Broadcom's AI chip financing has moved from an equity-market story to a credit-market problem. You don't have to think AI demand is fake to worry about guarantees that only appear when something goes wrong.
The AI buildout is now big enough that bond traders are asking a better question than stock investors. Not whether Nvidia, Broadcom, Microsoft or Meta can keep selling the future, but who gets stuck with the bill if the hardware behind that future is worth less than the financing assumes.
That is why the telecom comparison keeps coming back. The technology itself did not fail in 2001. Fiber mattered. Internet traffic kept growing. But the companies that borrowed against perfect demand curves still broke when prices fell, funding tightened and too much capacity arrived at once. A 2005 SEC filing tied to the old AT&T and SBC deal said telecom capital spending surged to more than $120 billion in 2000, more than three times the 1996 level. Federal Communications Commission Chairman Michael Powell told the Senate in July 2002 that the industry owed about $1 trillion, according to remarks later cited by The American Prospect.
Credit markets remember that kind of number.
The clearest current signal is Broadcom. Bloomberg reported on August 24 that yields on Broadcom's 5.15% bonds due in 2031 rose about 14 basis points so far in August, while the cost of five-year credit default swaps jumped 28 basis points, more than the moves in Oracle and SpaceX. That isn't a normal semiconductor story. It is the market pricing a chipmaker that is starting to look, in part, like a financial backer.
Amazon scraps AI leaderboard to stop workers boosting usage scores , Senior executive tells staff 'don't use AI just for - why Amazon shut down internal AI leaderboard - how to prevent AI metrics gaming at work
Broadcom struck an agreement earlier this year to backstop most of a $35 billion debt package in which investors including Apollo Global Management and Blackstone financed custom AI chips leased to Anthropic, Bloomberg reported. The company is also in talks over a financing package of more than $60 billion tied to AI chips expected to benefit Anthropic and other companies. If you own the debt, the sales growth is only half the story. The backstop is the part you have to read twice.
The Guarantee Is The Story #
Residual value guarantees sound harmless until demand slows. A special-purpose vehicle borrows money to buy chips, then leases that hardware to a customer. If the customer walks away and the chips cannot be sold or leased again for enough money, the guarantor covers part of the gap. That support can help the debt win an investment-grade rating, because the lender is leaning on Broadcom or Nvidia rather than only on the customer using the machines.
Frankly, that is a neat structure until it isn't.
Bloomberg reported on August 15 that bond traders were already focused on roughly $70 billion of these AI backstops, obligations that do not appear as ordinary debt on the guarantors' balance sheets. Meta described its own residual value guarantees in filings by saying the payments were not probable, so no liability had been recorded. That may be correct accounting. It still leaves investors with a judgment call that no spreadsheet handles cleanly: what looks remote during a shortage can look very real when used hardware prices fall.
Tony Trzcinka, an investment-grade portfolio manager at Impax Asset Management, told Bloomberg the rise in Broadcom's CDS looked specific to Broadcom's balance sheet rather than broad fear about AI spending. JPMorgan strategist Tarek Hamid put a sharper label on the same concern, writing that leases, purchase commitments, residual value guarantees and other backstops were adding a layer of phantom debt underneath the AI ecosystem that could stretch into the trillions.
That phrase earns its place.
The Debt Is Moving Around #
The Wall Street Journal recently put the broader issue in harder numbers. Its analysis of securities filings found that nine major technology companies, including Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, AMD and SpaceX, had about $3 trillion of AI-related obligations sitting outside conventional balance-sheet debt. The Journal's count included about $1.2 trillion of leases that had been signed but had not yet started, along with about $1.9 trillion of purchase commitments for chips, power and other infrastructure.
Hyperscaler Debt Flood Brings Derivatives Bonanza Hyperscaler Debt Flood Brings Derivatives Bonanza - hyperscaler AI infrastructure debt derivatives trading - how Wall Street profits from tech company borrowing
None of this proves AI demand is fake. Nvidia's data center revenue is real, and the shortage of compute is real enough that buyers are signing contracts years ahead of delivery. But debt does not need a technology to fail before it causes damage. It only needs the cash flows to arrive later, the resale values to fall faster, or the customer growth to come in below the financing model.
That is the useful lesson from telecom. The internet won, but many of the securities built around the first infrastructure rush did not. AI may follow the same split path: useful technology on one side, stretched credit structures on the other. If you are reading the market from stock charts alone, you will miss where the worry is showing up first.
Start with the bond desks. They are already asking who owns the risk when the guarantee stops being theoretical.
Also read: Apple's New Mac Studio Bets Builders Will Run AI Locally Instead of Renting Cloud GPUs, OpenAI Says Its First Chip Jalapeño Beats Nvidia's GB300 on Inference, ARIA Bans AI-Generated Songs From Australia's Official Music Charts