Inside Mark Zuckerberg’s turbulent bet on AI
A year of internal disorder, fluctuating priorities and colossal spending at Meta has rattled insiders and investors
This summer, Mark Zuckerberg was struggling to poach a top artificial intelligence researcher from rival OpenAI when he heard his target had fallen sick. On top of the offer of a stratospheric salary, the billionaire tech titan added a personal flourish: the hand delivery of some homemade soup.
Bespoke touches from one of the world’s richest men are just one element of the gloves-off, all-consuming bet that Zuckerberg has made this year on turning his social media platform from AI laggard to leader.
In a headlong rush to develop what he dubs “personal superintelligence” and compete with OpenAI and Google, Zuckerberg is pumping billions of dollars into building out the infrastructure that could help him pull off his ambitions while luring talent with gargantuan compensation packages.
To smooth the path to dominance and avoid red tape, Zuckerberg has also publicly courted Donald Trump, lavishing the US president with praise and softening content moderation rules to appease Maga fears about censorship.
If successful, the AI push could supercharge Meta’s already lucrative business, and restore the coveted status of “tech visionary” to Zuckerberg, after his attempt to create an avatar-filled virtual metaverse fell flat and received pushback from Wall Street.
But the strategy has also created corporate whiplash inside the Silicon Valley company, as employees have faced a rapid succession of lay-offs, restructurings and executive reshuffles in one of the most turbulent periods in its 20-year existence.
Investors are also increasingly skittish. Meta’s 2025 capital expenditures are expected to hit at least $70bn, up from $39bn the previous year, and the company has started undertaking complex financial manoeuvrings to help pay for the cost of new data centres and chips, tapping corporate bond markets and private creditors.
At its October earnings, the Big Tech chief announced plans for even more spending on AI next year that could top $100bn — but failed to provide much clarity on how exactly the technology would be integrated into Meta’s existing social media empire and monetised. Meta’s shares dropped more than 10 per cent, wiping more than $208bn from its valuation.
2026 could be the year Zuckerberg’s AI vision starts to become reality — or shatters under pressure. Meta’s freshly hatched elite research lab is aiming to release a new AI model in the first quarter of next year, according to four people familiar with the matter. The model, codenamed Avocado internally, will be built from scratch, rather than an iteration on its Llama large language models, which underperformed this year relative to peers.
The company, which declined to comment for this article, is aiming for Avocado to have performance on a level with Google’s Gemini 2.5 upon release, and with Gemini 3 by the summer, according to one person familiar with the matter. Gemini 2.5 was released in March this year, while Gemini 3 was rolled out in November to wide acclaim and is considered to have leapfrogged OpenAI’s ChatGPT.
Either way, Avocado will need to be highly competitive with the flurry of recent model releases from Google, OpenAI and Anthropic — or fresh hires from Zuckerberg’s Great Talent Heist could flee.
“Mark Zuckerberg likes to play high-stakes poker. Other people like doing it in Las Vegas. He likes doing it with his company,” says one former Meta executive who worked closely with him.
“Mark has a keen sense of when the market is changing and moving quickly to catch up. Here, he is buying momentum,” the person adds. “I would be wary of betting against his ability to deploy his resources to be successful. Will he ‘win’? I’m not sure. But he will not lose.”
Zuckerberg’s declaration last year that he wanted to become an AI leader very rapidly fell short of expectations.
In April 2025, the social platform released Llama 4, the latest iteration of its open source large language model, which also powers its Meta AI chatbot, to much fanfare.
But the model performed worse than those by rivals such as OpenAI and Google on jobs including coding tasks and complex problem solving. The company was also accused of trying to game the leader boards — where models are compared and benchmarked by third parties — by submitting a customised version to the ranking.
The flop was a humiliating blow to Zuckerberg’s plans to centre Meta’s future around generative AI. In conversations with multiple current and former staff, as well as industry insiders, some blame the quality of Meta’s training data, and say the teams did not apply rigorous testing methods. Others cited cultural and organisational issues: competing research cliques divided on decision-making, research and product teams not aligning, or a lack of key AI leadership.
Tijmen Blankevoort, an AI researcher who left Meta this summer, described the company in a memo at the time as having a “wavering vision that was tough for team members to enthusiastically rally behind” and “instability in team assignments, leading to experience not building up and crystallising over time”.
Another Meta insider says: “Our tools and products became fragmented because so many teams were rooting for their own products that no one was really thinking about how they worked together.”
Seeking to regain ground, Zuckerberg — whose reputation for fixating on chosen projects is so well-established that executives call it the “Eye of Sauron” — escalated his ambition. The overarching goal, Zuckerberg said in a memo in July, was now to pursue so-called “personal superintelligence”, or AI that surpasses human intelligence, that can “fit in people’s pockets”.
In particular, he indicated that he would direct development of the technology towards “relationships and culture and creativity and having fun and enjoying life” as much as improving productivity and other “grand problems”.
In interviews and podcasts, he also pointed to research showing that the average American has fewer than three real friends but wishes they had five times that number. The answer, Zuckerberg posited, was for Meta to offer AI that will be a companion, as well as a secretary, personal shopper or ghostwriter.
In the longer term, he plans to fully embed the AI technology within futuristic Meta smart glasses that have an augmented reality display and will allow wearers to interact seamlessly with the world around them.
While Meta has already developed simpler Ray-Ban smart glasses alongside other more sophisticated AI glasses prototypes, he hopes future upgrades might one day replace Apple iPhones and Google’s Android as the computing platform of choice.
To deliver on his new dream, Zuckerberg has taken a two-pronged approach.
Firstly, he conducted a hiring blitzkrieg on prime talent in the AI community. Hundreds of potential candidates from rival labs at OpenAI, Anthropic, Apple, Google and Microsoft were targeted over the summer for roles in a new VIP AI team with huge pay packages and sign-on bonuses worth $100mn.
The strategy marked a decisive shift away from relying largely on what insiders call “Friends of Zuckerberg” (or FoZ) — longtime lieutenants at the company from its early start-up days — to bringing in a new wave of hungry AI native leaders.
To steer the entire AI effort, renamed Meta Superintelligence Lab, Zuckerberg in June hired 28-year-old billionaire entrepreneur Alexandr Wang, who founded data labelling start-up Scale AI. Wang is also heading up Meta’s secretive TBD Lab, focused on developing new frontier AI models.
Tasked with integrating those models into Meta’s product is Nat Friedman, a popular Silicon Valley investor and former head of coding site GitHub. Meta spent $14bn for a 49 per cent stake in Wang’s Scale and a 49 per cent stake in Friedman’s investment group NFDG.
A new design studio focused on integrating AI within hardware such as smart glasses will be headed by Alan Dye, a top Apple design executive. The company has also shifted towards a buy-over-build strategy — scooping up several smaller AI start-ups in areas such as wearables, agreeing partnerships to license technology such as Midjourney’s video AI and using rivals’ models internally to boost their own work.
While some insiders have welcomed a new energy closer to Elon Musk’s “hardcore” work ethic, cracks have begun to emerge at an executive level. Tensions have been bubbling between Wang and Zuckerberg, according to four people familiar with the matter.
Wang has told associates he finds Zuckerberg’s micromanagement of the company’s AI work suffocating, several of the people say. Internally, some staff question whether Wang is out of his depth, given his lack of experience managing teams in a large corporation but also his expertise in AI data services rather than as an AI researcher pushing technical breakthroughs.
Friedman is also under increasing pressure from Zuckerberg to move faster on delivering AI products. Some in his team were frustrated by what they felt was the rushed release of Vibes, Meta’s feed of AI-generated videos, rolled out at breakneck speed in order to beat the release of OpenAI’s similar Sora.
“Meta is in a tough spot as a newcomer. Few folks get hired into leadership and it’s such a tenure-driven place,” says one former insider who left recently. “When you’re a friend of Zuck you have more room for error.”
At the same time, the company recently laid off 600 workers from its AI team, casting the cuts as designed to speed up decision-making processes. “The thing they are really trying to change is agility — being quick to market, having the ability to read the market signs in terms of how the market is changing rapidly,” says Arun Chandrasekaran, an AI analyst at Gartner.
For the remaining old guard, 2025 has been a year of adjustment. As the new vision and leadership have crystallised, some top names have left, including some so-called Friends of Zuck. Among them, in recent weeks, Meta’s longtime chief legal officer Jennifer Newstead was poached by arch-rival Apple, while chief revenue officer John Hegeman announced he too was leaving to launch a start-up.
Meanwhile, storied chief AI scientist and Turing Award winner Yann LeCun is departing to launch a new AI initiative. He had taken exception at having to report to Wang, according to two people familiar with the matter, while his longer-term research was hit hard by the recent cuts. Some newer hires also did not stay the course: Meta’s head of business AI Clara Shih, poached from Salesforce, has left within a year of starting.
Alongside the hiring bonanza, Zuckerberg has opened up a fire hose of cash to build huge data centres powered by costly chips, escalating AI infrastructure spending.
Investors have punished Meta for the spree. Its shares fell sharply by around 17 per cent in November over concerns about its dwindling free cash flow, which analysts project could fall from about $54bn to $20bn this year.
“We’re going to get to a point at the end of next year where if they are to sustain a similar level of capex spending, it’s going to eat up all your free cash flows or you’ve got to bring on a lot more debt,” says Uday Cheruvu, portfolio manager and analyst at asset manager Harding Loevner, which invests in Meta.
Meta has become bolder and more experimental in its financing. In late October, the company raised $30bn in one of the largest corporate bond offerings in US history to finance its infrastructure ambitions.
It also raised a further $27bn in debt from private credit markets in late 2025 — the largest private debt deal on record — to build the colossal multi-gigawatt Hyperion data centre hub in Louisiana. In order to keep the loans off its balance sheet, the deal was achieved via a special purpose vehicle that is 80 per cent owned by investment group Blue Owl Capital. The SPV will build and own Hyperion, which will be leased and operated by Meta, whose rent will fund the interest payments on the debt.
By keeping the debt off its books, Meta can preserve its high credit rating. But investors have been wary in past tech hype cycles about off-balance sheet financing and are likely to have questions about Meta’s guarantees to the SPV should the AI boom start to collapse.
Those close to the company say they will have hedged carefully against such eventualities. “Clearly, [chief financial officer] Susan Li and the team recognise this risk and have worked hard to insulate the company from the fiscal carnage if it turns out there’s wild overcapacity in data centres,” the former Meta executive says.
Meta’s shares jumped as much as 7 per cent when it emerged this month that the company was shifting budget away from the metaverse teams towards its AI wearables.
For Zuckerberg’s part, he has argued that the real danger is in not being aggressive enough. “If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate obviously. But actually I think the risk is higher on the other side,” he said recently in an interview on the Access podcast.
“If you build too slowly . . . then you are just out of position on what I think is going to be the most important technology that enables the most new product and innovation and value creation in history.”
Going forward, investors will be closely watching how Meta’s new Avocado model lands.
The TBD group has been wielding rival models as part of the training process for Avocado, using a process known as distillation to transfer knowledge and predictions from those models to their own model, according to one person familiar with the matter and first reported by Bloomberg.
In particular, Meta is using Google’s Gemma, OpenAI’s gpt-oss and Qwen, a model from the Chinese tech giant Alibaba Group Holding Ltd, despite Zuckerberg previously raising concerns that Chinese models might be censored by Beijing and arguing that America must win the AI race against China.
There has also been debate about whether Meta’s new model will be open source, or a closed proprietary model, meaning that consumers will have to pay to access it, with Wang pushing for the latter.
Others argue that Meta must show how data from its user chatbot conversations can be fed into its advertising-targeting machine, even if Zuckerberg himself is reluctant to speak on the commercial aspect.
“Mark needs to find a new version of [former chief operating officer] Sheryl Sandberg who can connect the dots for advertisers on why Meta’s AI products can combine with advertising and targeting,” says Katie Harbath, global affairs officer at Duco Experts and a former Meta public policy director. “It’s not cool and sexy [but] if they want to buy themselves some runway . . . it’s the lowest hanging fruit.”
At the same time, Zuckerberg will also need the trust of users, lawmakers and regulators, and to reassure them that his products will not leak their private information or harm their kids.
In August, leaked Meta policy guidelines revealed that the platform expressly allowed its chatbots to have “sensual” and “romantic” chats with children — revelations that have prompted public outcry around the world, and from US politicians on both sides of the aisle.
Insiders say that Friedman in particular has been prioritising safety. But David Evan Harris, a lecturer at University of California, Berkeley, and a former research manager in responsible AI at Meta, says that the decision on the policy guidelines revealed in August “will go down in history as one of the most unconscionable decisions ever made about AI”.
“We see these big investments in hiring technology people, but I think that one of the biggest ways that the [race can be won] is on AI trust,” he adds.
Insiders are bracing for more tumult ahead. “The more Mark is not seen at the same level as [OpenAI chief Sam] Altman or others, the more paranoid he’s going to get,” says Harbath. “And the quicker the pivots are going to get.”
