Over the past three years, Big Tech as a whole has spent hundreds of billions of dollars expanding AI infrastructure. For companies like Microsoft, Google, Amazon, and Oracle, the logic is straightforward. They are cloud providers. They build enormous data centers, fill them with NVIDIA GPUs, and rent that computing power to customers ranging from startups to governments. The capital expenditures are immense, but so is the revenue. Compute is the product. AI is the demand engine.
Meta, however, is different. It does not sell cloud services. It does not rent GPU capacity. It does not charge users for access to its AI models. Yet it is now spending at a pace that rivals or exceeds the companies that do.
In 2021, Meta spent roughly nineteen billion dollars in capital expenditures. By the first nine months of 2025, that figure had ballooned to forty-eight billion dollars, almost entirely driven by the construction of new data centers and the purchase of high-end NVIDIA hardware. Zuckerberg has since announced plans to spend up to six hundred billion dollars over the coming years on AI infrastructure alone—an amount so large that it effectively consumes Meta’s entire operating cash flow, which hovers around one hundred billion dollars annually.
The question that increasingly dominates investor calls, analyst notes, and internal debate is simple: why?
Meta’s AI journey did not begin with hype. Long before large language models captured the public imagination, Facebook was already deeply integrated with machine learning. In 2010, the company deployed facial recognition systems to automatically tag users in photos. Over the following decade, AI became central to content moderation, ad targeting, translation, and feed ranking. These systems were powerful but relatively efficient, requiring talented researchers more than massive compute clusters.
The intellectual foundation for this era was laid by Yann LeCun, the French-American computer scientist widely regarded as one of the pioneers of modern artificial intelligence. In the 1990s, LeCun developed convolutional neural networks at Bell Labs, where his work enabled the U.S. Postal Service to automatically recognize handwritten ZIP codes. After leaving Bell Labs, he became a professor at New York University, continuing his research while shaping generations of AI scientists.
In 2013, Zuckerberg recruited LeCun to lead Facebook AI Research, known internally as FAIR. The mandate was pragmatic: build AI systems that improve Facebook’s core products. For years, FAIR operated largely outside the spotlight, focused on applied research rather than grand philosophical claims about intelligence.
That restraint ended in late 2022.
When OpenAI released ChatGPT, the reaction inside Meta was immediate and intense. The product’s explosive growth made it clear that large language models had shifted AI from an internal tool into a consumer-facing phenomenon. Zuckerberg reportedly instructed Meta AI to develop a competitive model at any cost. Budget constraints were lifted. Speed became the priority.
The result was LLaMA—Large Language Model Meta AI—released in 2023. Unlike OpenAI and Google, Meta chose an open-weight strategy, allowing developers and researchers to download and run the model themselves. At the time, this move was celebrated across the open-source community as a challenge to closed, API-only systems. LLaMA rapidly found its way into third-party products, including the Brave browser’s AI assistant, Leo.
Meta itself integrated LLaMA into its ecosystem, launching a free chatbot at meta.ai, embedding generative features into Instagram, and releasing a standalone Meta AI app filled with endlessly scrolling AI-generated videos. None of it cost users anything. None of it generated meaningful revenue.
Even Meta’s only direct AI-adjacent hardware product—the company’s augmented reality glasses—failed to justify the spending. The Reality Labs division, which houses AR and VR efforts, generated roughly four and a half billion dollars in revenue over two years while incurring operating losses exceeding thirty-six billion dollars. Public demos of AI features frequently malfunctioned, including a widely shared video in which Zuckerberg struggled to answer a WhatsApp video call using the glasses.
Despite these setbacks, Zuckerberg’s ambition only grew. In mid-2025, he publicly articulated his belief that artificial superintelligence was within reach. In a video posted to Instagram, he announced the formation of Meta Super Intelligence Labs and declared, “AI keeps accelerating, and over the past few months, we’ve begun to see glimpses of AI systems improving themselves. So developing super intelligence is now in sight. I believe deeply in building personal super intelligence for everyone.”
The vision was sweeping, but the definition remained vague. What exactly constitutes “personal superintelligence” was never clearly articulated. What was clear was the scale of investment required. Meta began planning data centers of unprecedented size, including a proposed facility in Louisiana reportedly comparable in footprint to Manhattan.
At the same time, Meta escalated its hiring war. According to reporting from Bloomberg and The Wall Street Journal, the company offered signing bonuses as high as one hundred million dollars to lure top researchers away from competitors like OpenAI. Yet the most controversial hire was not a researcher at all.
In June 2025, Meta announced it had paid fourteen point three billion dollars for a forty-nine percent stake in Scale AI, a data-annotation company founded by Alexander Wang. Because Meta stopped short of acquiring a majority stake, the deal avoided antitrust scrutiny. As part of the transaction, Wang agreed to join Meta as its new Chief AI Officer.
The appointment stunned much of the AI research community. Wang, then twenty-eight years old, had dropped out of MIT at nineteen to start Scale AI with backing from Y Combinator, where Sam Altman was president. Scale AI never built foundational AI models. Its business was labeling data—paying hundreds of thousands of contractors a few dollars an hour to tag images and text so models could be trained. After the release of ChatGPT, demand for labeled data exploded, and Scale AI’s revenue surged. In May 2024, it raised one billion dollars at a valuation of thirteen point eight billion.
But data labeling is not research. Wang had no formal background in AI theory, no experience building large language models, and no track record of scientific leadership. Yet he was now responsible for Meta’s entire AI strategy, reporting directly to Zuckerberg.
For Yann LeCun, the situation was untenable.
LeCun had long warned that scaling large language models would not produce human-level intelligence, let alone superintelligence. In multiple podcast appearances throughout 2025, he argued that LLMs fundamentally lack reasoning and world models. On one prominent technology podcast, he stated bluntly, “We are not going to get to human-level AI by just scaling up LLMs. This is just not going to happen. There’s absolutely no way. The idea that we’re going to have a country of geniuses in a data center—that’s complete BS.”
He went further, cautioning investors directly. “If you think that we’re gonna get to human-level AI just by training on more data and with a few tricks, you’re making a mistake. If you invested in a company that told you that, that was probably not a good idea.”
These views clashed directly with Meta’s new direction. Reports from The Information and the Financial Times indicated growing internal frustration following the release of LLaMA 4 in April 2025, which was widely viewed as underwhelming. Worse, Meta was accused of gaming benchmark tests by tailoring specific versions of the model to perform well on certain evaluations. Zuckerberg was reportedly furious.
Shortly after Wang’s arrival, LeCun was sidelined. In November 2025, he announced he would step down from Meta at the end of the year. Days after officially leaving, he gave an interview to the Financial Times in which he addressed his successor with thinly veiled contempt. “He learns fast. He knows what he doesn’t know,” LeCun said of Wang, before adding, “There’s no experience with research or how you practice research, how you do it, or what would be attractive or repulsive to a researcher.”
Then came the line that crystallized the philosophical divide. “I’m sure there’s a lot of people at Meta, including perhaps Alex, who would like me to not tell the world that LLMs basically are a dead end when it comes to superintelligence. But I’m not going to change my mind because some dude thinks I’m wrong. I’m not wrong. My integrity as a scientist cannot allow me to do this.”
For Zuckerberg, however, the allure of brute force was irresistible. Meta generates enormous cash flow. Unlike startups dependent on venture capital, it can afford to pursue a vision unconstrained by near-term profitability. The belief that intelligence will eventually emerge from sheer scale has become the organizing principle of the company.
Whether that belief is visionary or delusional remains an open question. Critics argue that Meta is burning shareholder value on a speculative race with no finish line, betting everything on a theoretical breakthrough that its own former chief scientist insists will never arrive. Supporters counter that transformative technologies often look wasteful until they suddenly aren’t, and that controlling AI infrastructure at planetary scale could prove decisive in ways not yet visible.
What is undeniable is that Meta’s strategy is unprecedented. No other company is spending so much, for so long, with so little direct monetization. Zuckerberg is not merely investing in AI; he is wagering the future of Meta itself on the assumption that intelligence can be bought by the megawatt.
If artificial superintelligence emerges from Meta’s data centers, history will remember this era as visionary. If it does not, the six hundred billion dollar bet may stand as the most expensive miscalculation in the history of the technology industry.
No comments:
Post a Comment