Zuckerberg's trade-off between selling excess capacity and reserving it for internal AI work will shape Meta's next infrastructure cycle.
Mark Zuckerberg has been direct about a problem most hyperscalers prefer to obscure: when you build enough compute to win an AI race, you end up with more capacity than you can use at any given moment, and that surplus is either an asset or a liability depending on what you do next.
Meta is now weighing whether to lease that excess compute to external clients or hold it in reserve for its own AI projects. Zuckerberg has outlined this trade-off publicly, framing it as an unresolved strategic dilemma, according to CNBC.
The cost of building ahead
The dilemma exists because of how AI infrastructure investment works. Companies must commit capital years before they know exactly what workloads they'll run. Meta has been spending at a scale that makes the choice consequential, managing what Zuckerberg described as massive AI capital expenditures.
Building ahead of internal demand creates optionality for future model training and iteration. It also creates idle capacity in the near term, and idle compute is expensive compute.
Leasing that capacity to outside clients converts a sunk cost into revenue. The risk is that external commitments reduce flexibility precisely when Meta might need to surge its usage, during a major model training run or a rapid product deployment.
What selling compute actually means
Leasing compute to third parties would put Meta in more direct competition with Amazon Web Services, Google Cloud, and Microsoft Azure, all of which have built substantial businesses around exactly this model. Meta's infrastructure has been optimized for its internal workloads, which may not translate cleanly into a general-purpose cloud offering.
The internal-use case is also not static. Meta's AI ambitions span its core social platforms, its standalone AI assistant, and longer-horizon research. Each has different compute profiles and urgency levels, which makes it harder to define a stable floor of capacity that could safely be leased out without creating internal bottlenecks.
The company has not resolved this tension; it is actively managing it. That is a different posture than announcing a cloud business or committing to keep everything internal.
The monetization pressure underneath
Meta's AI spending has faced scrutiny from investors who want returns on infrastructure investment. Leasing compute is one of the more legible ways to demonstrate that the capital expenditure is productive, even before internal AI products generate direct revenue at scale.
That pressure cuts both ways. Committing capacity to external clients on multi-year contracts could constrain Meta's ability to respond to shifts in AI competition, and conditions have been shifting faster than most infrastructure contracts are written.
The near-term marker will be whether Meta makes any formal announcement about external compute availability. Zuckerberg's public framing of the dilemma suggests the decision is live, and Meta's next earnings call or infrastructure announcement will show which direction the company has moved.
Get weekly AI video rankings, evaluation updates, and industry news delivered to your inbox.