Open source was always a phase

Meta built the largest open-source AI ecosystem in history, mass-produced developer loyalty with the Llama family, and had Mark Zuckerberg publish a 2,000-word manifesto arguing that openness was a moral and strategic imperative. Then it shipped its best model behind a locked door, and the stock went up 9%.

The disconnect between what Meta said and what Meta did this week is worth sitting with, because it tells you something about the economics of open-source AI that the manifestos never did.

TechCrunch reported that Muse Spark, the first model from Meta's $14 billion Superintelligence Labs under Alexandr Wang, is fully proprietary. No downloadable weights. No community fine-tuning. Access comes through Meta's AI portal or an invite-only API. The model was built with over an order of magnitude less compute than Llama 4 Maverick and now powers features across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses, including a new Contemplating mode for multi-agent reasoning.

The developer reaction was predictable. The Register captured the cynicism well: the same company that positioned open-source AI as "the path forward" has closed the gates after the ecosystem was built. Meta says it "hopes to open-source future versions," which is the corporate equivalent of saying you'll definitely call. Developers who structured their workflows around the assumption that Llama's successors would remain open are now staring at an API waitlist.

But here's the part that should worry open-source advocates more than the model itself: the market loved it. Meta's shares surged nearly 9% on the announcement, the biggest single-day rally since January, adding tens of billions in market capitalisation. BofA Securities and William Blair both reiterated positive ratings. Wall Street didn't punish Meta for abandoning its open-source identity. It rewarded the company for finally acting like it wanted to capture the value it was creating.

The economics were always pointing here

The way I see it, Meta's open-source play was a market-making strategy, not a permanent commitment. When you're behind in AI, you open-source to commoditise the layer above you, build an ecosystem that routes developer attention your way, and make it harder for competitors to charge for the thing you're giving away. Llama did all of that brilliantly. It attracted researchers, startups, and enterprises who built on Meta's weights instead of paying OpenAI or Google.

But once you've built the ecosystem and your models are competitive at the frontier, the incentives flip. Open weights become a subsidy to your competitors. Every company fine-tuning Llama for free is a company not paying Meta for access. The strategic logic of openness depended on Meta being behind. The moment it caught up, the logic evaporated.

This isn't a story about Meta being dishonest. It's a story about open source as a go-to-market strategy rather than a durable commitment. And it won't be the last time it happens. Every major tech company currently publishing open weights is doing the same calculation Meta just resolved: is the ecosystem value still worth more than the revenue we're leaving on the table? As models get more expensive to train and more capable at generating direct revenue, the answer will keep tipping towards closed.

If you're building on open-weight models from any of the big labs, the question to ask isn't whether the current version will stay available. It's whether the next version will. Meta just showed you what happens when the answer changes.

---