In a dramatic strategic pivot, Meta is launching a new research lab dedicated to building “superintelligence” and is in talks to invest billions in data infrastructure firm Scale AI, whose founder and CEO, Alexandr Wang, will join the new initiative. The move, confirmed in a report by The New York Times, is part of a sweeping reorganization of the company’s artificial intelligence division as it scrambles to counter significant internal turmoil, a high-profile talent exodus, and mounting competitive pressure.
This aggressive play signals a fundamental shift for Meta, which has been grappling with product development delays and struggles to retain the core researchers behind its flagship AI projects. By bringing in Wang, a prominent figure in the AI infrastructure space, and, as the Times also reported, potentially acquiring a major stake in his company, Meta is making a costly bet that it can buy and build its way back to the forefront of the AI race.
The plan represents a significant escalation of a corporate restructuring that began in late May. At that time, Meta reorganized its AI division into “AI Products” and “AGI Foundations” teams. According to an internal memo from Chief Product Officer Chris Cox, that move was designed to increase team ownership and clarify dependencies. However, the creation of a new, high-profile lab focused on the still-theoretical goal of artificial general intelligence (AGI) suggests the initial changes were not enough to address the company’s deeper challenges.
The Price of AI Supremacy
Meta’s decision to potentially pour over $10 billion into Scale AI underscores the immense and rapidly escalating financial strain of the AI arms race. The deal would be a rare external investment for Meta and would likely be a pure cash transaction, as the company lacks a cloud platform to offer credits like its rivals. Malik Ahmed Khan, an analyst at Morningstar, estimates that “A $10 billion investment, not counting Meta’s existing stake in Scale AI, would likely yield a post-money stake of loosely 10%.”
This spending is set against a backdrop of exponentially rising costs. Research from Epoch AI shows that the expense of training frontier models has been growing by a factor of two to three times per year, with the largest projects projected to cost over a billion dollars by 2027. This financial reality marks a stark evolution from Meta’s previous strategy.
Last year, Meta had pitched a “Llama Consortium” to competitors like Amazon and Microsoft, seeking financial help to train its models. That lukewarm reception has evidently pushed Meta toward a new strategy: outright acquisition of critical infrastructure to secure its supply chain and compete with the massive AI investments made by its Big Tech rivals.
An Exodus of Architects
The reorganization is a direct response to a severe talent drain that has plagued Meta’s AI efforts. The company has lost 11 of the 14 original authors of the foundational Llama paper, with many departing for competitors like Paris-based Mistral AI.
Only three of the original architects curretnly remain at the company. This exodus was compounded by the high-profile exit of Joëlle Pineau, who led Meta’s Fundamental AI Research (FAIR) group for nearly eight years.
To stanch the bleeding, Meta has been on an aggressive hiring offensive, offering seven- to nine-figure compensation packages to top researchers from Google and OpenAI. The pursuit of Scale AI appears to be the capstone of this strategy. The deal is not just about technology but about acquiring a proven leader in Alexandr Wang and his highly educated workforce, which, according to Morningstar, includes a significant number of PhDs and graduate degree holders specializing in complex data.
This talent infusion is seen as critical for Meta to catch up in areas where it has lagged. The move follows an earlier, less drastic internal reshuffle in May, signaling that Meta’s leadership believes a more fundamental overhaul is necessary to rebuild its research foundation.
Code, Copyright, and Credibility
Compounding the talent issue are significant technical and legal headwinds. The company’s most ambitious model, the 2-trillion parameter Llama 4 “Behemoth,” has been significantly postponed until late 2025 or beyond due to development struggles. The delay, which has reportedly caused mounting frustration within the company, was characterized by Deepwater Asset Management’s Gene Munster with the words “It’s a black eye but doesn’t change their opportunity.”
Simultaneously, Meta is also fighting a significant copyright case from authors like Sarah Silverman over its use of pirated books to train its Llama models. The lawsuit poses an existential threat to its data practices, with Judge Vince Chhabria expressing deep skepticism of the company’s “fair use” defense.
During a hearing, the judge cut to the heart of the legal debate, questioning how Meta’s actions could be considered “fair use” if its model was “dramatically changing, you might even say obliterating, the market for that person’s work” without the company paying for a license.
These unresolved legal and technical challenges have created a turbulent environment for Meta’s AI ambitions. The immense financial pressure of the AI race, the internal chaos of talent loss, and the persistent risks from courts and its own development pipeline appear to have forced the company’s hand.
The creation of a new superintelligence lab and the landmark deal for Scale AI are therefore not just offensive moves to seize the future, but a high-risk, high-reward defensive strategy to fix fundamental problems, secure its supply chain, and reclaim lost momentum in a rapidly evolving technological landscape.
