Kai ChenMeta’s announced Alberta data center puts a roughly $9 billion price on 1GW of capacity that will take two to three years to build.
The balance-sheet boundary moved before the gigawatt arrived
Since 2022, Meta has raised $62 billion of debt, with roughly half raised in 2025. Separately, the company moved $30 billion of debt for building AI data centers off its balance sheet through special-purpose vehicles.
Those figures should not be added as though they describe two independent piles of money. Together, they describe a financing architecture: more corporate debt, accompanied by a structure that places a large portion of AI infrastructure financing outside the parent balance sheet.
The SPV is the mechanism. It creates a perimeter around a project, allowing its debt and ownership to be structured separately from the operating company. An SPV does not make the servers cheaper. It makes an enormous buildout less dependent on keeping every dollar of project debt in the same corporate container.
That distinction turns infrastructure from an internal capital requirement into a financeable asset. Meta still needs the capacity, but it no longer needs to organize every project as a direct extension of the parent company’s balance sheet.
A gigawatt changes the object being financed
Meta’s announced Alberta data center is designed for 1GW of capacity and an estimated cost of roughly $9 billion. The project will take two to three years to build, so the announced gigawatt is not near-term operating compute. It is planned capacity with a multiyear construction horizon.
That delay is not a footnote to the financing story. It is the financing story. Once a data center is measured in gigawatts, billions of dollars, and years of construction, the relevant object is no longer a server purchase. It is a utility-scale capital project: money is committed before the capacity becomes productive.
Meta has not disclosed that Alberta uses the same SPV structure as its other AI data-center financing. The structural pairing is broader: debt and off-balance-sheet vehicles have expanded Meta’s financing toolkit, while Alberta expands the geographic and physical scale of what that toolkit may support.
Hyperion is the stronger test and the weaker fact
Hyperion would make the asset-class logic more explicit. Sources report that Meta is set to secure an almost $30 billion financing package for the Hyperion data center in Richland Parish, Louisiana, while retaining a 20% ownership stake.
Neither the financing package nor the ownership stake has been confirmed.
If completed as reported, the structure would separate Meta’s need for compute from full ownership of the asset supplying it. That would be the cleanest expression of infrastructure finance: retain a stake, secure the capacity, and place most of the project inside a distinct capital structure.
But the broader case does not depend on Hyperion closing. Meta has already raised $62 billion of debt since 2022 and moved $30 billion of AI data-center debt into SPVs. Hyperion would extend the pattern; it did not create it.
The reusable block is now capital structure
For Meta, the reusable block is no longer only compute. It is the financing wrapper around compute: corporate debt, project debt, SPVs, and potentially partial ownership can be arranged around a discrete site.
The cost of AI infrastructure has not collapsed. A roughly $9 billion Alberta project and a reported nearly $30 billion Hyperion package show why these physical systems no longer fit comfortably as ordinary line items. What has eased is the constraint that every buildout must be financed and owned in the same way.
The strategic asset is access to the gigawatt, not necessarily ownership of every dollar beneath it.
That changes the unit of expansion. The question is no longer only whether an operating budget can absorb more servers. It is also whether a project can be given its own financing perimeter, ownership structure, and construction horizon.
The Alberta project’s two-to-three-year gap between roughly $9 billion of capital and 1GW of usable compute is not incidental to the asset; it defines it. For Meta, the gigawatt has become a unit of finance before it becomes a unit of compute.