Llama 2 is free and I think Meta knows exactly

Meta released Llama 2 with a commercial license. Free to use, free to fine-tune, free to deploy. Available on Hugging Face for anyone to download.

The AI community celebrated. Open source wins again. David beats Goliath. Knowledge wants to be free.

I celebrated too. Then I started thinking about why.

Why give it away?

Llama 2 cost millions of dollars to train. Compute, data, engineering talent, research time. Meta spent real money building this model. And then they gave it away.

Companies don’t do this out of generosity. They do it because free-as-in-beer serves a strategic purpose. The question is: what’s the strategy?

The Android playbook

Google gave away Android. Free operating system. Open source. Any phone manufacturer could use it. Samsung, LG, Huawei, everyone shipped Android phones.

The result: Android captured 72% of the global smartphone market. Google didn’t need to make money on the OS. They made money on the services that ran on the OS. Search, Maps, Play Store, Gmail. The OS was the distribution mechanism. The services were the business.

I think Meta is running the same playbook with Llama 2.

If Llama 2 (and its successors) become the default foundation model that companies build on, Meta doesn’t need to charge for the model. They need the infrastructure. Fine-tuned versions of Llama running in enterprises, in startups, in applications. All of them creating a dependency on Meta’s architecture, Meta’s tokenizer, Meta’s training approach.

When the next version of Llama comes out, the switching cost is lower if you’re already on Llama. Your fine-tuning data, your prompts, your evaluation pipelines, all tuned for Llama’s behavior. Migrating to a different architecture costs time and money.

Free creates lock-in. It just takes longer.

Why this matters

OpenAI charges for access. Anthropic charges for access. Google charges for access. These companies fund AI development through revenue from the models themselves.

Meta funds AI development through advertising revenue. They don’t need Llama to make money. They need Llama to prevent OpenAI or Google from controlling the AI infrastructure. If the default AI stack is OpenAI’s, Meta loses influence. If the default AI stack is Llama-based, Meta stays relevant even without a chat product.

It’s defensive strategy disguised as generosity. And it might work.

The open source community gets a powerful free model. Startups get an alternative to expensive API calls. Researchers get a base for experimentation. Everyone wins. But Meta wins the most, because ubiquity is the strongest form of power in technology.

I’m glad Llama 2 is free. I’ll use it. I just want to be clear-eyed about why it’s free. The gift is real. The motive is strategic. Both things are true.

Related thinking: