Broadcom's $100B AI Debt Bet Is Your Vendor Risk

On August 20, Bloomberg reported that Broadcom was in talks with lenders to raise more than $60 billion in debt for an AI chip financing package benefiting Anthropic and other frontier labs. By the next day, CNBC put the expected size upwards of $70 billion, raised through a Broadcom-backed special purpose vehicle. Add the junior tranche under discussion and the total lands near $100 billion.

I’ve written a lot on this site about AI vendor stability. OpenAI’s S-1. Anthropic’s S-1. Cloud commitments, price wars, gateway acquisitions. All of that looks at the vendor’s own balance sheet.

This is the layer underneath, and almost nobody buying AI is watching it. The chips your model provider runs on are increasingly not owned by your model provider. They’re owned by a financing vehicle, leased to the lab, and backstopped by the chipmaker.

Quick Verdict

Question	The Answer
What’s happening?	Broadcom is negotiating a debt package reported at $70B to $100B to fund custom AI chips and infrastructure for Anthropic and other labs.
Who reported it?	Bloomberg on August 20, then CNBC on August 21, 2026. Initial ask was reported above $60B.
How is it structured?	A senior secured tranche reported around $60B to $70B, plus a junior tranche of roughly $30B. Broadcom would guarantee part of the senior debt.
Who’s putting up money?	Blackstone and Apollo Global Management are in talks, alongside global banks.
Is this the first one?	No. Apollo led a $35B capital solution for the Broadcom AI XPV Platform on June 9, 2026.
Who owns the chips?	Not the AI lab. Investors finance the hardware and lease it to the lab.
What’s the capacity target?	More than 20 gigawatts of compute for frontier labs through 2028.
Does anything break today?	No. This is a multi-year structural risk, not an outage.
What actually changes for buyers?	Your vendor’s cost floor is now partly a debt service schedule set by credit markets.
The number to watch	Broadcom’s disclosed contingent lease exposure in its quarterly filings.

What is the Broadcom AI debt deal?

It’s a financing package, reported at $70 billion to $100 billion, raised through a Broadcom-backed vehicle so investors can buy custom AI chips and infrastructure and lease that hardware to AI labs including Anthropic. The labs get compute without buying it. Broadcom guarantees part of the debt. Credit markets fund the buildout.

That last sentence is the one worth sitting with. Credit markets fund the buildout.

The Structure Nobody Explains in the Headlines

Most coverage stops at the headline number. The mechanics are what matter to anyone with an AI line item.

Here’s how it works. A financing vehicle raises debt. The vehicle uses that money to buy Broadcom XPUs, networking gear, and the surrounding infrastructure. The AI lab signs a lease and pays over time. Broadcom provides credit support on the senior tranches, which makes the paper cheap enough for institutional investors to want it.

Anthropic doesn’t buy the chips. It rents them.

This isn’t a one-off. Apollo announced the $35 billion capital solution for Broadcom’s AI XPV Platform on June 9, with Blackstone and a syndicate of global banks. That vehicle targets more than 20 gigawatts of compute through 2028 and started with Anthropic’s expansion of over a gigawatt beginning mid-2026. The $70B-plus package now in talks is the same playbook at three times the scale.

Apollo partner Jamshid Ehsani described the logic plainly in the June announcement: AI compute is “rapidly emerging as one of the most compelling new asset classes in finance, characterized by contracted cash flows, mission-critical utility.”

Contracted cash flows. Those contracts are your vendor’s obligations, and your subscription is somewhere at the end of the chain that services them.

Why This Is Different From the IPO Stories

When Anthropic filed its S-1 and when OpenAI’s financials went public, the risk was legible. Public markets create quarterly pressure. Quarterly pressure moves pricing. You could read the filing, see the operating margin, and predict the direction of your renewal.

Lease obligations behave differently, in three ways that matter.

They’re contractual, not discretionary. A lab under margin pressure can cut research spend, delay a product, or slow hiring. It cannot skip a lease payment on the hardware running its inference. Debt service sits above almost everything else in the priority stack, which means it sets a floor under what your vendor can charge.

They’re long-dated. Twenty gigawatts through 2028 means obligations extending well past any contract you’re signing this year. The financing decisions being made in August 2026 will still be shaping unit economics when you’re negotiating your 2029 renewal.

They’re partly invisible. Credit backstops on customer leases don’t sit on Broadcom’s balance sheet as debt. Simply Wall St flagged this as a growing scrutiny point on August 16, noting that bond traders read these structures as hidden exposure. A separate Simply Wall St analysis put the vehicle’s potential senior debt at as much as $370 billion by mid-2029, with an estimated $29 billion in default risk tied to guaranteed lease obligations.

I want to be careful here, because $370 billion is a projection, not a reported balance. It’s the outer edge of what the structure could accumulate under current growth assumptions. But the direction is not in dispute, and the disclosure gap is real.

The Concentration Problem Runs Both Ways

Broadcom’s CEO has said the company has line of sight to more than $100 billion in AI chip revenue in 2027. SiliconANGLE reported that Anthropic could account for more than 40% of it.

Read that from both directions.

From Broadcom’s side, a single customer approaching half of a $100 billion revenue line is concentration risk of the kind that ends careers when it goes wrong. From Anthropic’s side, the primary supplier of its custom silicon is also a guarantor on the debt that funds its compute. Those two companies are now financially entangled in a way that a normal supplier relationship never produces.

I flagged something similar when Google committed $40 billion to Anthropic days after Amazon’s commitment: the vendor decision and the infrastructure decision were collapsing into one. This goes further. Now the vendor decision, the infrastructure decision, and the credit decision are the same decision, and only one of those three shows up in your procurement review.

What Would Actually Have to Go Wrong

I’m not predicting a collapse. Anti-hype cuts both directions, and the doom version of this story is as lazy as the boom version.

For this structure to hurt an AI buyer, a specific chain has to break:

Failure point	What it looks like	How you’d notice
Demand softens	Compute utilization drops below lease-coverage assumptions	Labs quietly stop announcing new capacity
Credit repricing	Rates rise or the junior tranche struggles to place	Deal size shrinks or terms leak as “revised”
Backstop triggers	A lessee misses payments, Broadcom’s guarantee activates	Contingent liability disclosure in a 10-Q
Cost passes through	Debt service gets recovered in per-token or per-seat pricing	Your renewal quote, with a vague explanation

Only the last row touches you directly, and it’s also the most likely one. The realistic bad outcome here isn’t a lab failing. It’s a lab that can’t discount as aggressively as it used to, because a fixed obligation sits underneath its cost structure and doesn’t care about your negotiation.

Which is the opposite of what buyers have been trained to expect. The last two years conditioned everyone to assume inference prices only fall. OpenAI’s price cuts reinforced it. Multi-year deflation is a reasonable base case on the compute-efficiency curve. It’s a much shakier one when the efficiency gains have to clear a debt service schedule first.

How do you assess AI vendor risk from infrastructure debt?

Six steps. A procurement lead and a technical owner can work through this in about two hours per vendor, and it costs nothing but the time.

1. Ask your vendor who owns the compute they run your workloads on. Owned, leased, cloud-committed, or financed through a third-party vehicle. The answer is rarely in the sales deck and is usually available if you ask the right person.
2. Search your vendor’s largest suppliers for lease-backstop language in quarterly filings. For public suppliers like Broadcom, contingent liabilities on customer leases are a disclosed line. Read it once a quarter and watch the trend, not the absolute number.
3. Model your renewal at flat pricing, not declining pricing. Most AI budgets assume per-token costs keep dropping. Build a second scenario where they hold steady for 24 months and see what breaks.
4. Get price-protection language into contracts longer than 12 months. A cap on increases costs you nothing to ask for and is the single highest-leverage clause available before an IPO or a financing event resets the terms.
5. Keep a second model provider integrated and tested, not just listed. Optionality is only real if someone has run production traffic through the alternate path in the last 90 days.
6. Put one calendar reminder per quarter on your top vendor’s funding news. Twenty minutes, one owner by name. Financing announcements precede pricing changes by quarters, which is exactly enough lead time to be useful.

Steps 1 through 3 are this month. Steps 4 through 6 are ongoing.

My Read

Three things I think are true.

This is the most important AI story of the month and it will get the least attention from buyers. Debt structures are boring. They don’t demo well, there’s no product page, and the news cycle moves on in 48 hours. But the financing layer determines the cost floor for every AI service sold on top of it, and cost floors outlive product cycles. In my experience, the constraints that eventually reshape a market are almost always the ones that looked like back-office plumbing when they were installed.

“AI compute as an asset class” is a genuinely new financial structure, and new structures get their stress test late. Apollo’s framing is honest and probably correct: contracted cash flows on mission-critical infrastructure are attractive paper. The question isn’t whether the logic is sound. It’s whether the lease-coverage assumptions hold if compute demand grows more slowly than the 20-gigawatt plan implies. Nobody knows the answer, including the people underwriting it, because there’s no prior cycle to check against.

The entanglement is the story, not the headline number. A chipmaker guaranteeing its customer’s lease obligations while that customer becomes 40% of its revenue is a circular arrangement. Not fraudulent, not hidden, just circular. Value moves in a loop between a small number of parties, and loops are efficient right up until one node slows down. If you want a single question to carry into 2027, it’s this: how many of the companies in your AI stack are financially dependent on each other?

Here’s what I’d tell a business owner reading this and concluding it’s above their pay grade. It isn’t. You don’t need to understand tranche seniority. You need to understand one thing: the price you pay for AI is now connected to a debt market, and debt markets don’t care about your budget cycle. That’s it. Everything else is detail for someone else to worry about.

The Bottom Line

Broadcom is arranging up to $100 billion in debt so that investors can buy AI chips and lease them to Anthropic and other labs, building on the $35 billion AI XPV vehicle from June. Blackstone and Apollo are at the table. Broadcom is guaranteeing part of the paper.

Nothing about this breaks your AI stack this quarter. What it does is add a fixed, long-dated, partly off-balance-sheet obligation underneath the vendors you already depend on. The AI vendor risk conversation has been about IPOs and acquisitions, and it’s been looking at the wrong layer. The chips came first, somebody borrowed to pay for them, and that loan is now part of your cost structure whether or not it appears anywhere in your contract.

Your Next Step: This week, pull up your two largest AI vendor contracts and check for a price-increase cap. If there isn’t one, put it on the list for your next renewal conversation and ask for a 12-month notice requirement on pricing changes while you’re at it. Then build the flat-pricing budget scenario. Thirty minutes in a spreadsheet tells you whether a market that stops getting cheaper is an inconvenience or a problem, and it’s much better to learn that now than in a renewal meeting.

---

Related Reading:

• Anthropic Goes Public. Lock In Your Contracts Now.
• OpenAI Filed. Their Financials Should Worry You.
• Google Bets $40B on Claude. Here’s Your Move.
• The $60B AMD Deal That Could Lower Your AI Costs
• The AI Price War Starts Now. Don’t Lock In Wrong.
• OpenRouter’s $7B Stripe Deal Kills the Lock-In Pitch