Anthropic's $36B chip deal: read the Broadcom backstop

Apollo is raising $36 billion in private credit to buy Google TPUs for Anthropic. Broadcom is guaranteeing the resale value of its own chips.

The Editors · 9 min read · Jun 3, 2026

Apollo Global Management and Blackstone are arranging $36 billion in private credit to buy Google TPU chips for Anthropic to lease. The structure is standard for AI infrastructure now: a special-purpose vehicle owns the hardware, Anthropic pays rent, the lenders collect. What is unusual is the backstop. Broadcom, the company that designs the chips Google sells to the SPV, is guaranteeing the resale value of those chips against the senior debt. If Anthropic stops paying and the used hardware does not fetch enough on resale, Broadcom covers the gap. A chipmaker is underwriting demand for its own chips.

That rhymes with the vendor financing that helped inflate the telecom bubble in 1999. The comparison is inexact, and the people structuring this deal know their craft. But the Broadcom backstop is the load-bearing piece of the senior debt, and "the chip designer is insuring the resale value of its own chips" is a sentence worth saying out loud before assuming this is routine plumbing.

What the deal is

Apollo is leading the syndication. Blackstone is co-arranging. The debt sits in a special-purpose vehicle that will purchase Google's tensor processing units and lease them to Anthropic for use in data centers in New York, Texas, Louisiana, and Indiana. Bloomberg and PitchBook both confirm the structure and the parties.

The debt is tranched in three layers:

A1 super-senior notes: ~$6 billion. First in line for repayment.
A2 first-lien senior notes: ~$25 billion. Price talk on all-in yield is 5.50%-5.75% per PitchBook.
B second-lien notes: ~$4.5 billion. Fixed-rate coupon at 8.5%.

For context, BBB-rated U.S. corporate debt is currently trading around 5.3%. The A2 senior tranche is being priced in line with investment-grade corporate paper, even though the underlying asset is depreciating silicon. That tight pricing is only possible because of the Broadcom guarantee.

The Broadcom backstop

Broadcom is providing a "residual value support agreement" covering the senior $31 billion of the deal. Per Reuters and Yahoo Finance: if Anthropic stops paying its lease, the SPV sells the used TPUs on the secondary market. If the proceeds do not cover the A1 and A2 noteholders, Broadcom pays the shortfall.

Broadcom does not own the chips. Broadcom designed them and helped Google produce them. By backstopping the resale value, Broadcom is taking on the credit risk of its own customer's customer.

This is structurally different from how Meta financed its Hyperion data center last year. In that deal, Meta itself provided the residual value guarantee on the $27 billion of debt the SPV issued, as the Bank for International Settlements documents. Meta was the user. Meta carried the risk.

In Anthropic's deal, the user (Anthropic) does not carry the residual risk. The supplier of the supplier (Broadcom) does. That is the move worth paying attention to.

Why the structure exists

Anthropic confidentially filed an IPO with the SEC on June 1, 2026 (NPR). Booking $36 billion of corporate debt against a company at a $30 billion annualized run rate would change how the equity prices and complicate the registration.

The SPV solves that. The debt sits outside Anthropic. The lease payments hit Anthropic's income statement as operating expense, not interest expense. Investors evaluating the IPO see a cleaner page.

This is legal and it is becoming the default. The Bank for International Settlements tracks at least $120 billion of AI infrastructure debt now sitting in SPVs across Meta, xAI, Oracle, and CoreWeave. The pattern is consistent: the hyperscaler or AI lab leases the asset, the SPV holds the debt, private credit funds the SPV.

The reason it works for credit funds: stable lease income from a creditworthy tenant, secured against hard assets, backstopped by a guarantor. The reason it works for the labs: capex moves to opex.

What is different here

Three things distinguish the Anthropic deal from the prior pattern.

Scale relative to revenue. Meta's $30 billion Hyperion debt sits behind a parent company with around $165 billion in 2025 revenue. Anthropic's $36 billion SPV sits behind a company at a $30 billion run rate as of Q1 2026 (Anthropic). The run rate is growing fast. The coverage ratio is still tighter than the Meta comparison.

Where the revenue goes. Anthropic separately committed to spend approximately $200 billion over five years on Google Cloud per CNBC, roughly $40 billion per year. The compute Anthropic is paying for under that contract overlaps with the compute the SPV is buying. Plain version: Anthropic earns revenue, Anthropic pays Google for cloud, the SPV pays for the TPUs Google sources from Broadcom, and Broadcom guarantees the resale of those TPUs. Four entities, one chain of dependency.

The guarantor is the chip designer. Broadcom's revenue depends on shipping more AI chips. Its solvency depends on a continued AI buildout. Its guarantee is worth what its balance sheet is worth at the moment of a call. If AI capex contracts, Broadcom's revenue contracts and the guarantee value contracts at the same time. The protection is most likely to be called precisely when it is hardest to honor.

Reflexive credit is the term that fits. The protection sold to noteholders is correlated with the risk it is supposed to insulate them from.

The 1999 comparison, honestly

In 1999 and 2000, telecom equipment makers Lucent and Nortel financed their customers' purchases of their own gear. The argument was the same one being made now: the equipment was sound, the customers were creditworthy enough with vendor help, the market was growing. When the telecom buildout stalled, the receivables turned bad, the customers collapsed, and the equipment makers wrote down billions. Lucent's share price fell from $84 to under $1 over two years.

What is similar: a hardware supplier taking on credit risk linked to sales of its own product, justified by the assumption that end-customer demand will compound.

What is different:

Anthropic's revenue is real customer usage. Lucent was lending money to its own customers so they could buy Lucent gear.
Broadcom is providing a residual value guarantee on hardware. The structure resembles aircraft leasing more than telecom vendor paper.
The senior debt rating reflects both the asset value and the guarantor's balance sheet. If you trust both, the math holds.

The honest read: the structure is more sophisticated than 1999. The fundamental dependency, supplier-funded demand for the supplier's own product, is similar enough to deserve a closer look than the senior tranche pricing suggests.

What to watch

The TPU secondary market is thin. Most TPUs have not been sold used because Google has historically operated them in-house. Pricing the residual value is an educated guess from a small data set.

Broadcom's gross margin trajectory is the single best leading indicator. If AI revenue stays strong, the guarantee is comfortable. If AI revenue softens for a quarter, the guarantee value moves before the chips do.

Anthropic's cost of revenue after compute is the third line to watch. The $200 billion Google commitment over five years averages around $40 billion per year. Against a $30 billion run rate, the gap is funded by equity and debt, not by operating cash flow. It is a normal growth-stage profile, and it is also why the SPV has to exist.

What this means for the rest of the market

The SPV pattern is going to keep growing. CoreWeave already runs more than $11 billion in similar structures. xAI raised private credit against its Memphis facility on the same template. Oracle is reportedly working on a similar vehicle for its share of the Stargate commitment.

The market is pricing this debt at investment grade. That implies the credit funds writing the checks believe one of two things:

The AI buildout is structurally durable and the chip residual values will hold.
The guarantors (the hyperscalers and chipmakers backstopping these deals) are good for the money even if the residual values do not.

If you are a retail investor with no exposure to private credit, this still matters. The same insurance companies and pension funds that buy investment-grade corporate bonds buy senior SPV paper. The exposure exists; it does not sit on a public balance sheet you can look up in five seconds.

What to do with this

If you hold Broadcom stock, the residual value support is a contingent liability worth tracking through the next several 10-Q filings.

If you are watching the Anthropic IPO, the SPV is the reason the registration statement will look cleaner than a normal company carrying $36 billion in debt. That cleanliness has a price; it is paid by the noteholders and the guarantor, not by the equity buyer.

If you are an operator or builder watching AI capex, the signal is that 2026 is the year private credit started financing the AI buildout at a scale previously the domain of public debt and equity. That changes how fast capacity gets added, who gets credit, and what happens if the underlying assumptions stop holding.

This deal is not a scheme. The people structuring it are good at their work. But "the chip designer is insuring the resale value of its own chips" is the load-bearing sentence here, and walking away from the headline number without saying it out loud is how the next mispricing starts.

Sources

This is not financial advice.