Broadcom Is Trying to Raise $60 Billion in Debt to Build AI Chips

Broadcom is in talks with lenders to raise more than $60 billion in debt for an AI chip deal that would benefit Anthropic and others, reported on 20 August. The senior secured tranche is being discussed at roughly $60 to $70 billion, with a junior tranche around $30 billion stacked on top. Total raise could approach $100 billion.

For scale, that is one company arranging more borrowing for chips than most countries spend on defence in a year.

What the Broadcom AI chip debt deal covers

Blackstone and Apollo are in talks to participate, following a partnership the three struck in June. It builds on a $35 billion deal that kicked off the same group's AI infrastructure financing push earlier in the year. The structure is the interesting part: senior secured paper first, then a junior tranche that carries more risk and prices accordingly.

Neither the size nor the participants are unusual for infrastructure finance. What is unusual is the asset. Toll roads and pipelines have thirty year lives and predictable demand. AI accelerators depreciate on a timeline nobody has agreed on yet, in a market where the performance leader changes roughly every eighteen months.

Why financing compute on credit changes the risk

Up to now, most of the AI build-out has been funded from the cash flow of extremely profitable companies. Hyperscalers wrote cheques out of operating income. If demand disappointed, they took a write-down and moved on.

Debt does not work like that. If demand softens, the chips do not simply sit idle in a rack while everyone shrugs. Somebody still owes the interest, on schedule, regardless of utilisation. That converts a capex bet into a fixed obligation, and it means the people to watch in AI capex are increasingly the lenders rather than the buyers.

It also ties the chip supply chain more tightly to credit conditions. Memory is already the constrained input, with HBM pricing climbing through the year, as covered in our explainer on HBM versus GDDR versus DDR and the newer High Bandwidth Flash spec.

What to watch next

Two things. First, where the junior tranche prices, because that number is the market's honest opinion on how risky AI hardware collateral actually is. Second, whether other chip vendors copy the structure. Vendor-arranged financing that routes capacity to a specific customer starts to look less like a supply agreement and more like the kind of entangled relationship we picked apart in the OpenAI and Microsoft piece.

None of this is a prediction that the deal goes badly. It is a note that the funding model changed, and funding models tend to matter most in the year nobody is looking at them.