Analysis
Most AI companies want you to pay them every time you use their models. Meta wants the opposite. It gives its Llama models away for free and bets it will make the money back somewhere else entirely.
That sounds like charity until you look at how Meta actually earns. The company makes almost all of its income from advertising across Facebook, Instagram and WhatsApp. Every developer who builds on free Llama instead of a paid model from OpenAI or Google is one more person locked into Meta's world, and one fewer paying a rival. The free model isn't the product. The ecosystem around it is.
Llama 4, released as an open-weight model in April 2025, is the centrepiece of that bet. For an Australian business deciding which AI stack to build on, the question is whether the free option keeps up with the paid frontier, or whether you get what you pay for. The rest of this piece walks through the economics behind Meta's choice, where the real risks sit, and what it means for the tools your team will end up using.
The Economics of Open AI
Meta's strategy rests on a view of where the AI market is heading. Zuckerberg and his team expect AI models to commoditise the same way operating systems, cloud infrastructure and mobile platforms did. In each of those markets, the winner didn't squeeze the most revenue out of each user. The winner spread adoption as wide as possible and made money on the services next to it.
The pricing gap tells the story. OpenAI charges $5 per million input tokens for GPT-5.5. Google's Gemini 3.5 Flash also carries a per-token fee. Meta charges nothing for Llama 4's weights. That's not generosity, it's a market-share play. A developer who builds on Llama instead of GPT isn't paying OpenAI, isn't deepening their ties to Google's cloud, and isn't feeding a competitor's network effects.
The cost to Meta is real. The company itself has described the training spend as an industry-scale investment, and outside estimates put the figure for Llama 4 in the hundreds of millions, though no firm number has been confirmed. (Earlier Llama 3 training was estimated around $25M, so treat the larger figures with caution.) Add the ongoing research, engineering and community support, and the annual bill runs into tens of millions more. Meta books this as the cost of acquiring customers, the price of building a platform that pays off later through advertising, commerce and enterprise deals.
The Llama Ecosystem Play
Llama 4 isn't only a model. It's the anchor tenant in a growing ecosystem. Meta has put serious money into the surrounding tooling: PyTorch, the dominant AI framework it created, the Llama Stack API for standardised model deployment, and a widening set of enterprise integration tools. Llama models are also broadly available across the major cloud providers, with the company reportedly working with AWS, Azure, Google Cloud and Oracle to offer optimised hosting, though the exact partnership terms aren't all publicly confirmed.
Those distribution channels matter. By making sure Llama runs well everywhere, Meta lowers the barrier to adoption and stops any single cloud provider from cornering the value of an open model. The cloud providers get hosting demand. Meta gets ecosystem growth. The incentives line up in a way that happens to suit Meta's long game.
Meta has also built standing in the AI developer community through conference sponsorships, research grants and open-source work. Its research lab, FAIR, publishes heavily and maintains several of the field's most-used open tools. That goodwill and the talent pipeline it feeds are hard to put a number on, but they're worth something.
The Integration with Meta's Products
The logic gets sharper when you see how Llama plugs into Meta's consumer apps. Meta AI, the company's assistant, runs on Llama and is being built into Facebook, Instagram, WhatsApp and Messenger. Meta AI itself passed roughly a billion monthly active users in 2025, and it sits inside an app family that reaches more than 3 billion people daily, a distribution channel no competitor can touch.
Better Llama means better Meta AI. Better Meta AI means more time spent on Meta's platforms. More time means more ad impressions, and advertising is where Meta earns about 97% of its revenue. The AI spend doesn't have to pay for itself directly. It pays off through better products driving more use of the ad-supported core.
Risks and Challenges
The strategy has weak spots. Open weights mean competitors can build on Llama without giving anything back. Chinese labs have already used earlier Llama versions as the base for their own models, and the restrictive clauses in Llama 4's licence have drawn fire from parts of the open-source community.
There's also the question of whether open weights can stay close to closed research. The most capable models on the market, Claude Fable 5 before its suspension, Opus 4.8 and GPT-5.5, are all closed-source. If the frontier of capability stays behind paywalled APIs, Meta's open platform could get boxed into commodity work while the high-value use cases flow to proprietary models.