구글은 테라울프의 AI 데이터센터 개발을 지원하기 위해 32억 달러의 백스톱 시설을 제공했으며, 이는 새로운 컴퓨팅 캠퍼스를 뒷받침하는 자금 조달 모델을 보여준다.

Google has guaranteed $3.2 billion in lease payments for an artificial intelligence data center it does not own, constructed by TeraWulf—a company formerly known for bitcoin mining. This arrangement is not an act of corporate generosity; it is financial engineering, and it represents the foundational model for the current AI infrastructure buildout.

The transaction centers on Maryland-based TeraWulf, which has spent the last two years converting a decommissioned coal plant site on Lake Mariner in western New York, near Buffalo, into an AI data center campus. According to CNBC and Carbon Credits, Google agreed to backstop $1.4 billion in new lease obligations atop existing commitments, bringing its total financial support to $3.2 billion. In return, Google’s warrants now confer approximately 14 percent ownership in TeraWulf, an increase from just 8 percent several months ago.

Google is not the tenant; Fluidstack is. The AI cloud provider executed two 10-year agreements with TeraWulf for more than 200 megawatts of critical IT load at Lake Mariner. The footprint was later expanded with a 160-megawatt facility designated CB-5, raising Fluidstack’s total contracted capacity at the site to 360 megawatts. Data Center Dynamics values the combined lease portfolio at $3.7 billion across the initial term, a figure that climbs to $8.7 billion if Fluidstack exercises its extension options.

So why must Google guarantee another company’s lease? Because TeraWulf cannot secure the billions required to construct CB-5 and subsequent phases on its own credit. Traditional lenders remain skeptical of a former cryptocurrency miner’s balance sheet but readily accept Google’s. The backstop enables TeraWulf to raise project-level debt backed by Fluidstack’s lease payments, with Google’s guarantee serving as the underlying insurance. The structure delivers warrants and a growing equity position to Google, a construction loan to TeraWulf, and a completed facility to Fluidstack—all without requiring Google to place any data center assets on its own balance sheet.

TeraWulf’s financing structure is modest compared to Meta’s recent development in Richland Parish, Louisiana. Meta’s Hyperion campus, a four-million-square-foot site projected to draw up to five gigawatts upon completion in 2029, is being financed through a joint venture with funds managed by Blue Owl Capital. Blue Owl holds an 80 percent stake in the venture, while Meta retains 20 percent. Morgan Stanley structured $27 billion in debt and $2.5 billion in equity through a special purpose vehicle, with Pimco anchoring the bond issuance. The debt carries an A+ rating from S&P Global Ratings and does not mature until 2049.

Meta restructured its $50 billion Richland Parish data center through a Delaware-registered holding company on the precise day state regulators approved the natural gas plants intended to power the facility. The arrangement, which allocated 80 percent of the joint venture to Blue Owl Capital and 20 percent to Meta, was never disclosed to regulators.

Meta will occupy Hyperion and pay rent for decades. Because Meta does not control the entity holding the debt, the liabilities never appear on its balance sheet. This accounting mechanism underpins nearly every major gigawatt-scale AI campus announced this year: a special purpose vehicle originates the debt, a private capital manager such as Blue Owl maintains majority ownership, and the hyperscaler signs a long-term lease classified as an operating expense rather than a corporate liability.

The magnitude of this off-balance-sheet restructuring has ceased to be a marginal accounting adjustment. A study cited by Forbes indicates that the five largest U.S. hyperscalers have relocated more than $1.6 trillion in AI infrastructure debt into financial statement footnotes instead of their primary balance sheets. Separately, Moody’s estimates that approximately $662 billion in hyperscaler lease commitments now reside entirely outside reported debt, with more than $120 billion of that shift occurring in less than two years.

Investors must recognize that the special purpose vehicle structure does not eliminate risk; it merely relocates it. Should Fluidstack terminate its agreement at Lake Mariner tomorrow, TeraWulf would inherit a partially constructed campus burdened with debt calibrated for a departed tenant, leaving Google’s backstop as the sole barrier between that debt and default. In the Meta framework, bondholders backed by Blue Owl and Pimco—not Meta shareholders—would absorb the initial losses if Meta were to walk away from Hyperion. Meta’s capped 20 percent equity stake structurally limits its downside exposure.

This asymmetry is the explicit objective of the financing model. The hyperscaler secures the necessary compute capacity, preserves a clean balance sheet, and transfers tail risk to a private credit fund or warrant-holding partner that has bet on sustained AI demand over a decade or longer. If that demand materializes, all parties profit. If it fails to materialize, the first claims will not target Sundar Pichai or Mark Zuckerberg; they will target the special purpose vehicles that operate entirely out of public view.