Meta Plans Meta Compute Cloud Service to Sell Excess AI Capacity

Meta plans to launch a cloud service called Meta Compute that will sell excess AI computing capacity and model access to external customers.

The move puts the company in direct competition with AWS, Google Cloud, and Azure. Meta has already committed 182.9 billion dollars to AI infrastructure over the next several years.

One data center in Ohio, sized at the scale of Manhattan, is scheduled to come online this year. The new business will be led by infrastructure head Santosh Janardhan, Superintelligence Labs lead Daniel Gross, and president Dina Powell McCormick.

Meta chief executive Mark Zuckerberg has previously described a cloud offering as "definitely on the table," according to a Bloomberg report. The company now appears ready to act.

The plan mirrors how SpaceX reuses launch infrastructure and capacity across missions rather than letting assets sit idle. Meta wants similar efficiency from its large AI clusters.

Meta has already trained and released several large models. Some of the newest ones remain closed source, including the Muse Spark family. The company may host these models inside Meta Compute in the same way AWS hosts third-party models today.

It could also follow the CoreWeave model and offer raw compute capacity without managed software layers. Customers would then bring their own code or frameworks.

The shift matters because Meta built its AI clusters for internal research and product needs first. External sales represent a new revenue stream and a way to raise utilization rates on hardware that would otherwise remain underused between training runs.

Pressure now falls on the three major cloud providers. They have long dominated rented GPU capacity. A well-funded new entrant backed by its own large training clusters could change pricing dynamics and contract terms for AI workloads.

Meta faces its own set of trade-offs. Running a multi-tenant cloud requires new skills in billing, security, and customer support. TechCrunch noted the company has limited public experience in these areas at the scale of AWS. Beyond those basics, Meta would need robust solutions for dynamic GPU allocation and workload isolation to prevent noisy-neighbor interference, similar to the orchestration frameworks Kubernetes-based clusters at CoreWeave or the tenant separation layers Google Cloud deployed for Vertex AI. Failure here risks latency spikes or data leakage across customers sharing the same silicon.

Analysts note that utilization rates on large GPU clusters often drop below 50 percent between major training jobs. Selling spare cycles could lift that number and improve capital efficiency, but only if demand materializes quickly.

Independent observers point out that Meta still needs to finish and stabilize its own model roadmap. Hosting customer traffic on the same clusters used for internal research could create scheduling and priority conflicts during peak periods.

The next signals to watch include the exact pricing tiers Meta publishes, any partnership announcements with existing model providers, and early customer adoption numbers once the Ohio site reaches full operation.