Broadcom's Backstop on Anthropic's $35B Chip SPV Concentrates a New Systemic Risk
A vendor guaranteeing residual value on the chips it sells re-concentrates $30B of AI accelerator risk onto one balance sheet. Markets are pricing it as investment grade.
Apollo Global Management and Blackstone have finalized a roughly $35 billion debt package to buy Google TPU accelerators for Anthropic — one of the largest private-credit transactions ever assembled, closing just days after Anthropic's confidential IPO filing at a reported $965 billion valuation. The structure, not the size, is what markets should be studying. A special-purpose vehicle borrows the money, buys the chips, and leases them back to Anthropic, keeping $35 billion of equipment debt off the lab's pre-IPO balance sheet. And underneath the two senior tranches sits a Broadcom guarantee on the residual value of the very chips Broadcom helped supply.
The tranching tells the story
The facility splits into three debt layers plus a thin equity sliver. Roughly $6 billion of A1 notes priced at about one point over Treasuries; roughly $24 billion of A2 notes near a 5.75% coupon; and about $4.5 billion of B notes near 8.5%. Apollo's Atlas SP Partners put up on the order of $800 million in equity, owning the SPV. Roughly half the debt was syndicated to outside investors.
Tranche coupons (%): the 275bps gap is the price of the Broadcom backstop
| label | value |
|---|---|
| A1 (backstopped) | 5.3 |
| A2 (backstopped) | 5.75 |
| B (no backstop) | 8.5 |
The A1 and A2 layers benefit from Broadcom's support; the B notes do not, which is why they yield roughly 275 basis points more and absorb technology-obsolescence risk directly. That spread is the market quoting, in real time, the cost of insuring against a collapse in used-TPU prices. The Broadcom backstop is worth several hundred million dollars a year in saved interest on the senior notes — and that is precisely why it exists.
The market-impact read
For investors, three implications matter more than the headline.
One: the senior notes are rated as a claim on Broadcom, not on Anthropic. The A1 tranche carries investment-grade private ratings despite ultimate repayment depending on an unprofitable AI lab, the resale value of the fastest-depreciating asset class in modern finance, and a vendor guarantee on its own product. Strip the backstop and these are a leveraged bet on used-chip prices. The rating is doing what the monoline wraps did in the mid-2000s — converting volatile collateral into apparent high-grade paper by leaning on a guarantor's balance sheet.
Two: the diversification is an illusion. SPVs exist to disperse risk. The residual-value backstop does the opposite for the senior slice, re-concentrating roughly $30 billion of tail risk onto one semiconductor balance sheet. Because the guarantee makes senior funding cheaper, the economics push every future deal toward the same structure. By late 2025, tech firms had already moved over $120 billion of AI infrastructure debt off their books through SPVs — Meta's ~$30 billion Hyperion deal with Blue Owl, Oracle's string of OpenAI-linked vehicles. If vendor backstops spread across that base, a handful of chip and arranger balance sheets become the implicit insurers of the entire accelerator stock.
Selected off-balance-sheet AI infrastructure SPVs (USD billions, approx.)
| label | value |
|---|---|
| Anthropic SPV | 35 |
| Meta Hyperion | 30 |
| Oracle/OpenAI | 55 |
| Other SPVs | 40 |
Three: the correlation is the danger. The scenario in which Anthropic can't pay its lease is the same scenario in which used-TPU prices have collapsed — that is when the residual-value guarantee gets called, across multiple deals at once. Michael Burry has argued hyperscalers are understating accelerator depreciation by roughly $176 billion through 2028. In an SPV, that soft accounting debate becomes a hard, contractual claim that triggers all at once.
The historical comp
The rhyme is with vendor financing in the telecom bust. Lucent committed about $8.1 billion in customer financing and took $3.5 billion of bad-debt provisions in 2001-2002; Nortel lent billions to carriers that failed and followed them into bankruptcy. A vendor taking on the credit and residual risk of its own customers' purchases flatters everyone's numbers — until demand turns, at which point the collateral and the counterparties weaken together.
The watch items are concrete: whether the next large AI SPVs (OpenAI, xAI) add vendor backstops; the secondary market for used H100s and TPUs as the leading residual-value indicator; any rating-agency move on a guarantor's structured-finance exposure; and Anthropic's own cash flows once the S-1 goes effective. The financial engineering is sophisticated and the builders are not naive. But the feature that makes the deal cheap — a vendor wrap on volatile collateral — is the same feature that turned sector-specific demand disappointments into financial-system events in 2001 and 2008. Frontier-AI capex is now structured credit, and the senior tranche is priced as if that is a settled, safe fact.
Sources: Investing.com, Fortune, Cryptopolitan, Real Investment Advice.