The $35B Blueprint: How Private Credit Became AI Infrastructure's Dominant Financing Mechanism in 2026

Six weeks ago, the prevailing view in AI finance was that equity was the instrument of the moment. Anthropic had just closed a $65 billion Series H. The venture funds were still writing the largest checks in the sector’s history. Then, almost without announcement, the financing model changed.

Apollo Global Management and Blackstone finalized a $35 billion private credit package on June 6 to fund Anthropic’s purchase of Google custom Tensor Processing Units, one of the largest private credit transactions ever structured. The deal doesn’t fit the standard AI funding narrative. There’s no lead venture partner. There’s no valuation headline. The press release isn’t from Anthropic. It’s a debt instrument, arranged by Wall Street’s two largest alternative asset managers, secured against physical hardware, and routed through a special-purpose vehicle.

That structure deserves careful attention.

How the SPV Model Works

According to reports citing Bloomberg, the deal uses a special-purpose vehicle that borrows the funds, receives an equity investment, purchases the Google TPUs, and leases them to Anthropic. Anthropic gets compute capacity. The debt obligation lives inside the SPV, not on Anthropic’s direct balance sheet. Broadcom reportedly provides a residual-value backstop on the senior notes, a guarantee structure that makes the senior tranche attractive to institutional investors who’d otherwise find the credit profile of a pre-revenue AI company difficult to underwrite. The specific dollar amount of Broadcom’s backstop position couldn’t be independently confirmed in accessed sources and should be treated as reported, not verified.

The mechanics matter because they’re replicable. Every component of this structure, SPV borrowing, hardware as collateral, corporate backstop on senior tranche, institutional syndication, is standard private credit architecture. It’s been used in aircraft leasing, real estate, and data center financing for decades. What’s new is the asset class: AI chips, specifically Google TPUs, structured as leaseable physical infrastructure with a predictable depreciation profile that institutional lenders can underwrite.

The Precedent Pattern

This deal doesn’t stand alone. CoreWeave closed a $3.1B GPU loan in late May, a more straightforward secured debt instrument, but one that established the principle that GPU capacity is bondable collateral. Alphabet executed a large-scale capital raise for infrastructure around the same period. Earlier, Crusoe and Bergen structured a 750MW on-site power deal that treated power capacity as a financeable asset rather than an operating cost.

The through-line across all four transactions: AI infrastructure components, chips, power, physical facilities, are being reclassified from operating expenses into capital assets that can be financed, collateralized, and syndicated. That’s an accounting and structural shift as much as a financing one.

Three infrastructure debt transactions in 30 days, from separate arrangers, targeting separate asset types (GPUs, TPUs, power), with separate borrowers. When the same financing model appears independently across three transactions in the same sector within the same month, it’s not a trend. It’s a market consensus forming.

What This Requires of Institutional Investors

Three distinct investor audiences face different decisions from this transaction.

Institutional investors in the debt tranches are evaluating a new collateral type. GPU and TPU resale markets are thin and concentrated. If Anthropic’s business deteriorates and the SPV needs to liquidate the TPU fleet, the recovery value depends entirely on Google (the manufacturer) and a handful of hyperscalers as potential buyers. Broadcom’s backstop on the senior notes partially mitigates this, it effectively inserts Broadcom’s creditworthiness as a buffer between the debt investors and the TPU resale market. But the backstop terms and duration aren’t publicly disclosed, and that gap matters for anyone modeling recovery scenarios.

Equity investors in Anthropic, pre-IPO, face a different question. The off-balance-sheet structure protects the income statement and balance sheet metrics that public market investors will scrutinize. But it doesn’t eliminate the dependency. Anthropic’s access to its core compute infrastructure depends on maintaining the SPV leasing arrangement and, indirectly, on Broadcom’s continued participation in the backstop structure. If that arrangement were disrupted, through credit stress, counterparty dispute, or restructuring, Anthropic’s operational continuity would be directly at risk. Public market investors will need to understand this dependency, and the S-1 will need to disclose it.

Enterprise customers are the third audience. Anthropic’s infrastructure is now concentrated in a specific chip architecture (Google TPU), financed through a single SPV arrangement, backstopped by a single semiconductor company. Counterparty concentration risk isn’t a theoretical concern here, it’s a documented structural feature of this deal. Organizations running mission-critical workloads on Claude should understand that their inference capacity ultimately depends on the stability of a three-party arrangement between Anthropic, Apollo/Blackstone, and Broadcom.

The Open Questions

Several things about this transaction aren’t publicly confirmed. The tranche structure, reportedly three tranches, including a reported $6 billion A1 senior notes tranche with approximately half syndicated to outside investors, didn’t appear in the accessed source excerpts for this pipeline and should be treated as reported detail, not verified fact. Pricing (interest rates and spread) is undisclosed. Broadcom’s backstop terms and duration aren’t public.

What to watch for: Whether OpenAI or xAI adopt comparable SPV structures for their own compute buildouts. Both companies face similar balance-sheet pressures as they approach public markets, and both have GPU fleets that could theoretically be restructured into leaseable assets. If the Apollo/Blackstone model becomes the standard template, the arrangers with the relationships and structuring expertise, primarily the large alternative asset managers, gain significant recurring economics from every frontier lab compute buildout. That’s a durable revenue stream that would justify further expansion into AI infrastructure debt.

Also watch for any SEC comment letters on Anthropic’s S-1 that address the SPV structure. If the SEC requires Anthropic to consolidate the SPV onto its balance sheet, which is possible under certain accounting interpretations, the off-balance-sheet benefit disappears and the debt becomes directly visible to IPO investors.

The TJS Synthesis

The venture capital industry funded the model development phase of the AI buildout. Private credit is funding the infrastructure phase. These are structurally different risk profiles for different investor types, and they require different analytical frameworks. VC investors bet on team and model capability. Private credit investors bet on hardware residual value and counterparty creditworthiness. The presence of Broadcom as backstop provider and Apollo/Blackstone as arrangers signals that Wall Street has developed enough conviction in the AI infrastructure thesis to put debt capital at risk, not just equity.

The test for this model comes when the first major SPV needs to restructure. Until then, expect the template to spread. Watch Q3 2026 for the next infrastructure debt announcement from a frontier lab, if OpenAI or xAI close a comparable deal before September, the SPV model will have achieved sector-wide adoption in under six months.