, showed $161.2B of largely non-cancellable Broadcom equipment leases inside a $518B decade of computing commitments, which is the demand side of a financing structure Broadcom has spent the summer assembling.
In August, Broadcom was reported to be seeking more than $60B in debt, with figures of up to $100B discussed, through a special purpose vehicle backed by Apollo Global Management and Blackstone. The structure splits into a junior tranche of about $30B and a senior tranche of $60B to $70B, with Broadcom guaranteeing part of the senior portion to secure an investment-grade rating.
The money buys AI computing hardware that is then leased to companies including Anthropic, which keeps the capital spending off the lab's balance sheet and moves the contingent liability onto Broadcom's.
The Supplier Has Become The Credit
That shift is the part worth understanding. In a normal equipment sale, the buyer's creditworthiness determines whether lenders will fund the purchase. Here, the vendor's guarantee is what makes the debt sellable, because institutional investors are lending against Broadcom's balance sheet rather than against a young AI company's revenue projections.
Broadcom therefore takes on exposure to an AI revenue cycle it does not control, which is what the $370B estimate measures and why the downgrade followed.
Nvidia has done something comparable, with a reported $105B guarantee behind OpenAI's Ohio campus. The chipmakers are no longer only selling into the AI boom; they are underwriting it.
Investors Are Starting To Charge More
The bond market has noticed. Amazon's $25B sale in July came with 18 to 21 basis points of extra yield on its longest-dated bonds and an order book covered 2.5 times, down from 3.2 times in March, which Bank of America read as investors pushing back.
The supply explains the resistance. Amazon, Alphabet, Meta and Oracle sold roughly $194B of bonds through early July, up 79% on all of 2025, and Goldman Sachs expects those four plus Microsoft to issue $250B this year and $400B next. A combined $75B offering from Nvidia, SpaceX and Amazon struggled to clear and traded down after pricing, according to Wall Street Journal analysis, and hyperscaler spreads have widened across maturities.
Apollo's chief economist framed the question directly: who becomes the marginal buyer of investment-grade paper when hyperscaler supply keeps climbing. BlackRock strategists have called the environment a rare capital war.
Amazon's Own Numbers Show The Strain
Amazon spent $43.2B on capital projects in the first quarter of 2026, and its trailing twelve-month free cash flow fell to $1.2B from $25.9B a year earlier, even as operating cash flow rose 30% to $148.5B. The company has issued about $92B of debt this year, more than Google, Meta or Oracle individually.
Those are the figures of a profitable company converting cash generation into fixed assets at a pace that leaves little margin. Amazon said it does not expect to issue more debt in 2026.
Why The Structures Keep Getting More Complex
The pattern across these deals is risk moving to whoever can carry it most cheaply. Oracle borrows and builds. Broadcom guarantees debt raised by a vehicle it does not own. Anthropic signs take-or-pay contracts rather than buying hardware. Australia's Sharon AI mortgages its GPUs through a special purpose vehicle at 9.95%.
Each structure answers the same question differently: how do you fund hardware that costs billions, depreciates fast and depends on revenue that does not exist yet. None of them removes the risk, and each one puts it somewhere less obvious.
That opacity is what should concern buyers and investors. When exposure sits in guarantees, lease commitments and off-balance-sheet vehicles rather than in plain borrowing, working out who absorbs a downturn takes more effort than reading a balance sheet.
What It Means For Companies Buying Compute
For businesses contracting AI capacity, the financing structure behind a provider is worth knowing. A provider funded by vendor guarantees and take-or-pay leases has strong incentives to fill capacity, which helps pricing now, and limited flexibility if demand softens.
Buyers signing multi-year compute agreements should ask who owns the hardware their workloads run on, what happens to their capacity if the financing vehicle behind it runs into trouble, and whether their contract survives a change of ownership. Those questions have become relevant in a way they were not two years ago.
The same pressure is visible across the stack, from Toshiba doubling hard drive production to the power constraints pushing governments such as Thailand to screen data centre applications before approving grid connections.
The Test Is Whether Revenue Arrives On Schedule
Nothing in these arrangements fails if AI revenue grows as projected. Anthropic reports a run rate in the tens of billions, the hyperscalers are filling capacity as fast as they build it, and lenders are still buying the paper, if at wider spreads.
The vulnerability is timing. Debt has fixed repayment dates, guarantees crystallise at awkward moments, and refinancing walls arrive whether or not the revenue has. Broadcom's potential exposure of roughly $370B by 2030 is the clearest statement of what the industry is betting, and the bond market's widening spreads are the first sign that someone is checking the arithmetic.
