Meta’s superintelligence push gains compute and data momentum

Meta’s AI reset has moved from recruiting splash to operational buildout, with the company leaning on data, talent and compute to close the gap with OpenAI and Anthropic. The effort followed the Llama 4 release and included a $14.3B Scale AI investment tied to recruiting Alexandr Wang and members of the SEAL team, as well as large compensation offers to AI researchers and engineers.

A central advantage is emerging around reinforcement learning data. Meta has begun tracking employee screens, keyboards and mouse movements, while adding privacy protections and an option to pause the tracker for 30min. A new applied AI engineering org will put ~3000 engineers, including 70% of new grads and many senior staff, on full-time RL task and environment creation.

The compute ramp is similarly aggressive. Meta is building 5 1GW+ titan clusters, including Prometheus in Ohio, Hyperion in Louisiana, and sites in El Paso, Iowa and Indiana. Hyperion has 1.5GW under construction, while Prometheus is described as a constellation of 27 datacenters across 6 campuses, supported by an AI-Backbone network designed to move training workloads across large clusters.

MSL’s Muse Spark debut lagged rival open-source models on benchmarks, and Muse Spark 1.1 is described as short of Anthropic and OpenAI for current internal use. The long-term case depends on whether Meta keeps funding compute, retains top researchers and sustains its new RL data operation.