Meta is repositioning its long-term research agenda around what chief executive Mark Zuckerberg calls “personal superintelligence.” In an open letter released ahead of the company’s Q2 2025 earnings call, he presented the concept as the next stage of consumer computing, distinct from earlier metaverse ambitions and from industry efforts that frame artificial general intelligence chiefly as a workforce replacement tool.
Building on this vision, Zuckerberg argues that AI should serve as an individually tailored extension of human capability rather than a centralized system of mass automation. The letter emphasizes agency: in Meta’s view, superintelligence must adapt to personal goals, daily context, and creative intent. That positioning implicitly challenges the enterprise-first strategies pursued by OpenAI, Google, and others.
Delivering such real-time, context-aware models at a planetary scale will require formidable infrastructure. Meta has therefore expanded its AI budget by several billion dollars, recruited specialists from leading research labs, and established the new Superintelligence Labs division under former Scale AI chief Alexandr Wang. The group is charged with advancing Llama-class foundation models and exploring novel architectures optimized for low-latency inference.
To support these models, hardware is being overhauled in parallel. Internal sources indicate that custom accelerators now operate alongside Nvidia H100 and A100 GPUs in Meta’s data centers, while a next-generation Meta Training and Inference Accelerator (MTIA) is reportedly taped out for later this year. Discussions of in-house silicon echo Google’s TPU strategy and suggest future integration in head-worn devices, where power efficiency is critical.
Meta’s metaverse push has already consumed more than $60 billion in Reality Labs losses. Yet, AI has immediate market traction, and the company appears willing to redirect capital accordingly. Whether personalized superintelligence becomes a defining consumer platform will hinge on Meta’s ability to scale both algorithms and custom hardware before rivals lock in their own ecosystems.
Source(s)
Meta (in English)