The $60 Billion Test: Wall Street Starts Syndicating the Largest Chip-Debt Deal Ever Built

Bank of America, Citigroup, and Morgan Stanley have begun syndicating a $60 billion debt package tied to Anthropic’s AI compute expansion, taking one of the industry’s most unusual financing structures from negotiated backrooms into the broad bank market. First reported by the Financial Times on October 5 and confirmed in syndication letters circulated this week, it is the largest chip-financing transaction ever structured — and the clearest signal yet that AI infrastructure has become an asset class the financial system is being asked to hold at scale.

The timing matters as much as the size. Anthropic is preparing a public offering that could value the company at as much as $2 trillion, and its IPO prospectus discloses the commitments underneath this deal for the first time. Wall Street is effectively being asked to underwrite the compute layer of the frontier-AI economy at the exact moment the sector’s finances are becoming public.

What is actually being sold

The package splits into two tranches with very different risk profiles.

The top layer is a $42 billion Class A senior-secured tranche, backed by Broadcom’s A-minus investment-grade credit rating through a residual value guarantee. An SPV uses the proceeds to buy custom tensor processing units co-developed by Broadcom and Google, leases them to Anthropic, and services the debt from Anthropic’s lease payments. Because Broadcom’s guarantee sits behind the paper, lenders in this tranche are effectively betting on Broadcom’s creditworthiness as much as on Anthropic’s ability to pay.

Below it sits an $18 billion Class B junior tranche led by Blackstone, which has committed roughly $9 billion from its own funds and is helping syndicate the rest. This slice carries no Broadcom guarantee — lenders here are directly assuming Anthropic’s credit risk. According to the FT, underwriters may wait for broader marketing of the junior piece until after Anthropic’s IPO, when the company’s financials will be publicly disclosed.

The capital funds chip orders scheduled for 2027 delivery, with lease payments beginning only after the hardware is fully deployed.

Two different $42 billion instruments — and why conflating them misreads the risk

Even financial reporting has blurred a detail that changes how risky this deal looks. There are two separate $42 billion instruments in this story:

1. The syndicated debt tranche — the $42 billion Class A senior-secured slice now being marketed to banks, described above.
2. The Broadcom convertible loan facility — a commitment disclosed in Anthropic’s IPO prospectus for Broadcom itself to lend Anthropic up to $42 billion directly against infrastructure spending, in notes that can convert into Anthropic equity. Broadcom can designate a financing partner for these notes, and Anthropic has said it does not expect them to be sold before its IPO closes.

Put together, that is roughly $102 billion in financing for which Broadcom is either the guarantor, the direct lender, or both — approximately equal to Broadcom’s entire fiscal 2026 revenue.

This makes Broadcom four things to Anthropic at once: chip supplier, lessor, direct lender, and debt guarantor. Anthropic’s own prospectus acknowledges the concentration, flagging it as a “potential conflict of interest.”

The $125.2 billion lease underneath it all

The debt package is the financing wrapper around a much larger obligation. According to Anthropic’s prospectus and Reuters’ October 1 reporting, the company has signed a five-year, $125.2 billion TPU lease commitment — and the $60 billion syndicated package covers roughly half of it. The earlier facility disclosed in filings could finance about one-third of that commitment.

The chips themselves are next-generation Google TPUs under the Broadcom-Google co-development program announced in April 2026, which committed “multiple gigawatts” of TPU capacity coming online starting in 2027. Broadcom’s AI XPV platform, launched in June 2026 with the initial $35 billion Apollo-and-Blackstone-financed deal, targets more than 20 gigawatts of capacity through 2028. This syndication is the next, far larger installment of that program.

Why a TPU is different collateral than a GPU

Lenders extending credit against AI hardware need to understand one engineering distinction with real credit implications: a TPU is not a GPU.

Nvidia’s GPUs are general-purpose parallel processors — a defaulted GPU-backed loan can be resolved by reselling the collateral into a deep secondary market of gaming, simulation, and inference buyers. A TPU is an application-specific integrated circuit built around Google’s systolic-array architecture, optimized for the matrix multiplication at the core of neural network training and inference. It cannot be repurposed for general compute.

A single seventh-generation “Ironwood” TPU rack connects thousands of chips into one unified system via high-speed interchip interconnects. At that scale, the hardware has essentially one buyer class: frontier AI developers. The collateral value of the senior tranche on default therefore depends almost entirely on the health of the AI market for TPU-compatible workloads at that moment — a concentration that credit frameworks developed around GPU collateral may not yet fully price.

The CDO comparison, handled carefully

Commentators have reached for a shorthand: “collateralized chip obligations,” by analogy to the CDOs of 2008. The comparison is instructive but imprecise.

What is similar: debt is being structured in tranches against a pool of long-lived physical assets, with a single dominant borrower class, and the risk is correlated — if AI demand disappoints, lease payments, residual values, and resale markets would deteriorate together.

What is different: the assets are productive equipment with a single creditworthy counterparty under long-term lease, a deep-pocketed guarantor behind the senior tranche, and payments that begin only after deployment. Robert Leitao, managing partner at Rothschild & Co., put the concentration plainly to Reuters: “It feels that there’s quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that’s happened.”

Why It Matters

The AI boom is creating a new asset class around compute. Instead of one company carrying tens of billions of dollars of hardware on its balance sheet, banks and institutional investors finance the assets while the AI lab leases the resulting capacity — a structure borrowed from aircraft leasing, energy projects, and commercial real estate. This deal spreads the capital requirements across the financial system, but it also creates new dependencies between AI labs, chip suppliers, banks, and private-credit investors. If the syndication goes well, expect more of them; if it stalls, the industry’s most ambitious compute buildouts have a funding problem. Either way, the answer arrives in the bank market over the coming weeks — and in Anthropic’s IPO filings, where the counterparty risk finally becomes public information.