Anthropic Borrowed $36B. Broadcom Cosigned $31B of It.

Key Takeaways

• The Structure: Apollo Global Management and Blackstone are syndicating a roughly $36 billion private credit package. A Special Purpose Vehicle (SPV) borrows the money, buys Google’s Tensor Processing Units (TPUs), and leases the hardware to Anthropic. Anthropic gets compute capacity. The debt stays off Anthropic’s balance sheet.
• The Cosigner: Broadcom, which co-develops the TPUs with Google, is providing a residual value support agreement on the $31 billion of senior A1 and A2 notes. If Anthropic stops paying its lease and the chips do not fetch enough at resale, Broadcom covers the shortfall in full.
• The Asymmetry: Anthropic just raised $65 billion at a $965 billion valuation and filed confidentially for an Initial Public Offering (IPO) on June 1, 2026. With $47 billion in run-rate revenue, the company is still not creditworthy enough on its own to price $31 billion of senior debt at investment grade. The deal works only because Broadcom’s name is on it.
• The Rhyme: Telecom vendor financing peaked at roughly $33 billion in 2000, when Lucent and Nortel funded their own customers to buy their own gear. This deal is bigger and uses a more sophisticated wrap. The structural question is the same: what happens to the vendor if the customer cannot pay?

The Cosigned Loan

On May 28, 2026, Bloomberg reported that Apollo Global Management and Blackstone were shopping a roughly $36 billion debt package to additional investors. The deal would rank among the largest private credit transactions ever assembled.

The cash will not flow to Anthropic. It will flow to a special purpose vehicle that buys Google TPUs and leases them to Anthropic for deployment at data centers in New York, Texas, Louisiana, and Indiana. The lease lets Anthropic take on compute capacity without adding the associated debt to its own books.

That part is standard sale-leaseback financing. The new part is the wrap.

Bloomberg’s reporting, confirmed by parallel coverage in Private Equity Wire and PE Insights, describes a debt stack of roughly $6 billion of A1 notes, $25 billion of A2 notes, and $4.5 billion of subordinated B notes. Broadcom is providing a residual value support agreement on the $31 billion of senior A1 and A2 paper. If Anthropic stops making lease payments for a defined period, the SPV sells the chips. If the resale falls short of what is owed, Broadcom covers the entire balance owed to A1 and A2 noteholders.

In plain English: the company that builds the chip is guaranteeing that the company buying the chip will be able to pay for the chip. The lender is taking Broadcom’s credit, not Anthropic’s. Broadcom carries an A- (Positive) rating from S&P Global Ratings. Broadcom’s existing corporate bonds trade in the 4% to 5% yield range. Anthropic, as a private company that has never tapped the public debt markets, has no senior unsecured rating at all.

Why Anthropic Cannot Borrow This on Its Own

The strangest part of the deal is that Anthropic does not look like a company that needs a cosigner.

On the same day the chip deal leaked, Anthropic announced a $65 billion Series H led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. The round valued the company at $965 billion post-money. Run-rate revenue had crossed $47 billion earlier that month. Four days later, on June 1, Anthropic filed a confidential S-1 with the Securities and Exchange Commission (SEC), setting up an October IPO at what would be a trillion-dollar debut if markets cooperate.

The growth curve is real. Run-rate revenue went from $9 billion at the end of 2025 to $14 billion in February, $30 billion in April, and $47 billion in May 2026. Altimeter’s Brad Gerstner wrote in an investor note that the underwriting now turns on “whether the mark is conservative,” not whether the company earns its trillion-dollar valuation.

But growth is not the same thing as creditworthiness. A lender writing a senior loan of this size does not get paid in market capitalization. The lender gets paid in cash flow over a fixed amortization schedule that runs longer than the chip lease itself. Anthropic does not have that cash flow yet. Reporting on Anthropic’s internal projections described $70 billion in revenue and $17 billion in cash flow by 2028, with gross margins improving from roughly 50% in 2025 toward 77% by 2028, supported by an estimated $80 billion in cloud infrastructure spending through 2029. The Information reported in early 2026 that Anthropic had already lowered its 2025 gross margin projection to roughly 40%, ten points below the earlier 50% estimate, after inference costs ran higher than expected.

Equity investors buy the upside. Senior creditors buy the floor. Anthropic has the upside. It does not have a floor that any private credit committee would write tens of billions of dollars of investment-grade paper against without help.

That is what Broadcom’s residual value agreement does. It substitutes Broadcom’s investment-grade floor for Anthropic’s missing one. The yields on the A1 and A2 notes can price near Broadcom’s existing curve, which means the spread Broadcom is implicitly providing is worth billions of dollars per year in interest savings versus what Anthropic would pay on a high-yield basis. That figure is not in any press release. It is the embedded subsidy that makes the deal exist.

What Broadcom Just Agreed To

Hock Tan, Broadcom’s chief executive, has been explicit about who his AI accelerator customers are. On the Q1 fiscal year 2026 earnings call in March 2026, he named Google, Anthropic, OpenAI, and Meta as the customers driving the company’s custom Extreme Processing Unit (XPU) business. AI revenue in that quarter was $8.4 billion, up 106% year-over-year. The company guided to $10.7 billion of AI revenue in the next quarter.

Anthropic is a Broadcom customer. Broadcom is now also guaranteeing Anthropic’s ability to pay for Broadcom’s product. The two facts are inseparable.

Broadcom’s market capitalization is roughly $2.28 trillion. A contingent liability of this size is a small fraction of that. The position will not break the balance sheet. What it does is concentrate Broadcom’s risk on a single counterparty whose financial profile cannot be stress-tested with public data, and on a secondary market for used TPUs that effectively does not exist at scale yet. Google designed the TPU as proprietary infrastructure. The asset’s resale value depends on who else wants to run JAX workloads on Google-Broadcom silicon outside Anthropic, and the answer in mid-2026 is: nobody at scale.

This is the part that does not show up in any covenant. The residual value guarantee is unfalsifiable until the moment it triggers. Broadcom’s analysts will project recovery values; the SPV will accept them; rating agencies will use them; nobody will be able to verify them until Anthropic actually stops paying.

The Lucent Rhyme

By the end of 1999, Lucent Technologies had committed or agreed to extend roughly $8.4 billion in direct vendor financing to telecom customers buying its switching gear and fiber equipment. Industry-wide telecom vendor financing peaked at roughly $33 billion in 2000. When the customers defaulted in 2001, Lucent wrote off most of those loans and its market value collapsed.

The pattern is familiar enough that the site has already documented its earlier 2026 iteration in The Silicon Debt Trap: Apollo’s $7 billion sale-leaseback for xAI, where Nvidia was an anchor limited partner in the SPV that bought its own chips. That deal closed in February. This deal is roughly five times larger and structurally more aggressive. Lucent loaned the cash directly. Apollo-Valor-xAI used a sale-leaseback with the chip maker as an equity participant. Apollo-Blackstone-Anthropic uses a sale-leaseback with the chip co-developer as the explicit guarantor of the senior tranches.

Each new iteration moves the chip vendor closer to the credit. Version one had the vendor as supplier. Version two had the vendor as equity holder. Version three has the vendor as named guarantor of the senior tranches. The risk does not disappear. It migrates closer to the manufacturer’s balance sheet at each step.

The Bank for International Settlements flagged this trajectory in Bulletin No. 120, warning that AI companies are shifting from cash-flow funding to debt funding and that the opacity of private credit markets makes the true systemic exposure impossible to assess. The Anthropic deal is what the BIS was worried about, written one layer more sophisticated than the warning anticipated.

The Steel Man

There are real arguments that this deal is fine and that the Lucent comparison is overdone.

Anthropic is not a speculative customer. Inference demand for Claude is documented in Anthropic’s own revenue figures and in Broadcom’s quarterly AI guidance. A senior Google AI infrastructure executive has stated that seven-to-eight-year-old TPUs in the company’s fleet still run at full utilization, suggesting the residual value assumptions Broadcom is underwriting are not crazy. The hyperscaler picture, though, is mixed. Meta extended its server useful life to 5.5 years effective January 2025, locking in roughly $2.9 billion of reduced 2025 depreciation expense. Amazon went the other way the same month, shortening the useful life of a subset of servers from six to five years effective January 1, 2025, citing the accelerating pace of AI development. Four-year lease amortization sits between the two.

Broadcom can credibly carry the contingent. The company’s debt-to-EBITDA ratio fell from 2.0x to 1.2x over fiscal 2025 and is forecast to drop below 1.0x in fiscal 2026, supported by a $73 billion AI backlog. S&P upgraded the company to A- with a Positive outlook on the strength of that profile. An A- name guaranteeing the senior notes is not an obvious mispricing.

And the deal is structured with delayed draws tied to chip delivery, meaning the SPV does not have the full debt outstanding on day one. Capital is released as the leases turn on. That mechanism alone removes much of the funding risk that hit telecom in 2001.

These are not bad arguments. The honest reading is that the deal probably gets done, the senior tranches probably price near Broadcom’s curve, and Anthropic probably continues to grow into its valuation. The question is what happens if any one of those assumptions breaks.

The Stress Test Nobody Is Running

The scenario that breaks this structure is not “AI fails.” It is “AI inference margins compress.”

Anthropic’s underwriting assumes the gross margin trajectory described in its own internal projections holds through 2028. That assumption depends on Claude pricing power holding while compute costs fall. If a competing frontier model commoditizes inference pricing within the lease term, Anthropic’s contribution margin per query compresses. Lease payments are fixed. Revenue is not.

If lease payments stop, the SPV liquidates the TPUs. The TPU resale market does not yet exist at any meaningful scale because TPUs run a Google-specific software stack. Broadcom pays the gap. Broadcom’s annual free cash flow in fiscal 2025 came in at roughly $26.9 billion, with quarterly free cash flow margins running at 40 to 44% of revenue. The contingent is absorbable, but it is not invisible.

The deeper risk is that the deal becomes the template. If Apollo and Blackstone close this transaction with Broadcom as guarantor for Anthropic, the next round of AI capex deals will probably arrive with Nvidia as guarantor for OpenAI and AMD as guarantor for a hyperscaler. Each iteration moves more AI capex onto chip maker balance sheets that the public equity market has been valuing as if those balance sheets were unencumbered.

That is the lesson worth taking from Lucent that the Anthropic deal has not yet had to learn. The risk did not appear in the telecom customers’ books when they defaulted. It appeared in the vendor’s footnotes, where it had been sitting for two years before anyone read them. Broadcom’s contingent will sit in the same place, until the day it does not.

For now, a startup approaching a trillion-dollar valuation needs a chip maker to cosign its loan. That sentence should be the headline of the story, not the structural detail buried in a Bloomberg paragraph about senior tranches.