Broadcom Is Raising $70 Billion Against Its Own Balance Sheet. The Market Already Priced What That Guarantee Is Worth: 275 Basis Points.

Bloomberg reported on Thursday, August 20, 2026 that Broadcom is in talks with lenders to raise more than $60 billion in debt for an AI chip financing deal benefiting Anthropic and other labs. By Friday morning CNBC had the number higher: $70 billion to $80 billion, with a senior tranche around $45 billion and a junior tranche around $35 billion. Apollo and Blackstone are again in the room.

The first deal on this platform closed in June at $35 billion. Ten weeks later the second one is roughly double. That is the headline everybody ran.

I want to argue the headline is the least interesting part. The June deal left behind something far more useful than a dollar figure: a tranche stack where one slice carried a Broadcom guarantee and one slice did not, and both cleared the market on the same day. Two prices, same collateral, same lessee, same five-year term. Subtract them and you get the only public quote anyone has on what the credit markets think a frontier AI lab is worth as a borrower on its own name.

The answer is 275 basis points.

What the AI XPV Platform Actually Is

Broadcom announced the AI XPV platform on June 9, 2026, with Apollo Global Management and Blackstone's Credit and Insurance business as anchor investors. The stated design goal is more than 20 gigawatts of compute capacity through 2028, built on Broadcom XPUs and networking, customized for frontier labs including Anthropic and OpenAI.

The mechanics are ordinary structured finance dressed in AI clothing. A special purpose vehicle borrows money from institutional investors, takes an equity slug, buys the chips, and leases them to Anthropic on a five-year term. The SPV owns the silicon. Anthropic makes lease payments. Investors get paid from those payments. Anthropic never puts $35 billion of accelerators on its own balance sheet, and Broadcom books a chip sale to a counterparty funded by somebody else's capital.

Hock Tan called it synchronizing the world's most sophisticated capital with Broadcom's technological roadmap. That is a fair description of what it does. It is also a fair description of what a vendor does when its customers cannot fund the purchase order.

The initial tranche was roughly $35 billion, led by Apollo with Blackstone, covering about 1GW that Anthropic deploys at Fluidstack sites starting mid-2026. Here is how it split.

Thirty billion of the $34.5 billion in notes carries a Broadcom residual value guarantee. The remaining $4.5 billion does not. Under the residual value support agreement, if Anthropic stops making lease payments and the proceeds from selling the chips fall short, Broadcom covers the full outstanding balance for A1 and A2 holders. Class B holders get whatever is left.

The 275 Basis Points Is the Only Honest Number in the Deal

A2 cleared at 5.75 percent. Class B cleared at 8.5 percent. Both sit against the same chips, the same lessee, the same lease, the same tenor. The only material difference between them is whose name is on the backstop.

I want to be careful here, because subordination alone carries a spread even with identical credit behind it. Some slice of that 275 basis points is payment priority rather than guarantee value. But the guarantee is not partial in the usual sense. It does not absorb a first loss and then run out. It covers the full outstanding balance on A1 and A2 if the chips do not cover it. Once you strip out a normal subordination premium, most of what remains is the market pricing the difference between lending to Broadcom and lending to Anthropic.

And 8.5 percent is a real number with a real meaning. That is high yield. That is where the market prices a chip-backed, asset-secured, five-year obligation from a company that is simultaneously reporting a $65 billion annualized run rate and reportedly seeking a public valuation of $2 trillion or more.

Both things are true at once and they are not in tension, which is the part I find genuinely interesting. Equity markets price the upside distribution. Credit markets price the downside one. An investor can rationally believe Anthropic is worth $2 trillion in the good case and still want 8.5 percent to lend against a five-year lease, because the equity case runs on growth and the credit case runs on whether the lease payments show up every month for sixty months at a company that did not exist at this scale eighteen months ago.

The credit markets have been quiet on frontier labs because frontier labs have funded themselves with equity. This is the first clean tick.

A Residual Value Guarantee on a Custom ASIC Is Not Really About Residual Value

This is the part I think is underread, and it is a hardware argument rather than a finance one.

Residual value guarantees are a normal instrument. Aircraft lessors use them. Auto lessors use them. They work because the underlying asset has a deep, liquid secondary market with observable prices. If a lessee defaults on a fleet of narrowbodies, there is a global bid for narrowbodies.

An XPU is not a narrowbody. It is not even a GPU. Nvidia hardware has a genuine resale market: neoclouds buy it, research labs buy it, crypto refugees buy it, and there is a price you can look up. A Broadcom XPU is a custom ASIC co-designed for one customer's stack, wired into one customer's network topology, and paired with one customer's compiler and kernel work. If Anthropic stops paying, the SPV is holding a gigawatt of accelerators configured for Anthropic.

Who is the second buyer? Possibly Google, since the TPU lineage is shared and Broadcom has built Google's tensor chips for over a decade. Possibly another XPV tenant. But that is a list of two or three names, and every one of them is a strategic buyer who knows the seller is distressed and knows the hardware fits almost nobody else. That is not a market. That is a negotiation.

Which means the residual value guarantee is doing something different from what its name suggests. In a deep secondary market, an RVG covers a gap between expected and realized resale value. Here, the realistic recovery in a default scenario is low enough that the guarantee is functionally a full credit guarantee with a collateral costume on. Broadcom is not insuring a price. Broadcom is insuring a customer.

Bank of America Noticed on August 11

Ten days before this week's reporting, BofA cut its credit view on Broadcom from Overweight to Marketweight, citing XPV specifically. Analyst Tom Curcuruto noted that Broadcom's bond spreads had widened roughly 20 to 30 basis points against other A-rated semiconductor issuers since early June, which is to say since the XPV launch.

The modeled exposure is the number that should stop people. Scaled across the full 20GW ambition with roughly 2GW deals landing quarterly, maximum residual value guarantee exposure reaches about $370 billion by mid-2029. BofA put maximum loss exposure at roughly $42 billion under a total default assumption and about $10.5 billion at a 25 percent default rate. Separately, Broadcom has extended around $29 billion in backstop guarantees on XPV lease payments.

Broadcom's entire AI-related order backlog was reported at roughly $73 billion. The contingent guarantee book is on a path to be five times that.

Note the shape change between tranche one and tranche two. The first deal was 87 percent guaranteed senior paper and 13 percent unguaranteed junior. The second is reportedly around $45 billion senior against $35 billion junior, which is closer to 56 and 44. If those proportions hold, a much larger share of the second deal is being sold to investors who are taking lab credit risk directly rather than Broadcom credit risk. That could mean the market got more comfortable with Anthropic. It could also mean Broadcom is rationing how much of its own balance sheet it is willing to keep pledging. Those read very differently and nobody outside the deal knows which one it is yet.

The Case That This Is Fine, Made Properly

The obvious comparison is vendor financing in the telecom buildout, when equipment makers lent customers the money to buy equipment and booked the resulting revenue as real. It ended badly and the phrase still carries a smell. I think the comparison is worth making and I also think it is weaker than it looks, in four specific ways.

One. Broadcom is not lending cash. It is providing a contingent guarantee. The money comes from Apollo, Blackstone, and institutional credit investors who are underwriting the deal on their own analysis. No receivable from a shaky customer sits on Broadcom's books inflating revenue quality.

Two. The exposure is contingent and collateralized rather than direct. Broadcom pays only if Anthropic defaults and the chips fail to cover the balance. Both conditions have to fire. My argument above is that the second condition is closer to automatic than the structure implies, but it is still a second condition.

Three. The demand is not speculative in the way dark fiber was. Anthropic is putting these chips against paying inference workloads today, at a run rate the company says passed $65 billion in July, with reported positive adjusted operating income in the second quarter. Telecom built capacity for traffic that had not arrived. Anthropic is buying capacity for traffic it is currently rate limiting.

Four. Somebody has to solve this. A five-year-old company cannot put $71 billion of compute commitments on a balance sheet that has never issued a bond. Either the compute does not get built, or the frontier labs get absorbed by hyperscalers who can fund it from operating cash flow, or a structure like XPV bridges the gap. The third option is the only one that preserves independent labs. That is a real argument and I do not want to wave it away.

What survives all four counterpoints is narrower but I think it holds: the demand signal for Broadcom's chips is now partially manufactured by Broadcom's own credit. Every rack sold through XPV grows the guarantee book. Revenue and contingent liability rise together, in lockstep, by construction. That is not fraud and it is not even unusual. It is just a feedback loop, and feedback loops are fine right up until the input changes sign.

This Is Not Happening in Isolation

XPV is the most structurally interesting deal in the category but it is far from alone. The financing of AI compute moved from equity to debt over about eighteen months, and the totals have gotten large enough that the shape matters.

The number in that table that should get more attention than it does: the five largest US hyperscalers ended 2025 with roughly $969 billion in total undiscounted future data center lease commitments, of which about $662 billion had not yet commenced and therefore sat entirely off the reported balance sheet. That off-sheet figure equals about 113 percent of those same five companies' combined adjusted on-balance-sheet debt.

So the pattern is consistent from the largest balance sheets in the world down to a five-year-old lab: the obligation is real, the structure keeps it off the primary financial statements, and the disclosure lives in a footnote or a private credit agreement that nobody gets to read.

Our Take

The interesting thing about the XPV structure is not whether it blows up. I do not think anyone can honestly forecast that, and pieces that claim to are selling something. The interesting thing is that it has produced a price where there was no price.

For three years, every argument about whether the AI buildout is rational has been conducted in equity terms, which means it has been conducted in terms of upside scenarios and terminal multiples and vibes. Credit markets do the opposite job. They ask what happens in the bad case and demand to be paid for it. The June tranche stack is the first time the bad case got quoted out loud on a frontier lab, and it printed at 8.5 percent against 5.75 percent for the same asset with a large chip vendor standing behind it.

That is not a scandal. It is data, and it is better data than anything the equity side has produced. It says the credit market believes these chips will be paid for, and also that it wants roughly 275 basis points of compensation for the possibility that they are not, and that Broadcom is currently willing to absorb that risk in exchange for demand it would otherwise have to wait for.

The thing worth watching is not default. It is whether Broadcom keeps writing the guarantee at the same ratio as the deals get bigger. A vendor that backstops 87 percent of a $35 billion deal is confident. A vendor that backstops 56 percent of an $80 billion deal is either being disciplined or being told no. Same behavior, opposite meanings, and the difference is everything.

Three things I am watching:

One. The guaranteed share of tranche two when it prices. If the Broadcom-backed portion falls as a percentage of total, the market is taking more lab credit risk directly, and the unguaranteed coupon becomes the number to track. If it clears meaningfully inside 8.5 percent, Anthropic's standalone credit improved in ten weeks. If it clears wider, it did not.

Two. Whether Anthropic's public S-1 discloses the XPV lease obligations as a quantified commitment schedule or as narrative risk language. A company reportedly carrying around $71 billion in compute commitments and seeking a $2 trillion valuation will be judged partly on how legible it makes that number. Reporting is already pointing at AI backlash showing up as a named risk factor in the filing, which suggests the disclosure posture is at least deliberate.

Three. Whether any second XPU customer signs a comparable structure. OpenAI is named in the platform's stated scope. A second lab borrowing against the same vendor's guarantee would turn a bilateral arrangement into a market, and would also concentrate a great deal of the industry's downside onto one semiconductor company's credit rating.

Nvidia sells chips. Broadcom is starting to sell chips and underwrite the buyer. Those are different businesses with different risk profiles, and only one of them is priced into a semiconductor multiple.