How China’s DeepSeek made Mark Zuckerberg lose faith in Meta’s AI team, and start the biggest-ever talent war in Silicon Valley |

Mark Zuckerberg is preparing to defend billions in AI spending as Meta faces its slowest profit growth in two years, with the CEO creating a secretive “Superintelligence lab” after the company’s latest AI model reportedly failed to meet expectations.Meta’s Llama 4 AI model attempted to copy techniques from Chinese startup DeepSeek but received such poor reception from developers, making Mark Zuckerberg authorise a complete overhaul of the company’s AI strategy, according to sources cited by CNBC and the Financial Times.Meta invested $14.3 billion in Scale AI, bringing CEO Alexandr Wang aboard as chief AI officer alongside former GitHub CEO Nat Friedman, ex-Safe Superintelligence CEO Daniel Gross, and ChatGPT co-creator Shengjia Zhao to lead the new Meta Superintelligence Labs. The unit operates in isolation at Meta’s Menlo Park headquarters, with Zuckerberg personally meeting new hires as they join the closely guarded project.The catalyst for this dramatic pivot came in January when DeepSeek’s R1 model caught Meta off guard. The Chinese lab’s mixture-of-experts (MoE) approach appeared cheaper to train and run than traditional dense models, prompting Meta executives to believe they had found a shortcut to leapfrog rivals like OpenAI, sources told CNBC.However, Meta’s rush to adopt MoE architecture backfired spectacularly. Developers found Llama 4 more difficult to customize and integrate compared to its predecessor, with many preferring the older Llama 3 model. The disappointing reception led Zuckerberg to lose confidence in his generative AI team’s leadership, particularly after controversies over whether Meta had manipulated industry benchmark tests.

Mark Zuckerberg’s Superintelligence Lab is Meta’s Manhattan Project-equivalent, and he’s on a spending spree for the best talent for it

The new Superintelligence Lab represents Mark Zuckerberg’s attempt to create a startup-like environment within Meta’s $1.8 trillion structure. “It’s like the Manhattan Project. They are throwing all their cash at AI and trying to work out what to do,” Uday Cheruvu, portfolio manager at Harding Loevner, told the Financial Times.The hiring blitz began as early as March, when Zuckerberg started drawing up a list of top AI researchers before embarking on his offensive. He has since wielded a $14.8 billion investment in Scale AI that brought CEO Alexandr Wang and nine-figure sign-on bonuses to poach approximately 50 top AI researchers from competitors including OpenAI, xAI, Tesla, Apple, and Google, including some of the biggest names in the industry.OpenAI CEO Sam Altman said Meta offered some of his employees as much as $100 million, and Wired’s reports also indicate the same that Zuckerberg is offering $100 million for the first year, with over $300 million over four years. Although Altman also claimed he was sure no one would leave, since they were dedicated to OpenAI’s mission, he turned out to be wrong. Last week, Meta hired ChatGPT co-creator Shengjia Zhao as the lab’s chief AI scientist, marking a significant coup in Silicon Valley’s intensifying talent war.Despite the extraordinary compensation packages, gains have been limited: only a fraction of high-profile researchers opted to join. Meta also explored acquiring two startups founded by former OpenAI executives, Safe Superintelligence and Thinking Machines Lab, but was rebuffed in both cases. Then, acording to Wired, Meta approached more than a dozen staffers at Mira Murati’s AI startup to discuss joining the lab. One researcher reportedly received an offer of more than $1 billion, but none accepted the offer, at least as of this writing.Zuckerberg’s spending spree hasn’t been an isolated event. Microsoft poached two dozen Google DeepMind AI researchers, including some working on the company’s Gemini model. Meanwhile, Google, close a $2.4 billion deal with WindSurf, which has been on OpenAI’s radar, nvested $2.4 billion into WindSurf, acquiring its CEO, key AI talent, and its core tech, just as Meta did with Scale AI.The recruiting frenzy has kicked off a talent race within the Silicon Valley, and while some believe this AI talent war benefits researchers, DeepMind CEO Demis Hassabis called it a rational move for Meta, saying the company is merely trying to catch up. Others, like Sam Altman, are less charitable, he has compared Meta’s tactics to those of the mafia.Zuckerberg’s ambitions go beyond catching up. His goal is Superintelligence: AI that surpasses human intelligence across all domains. This includes integrating advanced AI into Meta’s social platforms to offer digital companionship, generate unlimited personalized content, and enhance targeted advertising.The CEO goal believes assembling an elite team is essential to maintain Meta’s position as a top social and advertising platform, especially as rivals like Elon Musk’s X are absorbed into AI-first companies. “You actually kind of want the smallest group of people who can fit the whole thing in their head. So there’s just an absolute premium for the best and most talented people,” he told The Information.

Meta faces investor scrutiny over massive AI spending

When Meta reports second-quarter earnings Wednesday, analysts expect the company’s slowest revenue growth in seven quarters at 14.7% to $44.80 billion, with profit growth slowing to 11.5% as operating costs jump nearly 9%, according to Reuters data from LSEG. BNP Paribas estimates the Superintelligence hiring spree could drive $1.5 billion to $3.5 billion in additional annual research spending, with Meta already raising its capital expenditure forecast by 10% to between $64 billion and $72 billion for 2025.Despite the massive investments and internal tensions as existing AI staff worry about being sidelined, Meta’s stock has risen 20% this year as investors show more patience for AI spending compared to the metaverse push that drew sharp criticism in 2022. However, executives are now questioning Meta’s open-source AI strategy, with some considering abandoning the planned release of Llama 4’s “Behemoth” version in favour of developing proprietary models.Zuckerberg remains confident the bet will pay off, telling The Information that Meta’s “very strong business model that throws off a lot of capital” can support hundreds of billions in infrastructure and talent investments. The challenge now is convincing Wall Street that this secretive Manhattan Project will deliver the AI breakthrough Meta desperately needs.