Meta's Billion Users Are Playing a Different Game

September 16, 2025

Meta and its AI products are often an afterthought among the big AI companies. When talking to friends in SF, DC, London, about what chatbots they pay $20 for, the fierce debate is between ChatGPT and Claude, with a verbal minority for Gemini, and dissidents propagandizing Grok and DeepSeek. The only time my American friends have ever used any Meta AI products is on flights with free Whatsapp messaging.

Despite all of this, Meta almost certainly has the most users of any of the LLM companies, and their models are almost certainly the most popular LLMs in the world. In this post, I’m going to write about who these users are and how Meta’s reach affects its goals. I also share some speculative thoughts on what a world where Meta’s user-dominance continues looks like.

Meta probably has the most users interacting with its LLMs

None of the AI companies are public about their users, but we can make some guesses. OpenAI probably holds second place: in August, ChatGPT was on track to reach 700 million weekly active users. Google is harder to keep track of: they cheat by counting users of the AI Overview feature which shows up for every search. The Gemini suite has probably around 500 million monthly users. Anthropic, DeepSeek, and xAI are all under 100 million active users.

Meta has well over one billion monthly users. They broke that figure in the spring, and have had a good year since then.Some necessary caveats: OpenAI reports weekly, not monthly like the others. Also, it’s unclear if OpenAI and Google include enterprise clients – they used to, but this is all rather opaque. So where’s all this usage coming from?

There’s no puzzle here: Meta AI’s users mostly use it through Whatsapp’s AI channel, and the vast majority of the rest are through Facebook, Facebook Messenger, or Instagram. Although there is a Meta AI Assistant webpage, how often do you visit it? Users of Meta AI are not like the users of other LLMs. They’re mostly not early-adopters, mostly not tech-literate, mostly on smartphones. The majority live in poor countries. Last year, India became the largest user for Meta AI, and though they didn’t disclose specific numbers, India is really big.Could a Chinese LLM have higher usage than Meta’s products? I don’t think so, but it seems plausible. China has 300 million more smartphone users than India. Higher AI use in China would be through integration into one of the super-apps. I had assumed this had already happened; I don’t know why it hasn’t. But more importantly, most Chinese AI companies aren’t seeking international usage like Meta is, and don’t have the reach onto platforms outside of China. Meta AI’s users are the poorest of any of the LLM companies’.I was going to caveat this with “probably has…” but it’s true by definition: the majority of users of ChatGPT and Gemini are in rich countries; the majority of users for Meta are in poor countries. It is probably also true that Meta’s US-based users are poorer than the other companies, but this is harder to know.

But the most important difference between users of Meta AI and the other LLMs is how they use it. When Anthropic released data on how Claude is used, they classified only 23% of conversations as “non-work”. When asked for examples of what people used Meta AI for, Mark Zuckerberg lists social topics:
> “I want to talk through an issue”, “I need to have a hard conversation with someone”, “I’m having an issue with my girlfriend”, “I need to have a hard conversation with my boss at work”, “Help me role-play this”, or, “Help me figure out how I want to approach this”.

So while users of ChatGPT, Claude, Gemini are automating their jobs, users of Meta AI are mostly just talking to a buddy.

What can Meta do with this advantage?

Meta AI has over a billion users not because it’s the best AI product, but because of its network and integration effects. They have other clear advantages over their competition: They have far more chat data, both between users and between a user and the AI. They have far more data in more languages than their competition. They have a strong existing social network and integration ability [JL]. The only other AI company which can easily integrate AI into a social network is X – and they have 1/20th the users. The only other AI company which has non-AI products with billions of daily users is Google – and they don’t have a social network.Google+ RIP.

These advantages are going to make Meta’s AI very good at what Meta wants it to do, which is a) provide people with social connection, and b) use that social connection to sell ads. In a May interview, Zuckerberg categorized Meta’s four AI opportunities: improving ad targeting, making “the timeline” more engaging, increasing business marketing over chat, and generating direct AI revenue. To him, an AI that’s good at these four things has to integrate well into Meta’s existing ecosystem, alongside your Facebook friends, Whatsapp contacts, Instagram followers. [JL see steerability quote]

Having the most data on how friends chat will be helpful towards that goal. Beyond historical data collected by Whatsapp, Facebook, and Instagram, Meta’s massive user base provides live data for training and large populations for iterating and testing. Similarly, training a model to be good in a new language requires a lot of data in that language.

Meta can also integrate their AI characters into their existing social ecosystems subtly and convincingly. This doesn’t mean they’ll hide that they’re AI; right now, when I wanted to go chat with a Character on Messenger, I had to go under the “AI Characters” dialog. But once the conversation started, it moved into my normal list of chats with my cousin, high school friends, baseball team, and a dude I’m trying to buy an armoire from on Marketplace. Their AI tools are supposed to act just like anyone else we’d interact with on Meta’s platforms, and are placed in the same UIs as them.

Meta AI’s advantages, especially its massive headstart in terms of number of users, are going to help it towards its goals. But these goals are very different from what its competitors aim for. Meta isn’t trying to reinvent the future of work. Zuckerberg, in another spring interview:
> Imagine you were sitting at the beginning of the development of the internet and you asked, “What’s going to be the main internet thing? Is it going to be knowledge work or massive consumer apps?” You got both. You don’t have to choose one. The world is big and complicated. Does one company build all of that stuff? Normally the answer is no.

Meta’s approach has changed since these two interviews: both were before Meta was offering nine-figure pay packages to individual AI researchers, and likely before they started building data centers in tents. But Zuckerberg has reiterated this vision, in a word-salady August public letter:

> [Our vision] is distinct from others in the industry who believe superintelligence should be directed centrally towards automating all valuable work, and then humanity will live on a dole of its output. At Meta, we believe that people pursuing their individual aspirations is how we have always made progress expanding prosperity, science, health, and culture. This will be increasingly important in the future as well.

Anyway – having the most users doesn’t help much for automating the future of work. Knowing more obscure languages isn’t that economically useful – 5 languages make up 2/3rds of world GDP. Unlimited data of friends chatting isn’t going to be helpful in scaling training; most conversations are pretty boring, and don’t contain novel information.The last thing my most active group chat talked about was which US president had the coolest middle name (Warren G. Harding). How many versions of “How are you today?” do we really need in the training data?

Put together, Meta AI’s billion+ users may be very useful when combined with Meta’s platforms to achieve Meta’s goals. But these users are regular people talking about regular stuff, using AI to get through their lives. Even if the new Meta Superintelligence Lab is aiming for some more general AI product, or if they turn their attention toward competing with the other labs for economically productive AI, their user advantage won’t come in handy towards that.

So far, I haven’t mentioned Llama or any specific Meta AI model. First, they kind of suck. Second, Meta Superintelligence might soon partner with their competitor AI companies to incorporate their models onto the Meta AI platforms. Third, the advantages and challenges above apply whether they’re training a new model or adapting someone else’s model.

Llama 4’s was unexpectedly dominant on LM Arena, a platform which ranks LLMs based on human preference in head-to-head matchups. Zuckerberg puts this off as a fluke,“I guess there was a team that built, steered a version of it to be really good at LMArena and it was able to do that, because it’s steerable.” but it’s a good illustration of how Meta’s advantages make the model fundamentals relatively unimportant. Meta wants to be everyone’s favorite – it’s not a mistake that that’s how they ended up.

The paths might diverge

Machine learning is really good at generating engagement; who hasn’t scrolled longer than they wanted? Meta’s goal was always going to be increasing the attractiveness and engagement of their services. That’s what Meta AI’s goal will be: keeping their users scrolling, engaging with their products, and chatting directly to their business partners. All four of the opportunities Zuckerberg sees in AI are best served by a super-engaging AI product.

This is not what the other AI companies want. OpenAI, Anthropic, and Google are all aiming to integrate their tools into the entire economy, up to automating all knowledge work. Even if they fall short of automating 40% of global GDP, their AI tools are being integrated in business processes across rich countries.

Meta and its competitors are all AI companies, all using the same tools. But Meta is throwing these tools at a very different goal from their competitors, and if their paths diverge enough then they won’t be competitors any more. With caveats, what worries me is that the world will end up with two AI ecosystems: one maximizing eyeballs and one maximizing economic productivity.

If the world splits into two AI ecosystems, the “productive” AI will be biased towards rich economies, and the “eyeballs” AI will be biased towards poor economies. Rich countries, cities, and industries are the places where wages are higher, where the return to capital is higher, and where more workers are behind keyboards. All of these fundamentals are likely to make automation more attractive in rich countries than poor countries.

And where wages are low, there’ll be less drive to automate work. These are also the places where Meta is already dominant. Whatsapp is the most popular messaging app in the world because of its users in South Asia and Africa; Facebook is synonymous with the internet in parts of Southeast Asia.In a 2018 report, Facebook’s human rights impact assessors wrote: “For the majority of Myanmar’s 20 million internet-connected citizens, Facebook is the internet.” And strong AI tools will increase the lock-in power of social platforms. The same dynamics may also be present within countries. If AI increases unemployment or inequality in the US, unemployed Americans may spend their time with AIs optimized for ads, engagement, and friendliness, while the best AI work tools are limited to workplaces.

Mass automation could be the end of the great convergence. The economic integration of AI into rich economies will be traumatic, but increased productivity will increase inequality. Trapping poorer economies into an AI ecosystem less tailored for these outcomes will limit the effectiveness of automation even when it does occur, and keep growth relatively low.

But more importantly, there’s still so much we don’t know about how AI will supercharge social media, about AI-driven ads, and about AI relationships. Right now, Meta is rolling all of these out, and much faster in poorer countries. Exposure to social media is bad for you; exposure to economic growth is good for you. That’s what the choice might be.