Google's $200bn Chip Finance Machine Rests on Anthropic

Google has assembled roughly $200 billion of interlocking contracts to put its own AI chips into Anthropic's hands — a financing programme large enough that the Financial Times, which pieced it together this week, described it as one of the biggest infrastructure funding efforts ever built. More than $150 billion of it is tied to tensor processing units. Almost none of it sits on Google's balance sheet.

The mechanism is old; only the scale is new. A special-purpose vehicle, structured with help from Morgan Stanley, buys the TPUs and leases them to Anthropic. Outside investors fund the vehicle. That is vendor financing of the kind Boeing and GE built decades ago to move aircraft and jet engines — the manufacturer arranges the money so the customer can afford the hardware, and the asset itself is the collateral.

Four parties, one chain

The first tranche closed in June: $35 billion of private credit led by Apollo alongside Blackstone and a group of global banks. Broadcom, which co-develops the TPUs with Google, backstops roughly $30 billion of that — meaning it absorbs the residual-value shortfall if Anthropic stops paying. Broadcom has its own commitments running the other way, with purchase obligations of $55.2 billion for fiscal 2027 and $72.9 billion for fiscal 2028.

Underneath the chips sit the buildings, and underneath the buildings sits Google's credit. Google now guarantees up to $44 billion of third-party data-centre lease payments, according to The Information — up from about $6.5 billion at the end of September 2025, a near-sevenfold increase in nine months. The guarantees cover named projects: $3.2 billion at Lake Mariner in upstate New York, $7 billion at River Bend in Louisiana, $1.4 billion in Texas. Where Google has provided a guarantee it has sometimes taken penny warrants in return, converting a contingent liability into an equity stake.

The crypto miners got a second life

The developers on the other side of those leases are, to a striking degree, former bitcoin miners. TeraWulf, Cipher and Hut 8 all had the one thing the AI buildout cannot conjure quickly — energised sites with grid interconnects already approved — and all three have converted capacity into Anthropic-bound compute. TeraWulf signed a 20-year, roughly $19 billion lease with Anthropic in July; Hut 8 added a second 352 MW lease in Texas worth $9.8 billion weeks later. Across ten identified developments the FT counts about 2.4 GW, against a TPU deployment commitment of some 4.5 GW.

The reason Google bothers is competitive, not financial. Nvidia sells the default AI chip; Google sells an alternative that almost nobody outside Google had deployed at scale until Anthropic did. A guarantee from Alphabet's balance sheet is the sweetener that makes a TPU site bankable, and the discount is measurable — the FT puts the median borrowing cost on Google-backed projects at 7.1%, against 9.3% for comparable operators in the Nvidia ecosystem. Two percentage points on a multi-billion-dollar construction bond is the whole argument.

Everything depends on one tenant

Strip out the vehicles and the warrants and the backstops, and the structure reduces to a single question: does Anthropic pay its leases for twenty years? Roughly $200 billion of contracts are underwritten by the revenue of a company that did not exist five years ago. That is not a hidden risk — it is the risk, and the participants have priced it, which is what the Broadcom residual guarantee and the Google lease guarantee are for.

Ratings analysts are less relaxed about the accounting than about the credit. Moody's analysts David Gonzales and Alastair Drake noted that the recognised liability for these guarantees "is unlikely to reflect certain plausible future scenarios" — a guarantee costs nothing until the day it costs everything. Moody's counts $1.65 trillion of off-balance-sheet AI obligations across the five big US hyperscalers, including $662 billion of lease commitments that have not yet commenced, equal to 113% of their adjusted debt.

The concentration cuts both ways, and it is not unique to Google. Nvidia has weighed a $250 billion backstop so OpenAI can buy Nvidia chips, and Meta sold 80% of a data centre rather than borrow against it. Different structures, identical dependency. As Jefferies analyst Jonathan Petersen put it, an entire industry has been built underneath a handful of buyers, and if their appetite to invest slows, all of it slows with them.