Meta is spending billions to build a "superintelligence" unit and is aggressively hiring top AI researchers with massive compensation packages. In just a month, Apple has lost four core AI team members to Meta, but the new startup Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has so far resisted Meta's offers.
The latest to jump ship is Bowen Zhang, a leading multimodal AI researcher, who is leaving Apple for Meta's superintelligence team. He joins Ruoming Pang, the former head of Apple's Foundation Models group, along with Tom Gunter and Mark Lee. That brings the tally to four key Apple AI researchers who have switched to Meta in just a few weeks.
Meta CEO Mark Zuckerberg has broadened his recruiting beyond Apple, targeting Google DeepMind, OpenAI, and Safe Superintelligence, and is now pursuing talent from Thinking Machines Lab. According to WIRED, Meta has made offers to more than a dozen of Murati's roughly 50 employees, including at least one compensation package worth over $1 billion across several years. So far, no one at the company has accepted.
Meta's superintelligence group runs like a startup
The new hires are joining Meta's superintelligence unit, which Zuckerberg wants to operate like a startup inside the company. The group has its own building at Meta's Menlo Park campus, works in isolation from other teams, and has access to all of Meta's resources. Alexandr Wang, co-founder of Scale AI, and Nat Friedman, former CEO of GitHub, are leading the effort and working closely with Zuckerberg.
Even inside Meta, the project's actual strategy is kept under wraps. One insider described it as a high-stakes "Manhattan Project" for AI, with huge spending and little visibility into what the outcome will be.
The rise of the superintelligence group is also creating tension within Meta. Some employees, especially in generative AI, worry about being sidelined or replaced. Internally, there's a sense that Zuckerberg is reshuffling teams to accelerate the company's AI efforts.
Meta is making open source a central part of its strategy, aiming to turn core AI infrastructure into a commodity and challenge closed systems like ChatGPT. Behind the scenes, however, things have not gone smoothly. The company's Llama models have underwhelmed. Llama 4 was released too soon, and the larger Behemoth model remains on hold after failing to meet quality standards. The Financial Times reports that Meta now lets internal teams use outside models, which is unusual for a company aiming to lead in AI. As one ex-employee put it, speed is the goal, but Meta's own models are the bottleneck.
Investors watch for another Metaverse scenario
Despite the uncertainty, Meta's stock is up 20 percent since the start of the year, signaling investor confidence in Zuckerberg's AI push. Still, analysts caution that costs could spiral. BNP Paribas puts Meta's extra research spending on superintelligence between $1.5 and $3.5 billion per year, and in April, Meta raised its annual investment plan to as much as $72 billion.
Brent Thill from Jefferies sees the hiring spree as a sign of urgency, but warns it could squeeze profits. Without a clear commercial plan, some analysts expect investors may lose patience.
Zuckerberg has faced this kind of scrutiny before. He rebranded the company as Meta in 2021 and invested heavily in the "Metaverse," but the virtual world never lived up to expectations, and Meta's stock hit a six-year low. With the new push for superintelligence, Zuckerberg is trying to avoid repeating the same mistakes. Whether his all-star team can actually deliver remains to be seen.
