Meta has been involved in a deep career in artificial intelligence for years while the public conversation is shared by ChatGPT, Gemini and Claude. The company invests tens of billions of dollars in Llama infrastructure and models and yet many users still do not associate “AI” with Meta in their daily lives. That gap between spending and perception is exactly what begins to bother more than one investor.

Meta and its million-dollar bet on AI

In 2024, Meta raised its capital spending to a range of between $37 billion and $40 billion, with a focus on data centers, servers and GPUs to train generative AI models. Added to this is a broader transition in which the company spent nearly $28 billion to completely reorient itself around AI.

The plans for 2026 are even more aggressive and speak of a total expenditure of at least $115 billion, compared to about $72 billion the previous year, with the idea of ​​competing head-on with Google and other giants in the sector. In other words, Meta is betting historic amounts of money just to not be left out of the AI ​​conversation.

If we go down to the level of the models, Llama 3.1 of 405 billion parameters was trained with more than 16 thousand Nvidia H100 GPUs, chips that can cost between $25,000 and $40,000 each. That combination implies a potential investment of hundreds of millions of dollars in hardware alone for a model that almost no one identifies by name outside of the developer world.

The metaverse memory doesn't help either. Reality Labs, the division in charge of virtual reality and the famous metaverse, has accumulated operating losses close to $70 billion in recent years, a figure that continues to grow quarter by quarter. That's why many analysts look at current spending on AI and wonder if we're looking at another mountain of money that may never be recovered.

Llama moves in the shadow of ChatGPT and Gemini

Meanwhile, others take center stage. ChatGPT leads with hundreds of millions of monthly users and a clear dominance in usage and awareness, both in consumer and work environments. Gemini relies on Google's brutal distribution and already has hundreds of millions of users thanks to its integration into Android and the company's key services.

Claude has become the favorite choice for many advanced profiles, especially for programming, analysis and high-level tasks, although its user base remains smaller. In this map of attendees, Meta AI appears as the silent guest that lives within Facebook, Instagram and WhatsApp, but is rarely the first choice when someone thinks about “opening an AI chatbot.”

In the business field, the photo does not help Meta's ego either. A report on spending on language model APIs puts Anthropic at 32 percent of corporate usage, OpenAI at 25 percent, and Google at 20 percent. Meta and Llama are left with 9 percent of the market, far behind despite the rain of money destined for infrastructure.

The curious thing is that, in parallel, Llama models have been downloaded about 350 million times on platforms such as Hugging Face and have been adopted by companies such as Shopify, Zoom or Goldman Sachs for internal projects. That is, Llama is used, but its presence is discreet and fragmented, far from the massive branding effect that ChatGPT or Gemini have.

Lost money or future play for Meta?

This is where the discussion becomes more interesting. In its calls with investors, Meta admits that its generative AI products are not yet generating relevant revenue and that the economic impact will come later, if all goes well. For the market, that phrase sounds too similar to the promise of the metaverse, and feeds the idea that the company is spending today what it may recover in many years.

However, the strategy has another reading. By opening up Llama and making it a de facto standard within open weight models, Meta ensures something that the metaverse never really had a real base of developers and companies building on top of its technology. In that quieter layer, the company is achieving relevance.

What we see from the outside seems like déjà vu of the metaverse with a new name. What is at stake within the industry is who controls the infrastructure and models that will sustain the next AI-powered internet. Whether Meta's numbers will end up in the "lost money" or "visionary bet" chapter will depend on something that does not yet exist, concrete products that can transform this wild spending into experiences that people really want to use every day.