Anthropic's $35B Chip Financing: AI Capex Is Now Structured Credit

When a company worth nearly a trillion dollars on paper decides it would rather not put $35 billion of equipment debt on its own balance sheet, the financial engineering that results tells you more about the state of the AI industry than any benchmark score. In early June 2026, days after confidentially filing for an IPO at a reported $965 billion valuation, Anthropic finalized one of the largest private-credit transactions ever assembled: a roughly $35 billion debt package, arranged by Apollo Global Management and Blackstone, to buy Google TPU accelerators that a special-purpose vehicle then leases back to Anthropic.

The headline number is impressive. The structure is the actual story. This is not a loan to Anthropic. It is a securitization of compute. A bankruptcy-remote shell company borrows the money, takes title to the chips, and rents them out. Anthropic never owns the silicon and never reports the debt. And sitting underneath the two largest tranches is a guarantee from Broadcom — the very vendor whose custom accelerator program supplies the chips — promising to make investors whole if the hardware is ever sold for less than the notes are worth.

Read that sentence again, because it is the whole article in miniature. The chip vendor is backstopping the credit risk on the chips it sells. We have seen this movie before. It played in the late 1990s, it was called vendor financing, and the closing credits ran through the bankruptcy courts of Lucent, Nortel, and a generation of telecom carriers. This piece is not about whether Anthropic is overvalued — I have written about the valuation question elsewhere. It is about the debt structure underneath the valuation, and why turning frontier-AI capex into asset-backed credit with a vendor backstop concentrates systemic risk in a way the market is pricing as if it were investment grade.

The deal, decoded

Let's start with what was actually sold, because the tranching is where the risk lives. According to reporting on the completed transaction, the $35 billion facility breaks into three layers of debt plus a thin equity sliver.

The A1 notes — roughly $6 billion — carry investment-grade private ratings and priced at about one percentage point over Treasuries. The A2 notes — roughly $24 billion, the bulk of the deal — carry a coupon near 5.75%. Both the A1 and A2 layers benefit from Broadcom's support. The B notes — about $4.5 billion — carry a substantially higher coupon, reported near 8.5%, precisely because they sit outside the Broadcom backstop and absorb technology-obsolescence risk directly. Underneath all of it, Apollo's Atlas SP Partners structured-finance arm contributed on the order of $800 million in equity, effectively making it the owner of the SPV. Roughly half of the total debt was syndicated out to outside investors rather than held by the arrangers.

That spread between tranches is not arbitrary. It is the market pricing the single most important variable in this entire structure: what happens to the value of a Google TPU two, three, or four years from now.

The 275-basis-point gap between the A2 coupon and the B coupon is the cost of the Broadcom guarantee, expressed in yield. Investors will accept 5.75% on $24 billion when a $500-billion-market-cap semiconductor company stands behind the residual value of the collateral. Strip that guarantee away, and the same risk demands 8.5%. The backstop is worth, very roughly, several hundred million dollars a year in saved interest — which is exactly why it exists, and exactly why its existence should make you nervous.

Bankruptcy-remoteness and the recourse illusion

To understand why this structure is attractive — and where it is brittle — you have to understand what an SPV is actually for. A special-purpose vehicle is a legally distinct shell entity, designed to be "bankruptcy remote." That is a term of art with real consequences. It means the vehicle is structured so that if Anthropic itself goes bankrupt, the chips owned by the SPV are not dragged into Anthropic's bankruptcy estate, and the SPV's noteholders have first claim on the hardware. Conversely, the SPV's debt is not Anthropic's debt; it sits in a separate box, invisible to Anthropic's own credit metrics and, crucially, to the balance sheet it will show prospective IPO investors.

This is the same legal technology that powers aircraft leasing, equipment finance, and the mortgage securitizations of the 2000s. The lender's comfort comes from two places: the contractual lease payments, and recourse to a physical asset they can repossess and sell. On paper, that dual protection looks robust. In practice, the second leg — recourse to the asset — is only as strong as the asset's liquidation value, and that is precisely the variable the Broadcom backstop exists to paper over.

Here is the subtle point that gets lost in the elegance of the structure. "Recourse to physical collateral" sounds reassuring, the way "backed by real estate" sounded reassuring in 2006. But collateral protection is worthless if the collateral cannot be sold for what you are owed. A repossessed Boeing 737 can be re-leased to dozens of carriers worldwide; there is a liquid, decades-deep market. A repossessed cluster of two-year-old TPUs can be sold to... whom, exactly, in the scenario where the largest buyer of frontier accelerators just defaulted? The recourse is real, but the realizable value behind it is speculative. The SPV converts a speculative residual into the appearance of hard collateral, and the vendor backstop converts the appearance of hard collateral into the appearance of investment-grade credit. Each layer launders a little more uncertainty into apparent safety.

The thin equity sliver matters here too. With roughly $800 million of equity beneath $34 billion of debt, the SPV is levered on the order of 40-to-1 against the value of the chips. That equity cushion absorbs the first losses, but it is so thin that even a modest shortfall in residual value blows through it almost immediately and lands on the noteholders — and, for the senior tranches, on Broadcom. The whole point of the capital structure is that the equity holder (Apollo's Atlas SP) is paid a handsome return for taking a sliver of first-loss risk, while the catastrophic tail is exported to the senior noteholders and their guarantor. That is efficient risk-transfer right up until the tail event, at which point it is a very large bill arriving at a single address.

What a "residual value agreement" actually promises

The Broadcom backstop is not a guarantee of Anthropic's lease payments in the ordinary sense. It is structured as a residual value agreement. Here is the chain of events it is designed to cover.

The SPV owns the TPUs and leases them to Anthropic. Anthropic's lease payments service the debt. If Anthropic keeps paying, everyone is fine and Broadcom's guarantee is never invoked. But if Anthropic defaults — stops paying the lease — the SPV repossesses the chips and sells them into the secondary market to repay noteholders. If those chips fetch less than the outstanding balance on the A1 and A2 notes, Broadcom covers the shortfall.

So the guarantee is a bet on two correlated events happening together: Anthropic failing to pay, and used TPUs being worth less than expected at the moment of failure. The structure assumes those two risks are independent. They are not. The scenario in which Anthropic can no longer afford its compute lease is, overwhelmingly likely, the same scenario in which demand for frontier-training accelerators has cooled — which is precisely the scenario in which the secondary market for those accelerators is glutted and prices have collapsed. The backstop is most likely to be called in exactly the state of the world where it is most expensive to honor.

This is the crux. Asset-backed credit works when the underlying asset has a stable, well-understood liquidation value — think aircraft leasing, where a used 737 has decades of price history and a deep global resale market. AI accelerators have none of that. They are perhaps the fastest-depreciating multi-billion-dollar asset class ever to be securitized at this scale, and the people structuring the deal know it. That is why the equity is thin, the term is short, and the vendor itself had to step in to make the senior notes pencil.

The depreciation problem nobody can price

How fast does a Google TPU actually lose value? Nobody knows, and the honest answer matters enormously because it determines whether the Broadcom backstop is a formality or a live grenade.

The public debate over AI-hardware depreciation has been running hot for months. The investor Michael Burry has argued that hyperscalers depreciating Nvidia-class accelerators over five or six years are overstating the economic life of the hardware, which he pegs closer to two or three years given the cadence of new silicon. He has put a number on the alleged distortion: on the order of $176 billion of understated depreciation and overstated profit across the industry between 2026 and 2028. The counterargument, advanced by analysts who track the secondary market, is that an H100 still retains somewhere between 60% and 83% of its value 18 months after purchase because inference workloads cascade down the stack long after a chip falls off the training frontier.

Both sides are partly right, and that is the problem. The truth is that residual value depends entirely on a variable nobody can forecast: whether, in 2028 and 2029, there is still ravenous demand for frontier-scale training and inference compute. If the demand curve keeps bending up, used TPUs hold value and the backstop is decorative. If demand plateaus or the next architecture makes today's accelerators uneconomic to run, residual values fall off a cliff and the guarantee gets called. The chip cadence makes this acute. Nvidia alone has moved to an annual rhythm — Hopper in 2022, Blackwell in 2024, Rubin in 2026, Rubin Ultra in 2027 — and Google's TPU roadmap is on a comparable pace. Every new generation is a downward shock to the resale value of the prior one.

Here is what makes the SPV structure genuinely novel and genuinely dangerous: it converts a soft accounting question — how aggressively should we depreciate these assets — into a hard contractual one. In a hyperscaler's financials, an over-optimistic depreciation schedule shows up as gradually disappointing earnings. In this SPV, an over-optimistic residual assumption shows up as a multi-billion-dollar claim against Broadcom's balance sheet, triggered all at once, at the worst possible moment. The accounting debate is academic. The residual value agreement is a legally enforceable promise.

Off the balance sheet, into the system

Anthropic is not doing anything its peers haven't done. The off-balance-sheet financing of AI infrastructure has become the dominant capital structure of the buildout, and the numbers are staggering.

In October 2025, Meta completed what was then the largest private-credit data-center deal in history: a roughly $30 billion package for its Hyperion facility in Louisiana, routed through an SPV co-owned by Meta and Blue Owl Capital, with about $27 billion in A+-rated debt anchored by PIMCO and BlackRock. Meta leases the finished campus back under a long-term operating lease — converting what would have been capex into opex and keeping the debt off its corporate balance sheet. Oracle has structured a string of similar vehicles to fund the compute it leases to OpenAI, including a Blue Owl/JPMorgan SPV for its Abilene, Texas site and tens of billions more for facilities in Texas, Wisconsin, and New Mexico. By late 2025, analysts estimated that tech companies had moved over $120 billion of data-center financing off their books and onto Wall Street's through these structures.

The appeal is obvious. For a company about to go public, keeping $35 billion of equipment debt off the S-1 makes the balance sheet look dramatically cleaner. Lenders, meanwhile, get recourse to physical collateral and a yield well above corporate investment grade in a market starved for it. Private credit, sitting on enormous dry powder, gets to deploy at scale into an asset class with a compelling growth narrative. Everyone at the table is rationally optimizing their own position. That is precisely how systemic risk accumulates — not through villainy, but through a chain of locally sensible decisions whose correlations only become visible under stress.

What the Anthropic deal adds to the Meta and Oracle template is the vendor backstop. Meta's Hyperion SPV did not have Nvidia guaranteeing the residual value of its GPUs. Anthropic's does have Broadcom guaranteeing the residual value of its TPUs. That is the new ingredient, and it is the one that rhymes most loudly with history.

The ghost of vendor financing

In the late 1990s, telecom equipment makers discovered they could sell far more switches and routers if they simply lent their customers the money to buy them. Lucent committed something like $8.1 billion in vendor financing. Nortel extended billions more — by some accounts it lent over $7 billion to help start-up carriers buy its equipment, much of it unsecured and tied to future purchases. Cisco played the same game on a smaller scale. The mechanism flattered everyone's numbers: the vendors booked the revenue, the carriers got their gear, and the build-out roared.

Then financing conditions tightened, the carriers slashed capex, and the start-ups that had borrowed to buy began to fail. WorldCom, Global Crossing, 360networks — the customer base evaporated. The vendors were left holding loans to bankrupt counterparties and warehouses of unsellable inventory. Lucent took bad-debt provisions of $2.2 billion in 2001 and $1.3 billion in 2002 — $3.5 billion in customer-loan losses on financing it had extended to juice its own sales. Nortel's customers failed, and Nortel followed them into bankruptcy; a stock that peaked above $86 in July 2000 was worth pennies a decade later.

The structural rhyme is exact. In both cases, a hardware vendor with strong incentives to move product takes on the credit and residual risk of its own customers' ability to pay. In both cases, that risk is hidden from the buyer's headline financials and dispersed to third-party investors who are pricing it as if the collateral has a stable resale value. In both cases, the risk is most likely to materialize precisely when the underlying demand thesis fails — which is also the moment the collateral is worth the least and the counterparties are weakest.

The differences matter too, and they are not all reassuring. Broadcom is far larger and more diversified than Lucent or Nortel ever were, and its backstop covers a residual-value shortfall rather than an outright loan, which is a narrower exposure. But the modern structure is also far more leveraged, far more concentrated, and far more interconnected. A handful of arrangers — Apollo, Blackstone, Blue Owl, PIMCO, BlackRock — appear across nearly every one of these deals. The collateral across all of them is the same narrow category of fast-depreciating AI accelerators. And the demand underpinning all of it rests on a small number of frontier labs whose own revenues are, themselves, substantially recycled among one another and their cloud providers.

It is worth dwelling on the mechanism of the telecom collapse, because the sequence is instructive. Vendor financing did not blow up because anyone committed fraud at the outset. It blew up because of a feedback loop. Cheap vendor credit let marginal carriers buy equipment they could not otherwise afford, which inflated the vendors' reported revenue and stock prices, which lowered the vendors' cost of capital, which let them extend even more credit to even more marginal customers. Each turn of the loop looked like growth. The loop ran in reverse the moment the marginal customers could no longer service their debt: capex was slashed, vendor revenue collapsed, the loans went bad, the inventory became unsellable, and the stock prices that had financed the whole thing evaporated. The system was not fragile because any single decision was reckless. It was fragile because the same risk — "will telecom demand keep growing" — sat under the revenue, the collateral, and the customers' solvency all at once. They were one bet wearing three costumes.

The AI accelerator complex has the same hidden monoculture. The lab's ability to pay its lease, the resale value of the chips, and the demand thesis that justifies the whole capex wave are not three independent risks that a diversified pool can net against. They are one risk — "will frontier-AI compute demand keep compounding" — appearing as a lease payment, a residual value, and a guarantor's exposure. Diversification across many SPVs does nothing to reduce a risk that is identical in every SPV. This is the precise error that made mortgage securitization so dangerous: pooling thousands of loans feels like diversification, but if every loan depends on the same thing — house prices never falling nationally — then the pool is not diversified at all. It is one bet, finely sliced.

The equipment-leasing and securitization lineage

Strip away the AI gloss and what Anthropic has done is a sophisticated equipment lease wrapped in a securitization wrapped in a credit guarantee. None of these three layers is new; what is new is stacking them on an asset class with no resale history and a brutal obsolescence curve.

Equipment leasing with a manufacturer guarantee is decades old. Auto-leasing securitizations routinely rely on residual-value assumptions, and when those assumptions prove too rosy — as they did when used-car prices softened in the late 1990s and again periodically since — the residual-value insurance written against them takes losses. The entire residual-value-insurance industry exists because lessors learned, painfully, that betting on the future resale price of a depreciating asset is a real and frequently mispriced risk. What Broadcom has effectively done is write a giant slug of residual-value insurance on TPUs, an asset class for which there is no actuarial history to price against. Auto-lease residual insurers have decades of used-car price data. Broadcom is underwriting the 2028 price of a 2026 accelerator with essentially a guess, however informed.

The securitization layer adds a second well-known hazard: the originate-to-distribute incentive. When the party structuring a deal expects to syndicate most of the risk to outside investors — and here, roughly half the debt was sold on — its incentive to scrutinize the underlying collateral quality weakens. This is not a hypothetical; it is the central lesson of 2008. The arrangers who keep only a thin slice have less skin in the residual-value assumption than the noteholders who buy the distributed paper and the guarantor who wraps it. The structure quietly separates the people making the optimistic residual assumption from the people who eat the loss if it is wrong.

The circularity nobody wants to name

There is a final, uncomfortable feature of this ecosystem that the telecom comparison surfaces: the money goes in circles. Google supplies the TPUs and is also a major investor in Anthropic and a beneficiary of the compute spend. Broadcom co-develops the accelerators and now guarantees their residual value. The same private-credit shops arrange the deals, hold slices, and raise the funds that buy the rest. The frontier labs buy compute from cloud providers who are also their investors, and sell services back into the same ecosystem.

This is not necessarily round-tripping in the fraudulent sense. But it is circular financing, and circular financing has a well-documented failure mode: it makes the whole structure look more robust than it is, because the same dollars are counted as revenue, investment, and collateral support at different points in the circle. When the circle is spinning forward, every participant's numbers reinforce every other participant's numbers. When it stops, the reinforcement runs in reverse — exactly as it did when the telecom carriers stopped buying and the vendors who had financed them collapsed. The Anthropic deal is one node in a dense web where chip vendors, cloud providers, labs, and private-credit arrangers are all financially entangled with one another. The vendor backstop does not stand outside that web. It is woven into it.

Where the risk actually concentrates

Trace the exposure and a troubling picture emerges. The point of an SPV is to disperse risk — to move it off Anthropic's books and spread it among diversified credit investors. But the Broadcom backstop does the opposite for the senior tranches: it re-concentrates the tail risk of $30 billion in notes onto a single corporate balance sheet.

In the ordinary case, this is fine — Broadcom is enormous and the guarantee is never called. But systemic risk is not about the ordinary case. It is about the correlated bad state. Imagine a 2028 in which AI training demand has plateaued, accelerator resale values have fallen faster than the depreciation curves assumed, and one or more frontier labs are struggling to cover their compute leases. In that world, the residual-value guarantees on multiple of these deals come due at once. If the vendor-backstop model spreads — and the economics strongly incentivize it to spread — then a small number of semiconductor balance sheets become the implicit insurers of the entire AI accelerator stock.

That is the systemic concern in one sentence: a financing innovation designed to make individual deals safer for individual buyers may be making the system as a whole more fragile, by routing correlated residual-value risk through a tiny set of vendor and arranger balance sheets. The diversification is an illusion when the collateral is identical and the demand driver is shared.

The interest-coverage question

There is a simpler, more immediate question that the elaborate structure tends to obscure: can Anthropic actually afford the lease payments? The whole edifice rests on a stream of payments from a company that, however richly valued, is still deeply unprofitable and burning cash to fund its own growth. I have written before about how compute economics and agent demand strain even the best-funded labs.

Consider the order of magnitude. A $30 billion blended debt load at roughly a 6% blended coupon is on the order of $1.8 billion a year in interest alone, before any principal amortization or the lease margin that makes the SPV's equity worthwhile. Anthropic's revenue has grown spectacularly, but its compute costs have grown alongside it, and the lease payments on this facility are senior, contractual, and unforgiving in a way that equity-funded capex is not. The flexibility a company normally has to throttle capex in a downturn is precisely what a long-dated lease takes away.

If Anthropic's revenue trajectory continues vertical, the lease is trivially serviceable and this entire analysis is a footnote. The deal is a bet — by Apollo, by Blackstone, by Broadcom, by the syndicated noteholders — that it will. The structure is sound only as long as the growth story is sound. That is true of every leveraged bet, but it is worth saying plainly because the layers of legal insulation and the investment-grade rating on the senior notes are doing a lot of work to make a bet on uninterrupted frontier-AI demand look like a safe bond.

The disclosure problem the IPO will force

The off-balance-sheet design is most valuable precisely at the moment it is most consequential: the IPO. Keeping $35 billion of equipment debt out of the consolidated balance sheet flatters every leverage ratio a prospective public-market investor will scrutinize. But disclosure rules around variable-interest entities and off-balance-sheet arrangements are not nothing, and the S-1 process will force at least some sunlight onto the structure. The interesting question is how much.

A lease is still a lease. Under modern lease-accounting standards, a long-dated, non-cancelable operating lease shows up as a right-of-use asset and a corresponding liability — so the obligation does not vanish entirely from Anthropic's financials even if the SPV's $35 billion of debt does. The accounting magic is narrower than the headlines suggest: it moves the leverage from "debt the company owes" to "lease payments the company is committed to," which carries different optics and different covenant implications but is not free money. What the structure genuinely achieves is removing the chip vendor's and lenders' credit exposure from Anthropic's own credit profile, and shifting the residual risk to Broadcom. When the S-1 becomes public, the most valuable pages will be the related-party and commitments footnotes, where the true contours of the lease obligation and the Broadcom arrangement have to be described. This connects directly to the broader question of how to mark these AI assets and liabilities to market when the accounting gives management so many levers to pull.

A simple stress test makes the stakes concrete. Suppose by 2028 a new accelerator generation halves the resale value of today's TPUs and frontier-training demand flattens. The SPV's thin equity is wiped out immediately. The B-note holders, who priced obsolescence risk directly at 8.5%, take losses next. Then the repossession-and-resale shortfall on the $30 billion of senior notes lands on Broadcom's residual-value guarantee — potentially billions of dollars, arriving at once, on a single balance sheet, in the same quarter that the AI-demand narrative is unraveling and Broadcom's own equity is under pressure for exactly that reason. That correlated timing is the whole risk in one paragraph.

Why the rating is the most interesting number

The single most revealing fact in the whole deal is that the A1 notes carry investment-grade private ratings. Investment grade is a statement about the probability of loss. It says, in effect, that a rating agency looked at a structure whose ultimate repayment depends on (a) the continued solvency of an unprofitable AI lab, (b) the residual value of the fastest-depreciating asset class in modern finance, and (c) a vendor guarantee from a chip company on the chips it sells — and concluded the senior slice is about as safe as a high-quality corporate bond.

That conclusion rests almost entirely on the Broadcom backstop. Remove it and the senior notes are a leveraged bet on used-TPU prices. With it, they are rated as a claim on Broadcom. So the question that should keep risk managers up at night is not "is Anthropic good for it" but "how many of these residual-value guarantees can Broadcom write, across how many labs, before the rating agencies have to start treating its own balance sheet as the concentrated AI-collateral risk that it is becoming." We saw with the monoline insurers in 2008 what happens when a guarantor's AAA is the load-bearing assumption beneath a whole asset class, and the guarantor turns out to be undercapitalized for the correlated scenario.

I am not predicting that outcome. Broadcom today is a healthy, diversified company and a single residual-value agreement is well within its capacity. The point is structural: the model only delivers cheap senior funding because it concentrates risk on a guarantor, and the model's own economics incentivize that guarantor to write more and more of it. The thing that makes the deal work is the thing that makes the system fragile.

What to watch

A few concrete signposts will tell us which way this is breaking, and I have laid out a specific, dated prediction about how this financing model spreads.

First, watch whether vendor backstops become standard. If the next several large AI-accelerator SPVs — for OpenAI, xAI, or others — also feature residual-value guarantees from Nvidia, Broadcom, or AMD, then the concentration thesis is playing out and the vendor balance sheets are quietly becoming the system's insurers.

Second, watch the secondary market for accelerators. A sustained softening in used-H100 and used-TPU prices is the leading indicator for residual-value stress, and it will show up in resale listings long before it shows up in any SPV's financials.

Third, watch the rating agencies. Any move to put a vendor guarantor's structured-finance backstop exposure on credit watch would be the 2026 equivalent of the first monoline downgrade — a small headline with enormous implications.

Fourth, watch Anthropic's own cash flows once the S-1 goes effective and the financials become public. The whole structure is a derivative of one question: can the lab pay its compute lease through a full demand cycle. Everything else is plumbing.

The financial engineering here is genuinely sophisticated, and the people who built it are not naive. They have simply solved the problem in front of them — how to fund an enormous, fast-depreciating capital base for a pre-profit company without wrecking its IPO balance sheet — by reaching for the most powerful tool in modern finance: securitization with a credit wrap. It is elegant. It is also exactly the tool that, in telecom in 2001 and in structured credit in 2008, turned a sector-specific demand disappointment into a financial-system event. Frontier-AI capex is now structured credit. Whether that is innovation or the setup for the next writedown depends entirely on a residual-value curve that nobody can see.

Further Reading

• Anthropic's $965B IPO and the S-1 Bubble Test
• Anthropic, SpaceX, Colossus: Compute Rent and Agent Demand
• Prediction: Frontier labs fund the majority of new accelerator purchases via structured-credit SPVs by 2027