A year after Meta overhauled its AI division following the troubled Llama 4 launch, Max Kan and his co-authors write for SemiAnalysis that Meta Superintelligence Labs (MSL) may be closing the gap on OpenAI and Anthropic. The authors argue Meta is the only company positioned to excel at all three ingredients of frontier AI: data, talent and compute.
On data, Meta has started recording employees’ screens, keyboards and mouse movements to generate reinforcement learning material. Such recordings capture realistic workflows, which the authors say makes them more valuable than the artificial, over-engineered tasks used in many industry benchmarks. Meta has also created a roughly 3,000-person “applied AI engineering org” dedicated to building these training environments, drawing on staff across its many in-house business functions.
On compute, Meta is simultaneously constructing five gigawatt-scale “titan” data centers, including Prometheus in Ohio and Hyperion in Louisiana. According to SemiAnalysis projections, Meta will have more AI computing capacity than both OpenAI and Anthropic by the end of the year. To link these massive sites, Meta has built a new networking system called AI-Backbone.
On talent, Meta has continued poaching researchers from OpenAI, Anthropic and the shuttered Thinking Machines, alongside earlier high-profile hires like Alexandr Wang.
Still, the authors caution that Meta’s first public model, Muse Spark, trails competitors, and true success remains far from guaranteed.
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
