Kai ChenMeta moved $30 billion of debt used to build AI data centers into special-purpose vehicles. Now Anthropic may become a customer of Meta’s infrastructure. The contradiction starts below the models, in the financing beneath them.
Infrastructure-scale finance changes the rational product
Meta’s debt issuance accelerated in 2025, when it raised roughly half of the total issued since 2022.
The $62 billion debt tally and $30 billion SPV shift may overlap, so adding them into a theatrical $92 billion would misstate the scale. Through the SPVs, Meta moved AI data-center debt off its balance sheet. Together, the figures describe the same structural move: compute is being financed less like an internal software expense and more like long-lived infrastructure.
That changes the incentive. When compute is scarce and acquired project by project, reserving it for internal models makes sense. When capacity is financed at data-center scale, the economic question becomes broader: which workloads can earn a return on the installed base?
Internal models and ads remain answers. They are no longer necessarily the only answers.
Third-party hosting gives compute a second identity
The July 4 report explicitly frames Meta’s compute as usable for neocloud-style deals and third-party model hosting. It also says Meta may be near an arrangement involving Anthropic. The verbs matter: Meta could host outside models, and a deal may be near. Meta has not launched a public cloud, and the report does not establish a completed Anthropic contract.
But the option itself marks a structural boundary. Compute built to strengthen Meta’s own AI position is being considered as infrastructure for a model provider outside Meta. Under that architecture, Meta can derive value from model demand without requiring that demand to flow through a Meta model.
That is platformization. The asset stops being valuable only for what its owner builds on top of it. It becomes valuable because other builders can bring workloads to it.
Meta does not have to win every model choice to monetize the demand behind it.
Manus makes the boundary porous from above
Meta’s acquisition of Manus pushes the same shift from a different layer. Manus talent is joining Meta to help deliver agents across Meta products, including Meta AI. Yet Meta also plans to continue operating and selling the Manus service after the deal.
That is not proof of a cloud business. Manus is an agent service, not a public infrastructure launch, and suggestions that the acquisition could support cloud offerings remain speculative. Still, the operating choice matters: Meta is integrating the builders while preserving an external route to market for what they built.
Put that beside possible third-party model hosting and the architecture becomes clearer. At the application layer, Meta can own a service that remains available outside its existing products. At the infrastructure layer, it can potentially serve models it does not own. Different layers, same incentive: assets assembled for internal strategic advantage gain another revenue path when they remain accessible to outside demand.
Platformization starts before a cloud launch
The counter-evidence is straightforward. Meta’s compute can still primarily serve its own models and advertising business. The July 4 report establishes neither a general-purpose public cloud nor a completed Anthropic arrangement. Hosting one outside provider would not, by itself, make Meta a full peer to established cloud companies.
Platformization begins earlier. It starts when the owner’s resource-allocation logic changes from “capacity for our products” to “capacity that can also carry external workloads.” A public launch is the announcement layer. The hosting option is the structural layer.
This distinction also explains why the financing matters more than branding. Special-purpose vehicles do not make servers more capable. They make infrastructure-scale construction financially workable while preserving pressure to find productive uses for what gets built. The capital structure does not demand a “Meta Cloud” logo. It rewards utilization. Logos, as usual, arrive after the incentive has completed the paperwork.
Debt makes outside demand rational
Meta need not become neutral about models to monetize outside demand. It only needs to make some capacity available on terms an external provider will accept. That is narrower than launching a cloud but deeper than signing a hosting deal: infrastructure economics can reward workloads that strategic competition would otherwise keep outside Meta’s walls.
The $30 billion SPV shift made Meta’s AI buildout look like infrastructure finance. Once an outside model can become a customer, it starts to look like infrastructure in the commercial sense too. The data center wants customers because infrastructure-scale finance makes them rational.