Meta is building a cloud business to sell its spare AI computing power to outside companies. The plan, first reported by Bloomberg and confirmed by CNBC, would let businesses pay to run AI workloads on Meta's own data centers, or simply rent raw computing capacity the same way they might from a cloud provider.
This is not a small side project. Meta spent $72 billion on AI-related infrastructure in 2025. The company has planned a further $125 to $145 billion in 2026. When you spend at that scale, the pressure from investors to show a return becomes intense. Selling spare capacity is one logical answer, and Mark Zuckerberg had already signaled in May that it was, in his words, "definitely on the table."
The internal unit driving the effort is called Meta Compute, led by Meta's head of infrastructure alongside a leader from its AI research division and Meta's president. One option under discussion mirrors Amazon's cloud model: letting developers pay to run queries against AI models, including Meta's own recently launched Muse Spark model, all hosted on Meta's own servers. A second option is simpler: rent out raw computing capacity, the same model used by specialist providers like CoreWeave.
That second option is what rattled the market. CoreWeave dropped roughly 14% and Nebius dropped roughly 17% on the day the news broke. The concern is not just new competition. Meta is one of their largest customers: CoreWeave holds a $21 billion commitment from Meta, and Nebius holds an agreement worth up to $27 billion. A Meta that builds enough capacity to sell the excess is a Meta that may eventually need to buy less from them.
The broader computing shortage, however, is real and not going away. Data center vacancy rates in North America have hit a record low of around 1%, and 92% of capacity currently under construction is already pre-leased. GPU orders for Nvidia chips now stretch nearly a year in advance. On-demand computing capacity is effectively sold out. Meta entering the market as a seller adds supply to a market that has been badly short of it, which is genuinely useful for any organization trying to run AI projects.
For most organizations outside the tech industry, the real signal here is a shift in what the AI challenge actually is. For the past three years, the central problem was simply getting access to enough computing power. That problem is not solved, but it is slowly easing at the edges. The next problem, already arriving for many organizations, is what to do with the access once you have it.
Organizations that have spent the last two years building proper data foundations, clear operating models, and real governance over their AI outputs are positioned to take advantage of cheaper, more available computing power. Those that have not done that groundwork will find that cheaper access mostly means a cheaper way to generate outputs nobody can fully trust or verify.
There is a pattern here that repeats across technology history. When electricity became widely available, the companies that won were not the ones with the most power: they were the ones that figured out what to do with it. Cheap computing power follows the same logic. The organizations that will benefit most are the ones that already know which of their processes are worth automating, which AI outputs need human review, and which projects should stay as pilots rather than scaling into production.
Meta's move is a useful early signal that the AI infrastructure market is maturing. It does not mean the shortage is over. It means the shortage is beginning to be a solvable problem for large enough players, and that over time, the real differentiator will shift from who has computing access to who knows how to use it with discipline.