At the core of any large-scale AI infrastructure buildout is a resource allocation problem that does not resolve itself. Once a hyperscaler has provisioned GPU clusters and the networking fabric around them, those assets must be directed somewhere. The choice is binary: keep capacity internal to run proprietary models and products, or sell access to external customers and recover capital faster. Mark Zuckerberg put this tension on the table in remarks addressed to investors who have been pressing Meta for clarity on how the company's large AI spending eventually pays back.
The constraint Zuckerberg named
The sell-versus-keep question is not cosmetic. Compute sold externally generates revenue against a capital expenditure that has already been made. But that same compute, if retained, provides the throughput headroom that internal AI products depend on. Running inference at scale for Meta's user base and selling capacity to outside tenants draws from the same physical substrate. Those two uses compete. Zuckerberg's reported framing treats this as a genuine trade-off rather than a phased plan where both sides are eventually satisfied.
The investor backdrop
Meta's investor base has been waiting for a monetization answer to accompany the company's AI investment program. The spending has been large enough to draw sustained attention from shareholders, and Zuckerberg's comments were directed at that audience. What the reported remarks establish is the shape of the problem: a competition for the same compute between internal use and external sale, with no disclosed resolution as to which side captures how much.