February 2023: Meta releases LLaMA model weights to academic researchers under a restricted license. Within a week, they leak to 4chan. By the weekend, the entire internet has the model.
Meta could have locked it down. Filed takedowns. Sued. Instead, Zuckerberg went the other direction.
July 2024: Llama 3.1 405B — a 405-billion parameter model that matches GPT-4 on most benchmarks — is released for free download, commercial use permitted. You can run it on your own infrastructure, fine-tune it on your own data, and deploy it without paying OpenAI a cent.
Why would Meta give this away?
The logic is cleaner than it looks:
Meta doesn’t sell AI. Meta sells attention and advertising. Their revenue doesn’t come from charging developers for API calls. It comes from keeping 3.2 billion people engaged on their platforms.
For Meta, open-sourcing AI is a competitive weapon aimed directly at OpenAI and Google. Every developer who builds on Llama is a developer not deepening their dependency on GPT-4. Every company that runs Llama in-house is a company that didn’t build a procurement relationship with Microsoft.
Destroying your competitor’s pricing power costs you nothing when you weren’t charging in the first place.
The Llama ecosystem now includes: Groq (the fastest inference hardware on the market, built specifically to run Llama efficiently), Perplexity (which uses open models for retrieval), hundreds of enterprise deployments, and the majority of open-source AI research happening globally.
By some estimates, Llama is now running on more total compute than GPT-4 — it just isn’t running on OpenAI’s compute.
This is the classic platform play. Make the infrastructure free, make money on what sits on top. Facebook did this with React. Google did this with Android. Both decisions initially looked strange. Both turned out to be defining strategic bets.
Zuckerberg’s AI bet isn’t “build the smartest model.” It’s “become the OS that AI runs on.”
If AI capabilities become a commodity — and DeepSeek suggests they might — then the biggest social network in the world, running on the most widely-deployed open model, has a very good position.
It’s starting to look like it was the right call.