Google-Broadcom-Anthropic-Apollo TPU Financing Structure Detailed

Google-Broadcom-Anthropic-Apollo: What the TPU Financing Structure Actually Says

Google, Broadcom, Anthropic, and Apollo have structured a way to fund Anthropic's access to Google's custom AI chips (TPUs) by spreading default risk across multiple deep-pocketed guarantors. If Anthropic can't make its lease payments, Broadcom covers lenders by taking over the lease or selling the racks. Google separately backstops the data-center rent. SemiAnalysis reports Nvidia is building a thinner version of this for its own customers.

The structure matters beyond the deal itself. It tells you how frontier AI compute is actually being financed, and who's really propping up whom.

The subsidy is the story

Strip the guarantees and this deal doesn't close at cheap rates. Google's rent backstop and Broadcom's residual-value guarantee are what make lenders comfortable. Without them, no conservative lender touches the paper. That means Anthropic's compute access isn't a market relationship; it's a strategic dependency on Google's willingness to stay in the structure.

The Nvidia variant makes this plain. Nvidia's version carries a 25% residual-value guarantee and no rent backstop. That's what the market actually prices when a hyperscaler isn't absorbing the tail risk. The delta between the two structures is a measure of the implicit subsidy Google is handing Anthropic that Nvidia won't hand its customers.

Google doesn't need an equity stake or a board seat to constrain Anthropic. It holds the infrastructure Anthropic can't replace. Whoever controls the compute backstop controls a quiet off switch, and that's not something Anthropic's safety commitments or independence framing addresses.

The residual guarantee is unproven

Broadcom is asserting that used TPU racks have enough recovery value to make lenders whole in a default. That claim has never been tested. TPUs are Google-proprietary chips with essentially no secondary market. The guarantee is a credit promise, not a reflection of what the hardware fetches at auction. Even the Nvidia version at 25% is generous if everyone is dumping racks at once in a downturn. The first actual default will be instructive.

What this means for anyone buying compute at scale

Large-scale compute reservations are increasingly a credit-market product. Getting access at good terms is less about whether you can pay and more about whether you have a creditworthy backer willing to get into the structure alongside you. A startup without a Google, Microsoft, or Amazon absorbing counterparty risk is competing for un-backstopped compute at worse terms. The Anthropic deal isn't a menu item; the Nvidia program is closer to what an independent buyer can actually access.

The call

Nvidia's capital-partnership program will close its first publicly reported deal only with a hyperscaler or comparable investment-grade party also on the hook, not on Nvidia's 25% guarantee alone. Conservative lenders won't fund a chip vehicle backed by a single guarantee on hardware with no proven distressed resale market. The pattern the Anthropic deal established, layering creditworthy names into the structure to absorb tail risk, is what makes the math work. Expect that pattern to repeat. Call revisits by mid-2027.

Google, Broadcom, Anthropic, and the private-equity firm Apollo have wired together a way to fund Anthropic's AI chips that spreads the risk of default across several deep-pocketed backers. If Anthropic can't make its lease payments on Google's custom AI chips (TPUs), Broadcom takes over the lease or sells the racks and covers the lenders. Google separately promises to keep paying the data-center rent. SemiAnalysis says Nvidia is copying a thinner version of this for its own customers. The bigger point for anyone building with AI: the way frontier labs get their compute is now a credit-market product, and the terms tell you who's really propping up whom.

This is a briefing, not a decision you can undo. What it changes is how you read the compute market you depend on. No deadline forces a move. But the structure is a signal worth decoding before the next round of "we secured X billion in compute" press releases lands.

The Skeptic Strip the guarantees and this deal doesn't close. Google backstops the data-center rent. Broadcom guarantees the racks are worth something if they get sold. Take either one away and no conservative lender touches the paper at a cheap rate. So the real content here is that Anthropic's compute depends on Google staying willing to pay. That's a strategic relationship, not a market. Calling this a "template for AI compute financing" flatters a one-off into a general model. The Nvidia copy tells you the truth: 25% residual guarantee, no rent backstop. That's what the market actually bears when a hyperscaler isn't on the hook for the whole thing.

The Safety Lens Here's the part that doesn't show up in Anthropic's safety docs. If Anthropic's ability to train frontier models runs through Google's willingness to service this vehicle, Google holds a quiet off switch. Not through owning equity. Through owning the infrastructure Anthropic can't replace. Constitutional AI, the safety commitments, the independence framing, none of it addresses who can cut the power. Regulators worried about frontier model control keep looking at ownership and board seats. The real leash is a private credit vehicle most of them will never read. Ask who can actually constrain Anthropic's compute, and the answer is Google and Broadcom.

The Compute Pragmatist The residual value guarantee is where this either holds or breaks. Broadcom is asserting that used TPU racks are worth enough to make lenders whole in a default. That claim is untested. TPUs are Google-proprietary chips with essentially no secondary market. Nobody else runs them. So Broadcom's guarantee is a credit promise, not a reflection of what the hardware fetches at auction. The Nvidia version at 25% is more honest, because GPUs at least have real resale demand. But 25% recovery on distressed AI hardware in a downturn, when everyone's dumping racks at once, is still generous. These guarantees have never been tested by an actual default. The first one will be educational.

The Researcher The interesting move is the disaggregation. This structure splits AI compute risk into separate buckets tied to separate guarantors. Lenders to the chip vehicle carry Broadcom's credit. Lenders to the data center carry Google's credit. Neither takes direct exposure to whether Anthropic's revenue holds up. That's clean, and it's why capital markets can price this at all. The gap between the two structures is itself the signal. Google's version has a full rent backstop and an uncapped residual plug. Nvidia's has a 25% cap and no rent guarantee. That delta measures how much implicit subsidy Google is handing Anthropic that Nvidia won't hand its customers.

The Enterprise Buyer If you're signing for compute at any real scale, this is the map of who's actually creditworthy behind your capacity. The lesson is uncomfortable: securing large reservations is becoming less about whether you can pay cash and more about whether you can get a creditworthy backer into the structure. A startup without a Google, Microsoft, or Amazon willing to make lenders whole is competing for a shrinking pool of un-backstopped compute at worse terms. The bespoke Anthropic deal isn't a menu you can order from. The Nvidia program, thinner as it is, is closer to what an independent buyer might actually access.

Where they disagree Two real splits. First, is this a template or a one-off? The Researcher sees a repeatable way to slice AI credit risk. The Skeptic says the only reason it prices cheaply is Google's subsidy, and the Nvidia variant already shows what happens when you remove it. They're both right, which is the point: the structure generalizes, but the good terms don't.

Second, the residual value guarantee. The Compute Pragmatist says Broadcom is guaranteeing racks with no resale market, so the promise is pure credit. The Researcher treats that same guarantee as a clean risk transfer. The difference matters only in a default, and there hasn't been one yet.

The Safety Lens sits outside both arguments and lands the part nobody's pricing: whoever holds the backstop holds the leash.

What it hinges on Whether this is a durable financing model or a Google favor comes down to one thing: can anyone build a version of this without a hyperscaler eating the counterparty risk? Nvidia's program is the live test. It has a 25% residual guarantee and no rent backstop. If conservative lenders fund it at reasonable rates, the model generalizes. If they demand a hyperscaler co-sign anyway, then the Skeptic is right and this was always about Google keeping Anthropic alive.

Prediction: Nvidia's capital-partnership program, with its 25% residual-value guarantee and no data-center rent backstop, will close its first publicly reported deal only with a hyperscaler or comparable investment-grade party also on the hook alongside Nvidia, not on Nvidia's guarantee alone, by the time SemiAnalysis or the FT reports the first funded transaction under it in the first half of 2027.

Confidence: Medium. The thin guarantee won't clear conservative lenders on its own.

Why: SemiAnalysis lays out that the Nvidia version drops the residual guarantee to 25% and removes the rent backstop that made the Anthropic deal fundable at low rates, and the whole reason the Google structure prices cheaply is that Google and Broadcom absorb the tail risk lenders won't. AI hardware has no proven distressed resale market, so a 25% recovery promise from Nvidia alone leaves lenders carrying real loss risk in a downturn. The way lenders close that gap is by demanding another investment-grade name in the structure, which is exactly the pattern the Anthropic deal established. The opposite outcome, lenders funding a chip vehicle on a single 25% guarantee with no secondary market to test it, would require the credit market to price untested distressed AI infrastructure more generously than it prices assets with real resale liquidity, and that isn't how conservative capital behaves.

Revisit by 2027-06-30: We're right if the first funded deal under Nvidia's program includes a hyperscaler, cloud provider, or comparable investment-grade backer sharing the counterparty risk. We're wrong if a deal closes on Nvidia's 25% guarantee alone with no other creditworthy party in the structure.