Meta is considering closing its AI gates

For the past two years, Meta has played the role of open-source savior, pouring billions into frontier AI and giving the weights away for free.

Now, that entire philosophy might be coming to an abrupt end.

According to a report from The New York Times, Meta’s newly formed superintelligence lab is internally debating a massive strategy shift. Top leaders are actively discussing whether to abandon the company's open-source ethos in favor of building closed, proprietary models.

Last week, a small group of senior members from the new lab held talks about dropping the company’s most powerful open-source model under development, codenamed Behemoth. Among those in the room was Alexandr Wang, the 28-year-old newly appointed chief A.I. officer at Meta.

Trouble with the Behemoth

Meta had already finished the data training process for Behemoth, which was supposed to be its next flagship frontier model. However, sources familiar with the matter revealed that its release was put on hold due to underwhelming internal performance.

After Meta announced the creation of the superintelligence lab last month, teams assigned to Behemoth reportedly halted new test runs on the model altogether.

For years, Meta executives championed open innovation, with Chief AI Scientist Yann LeCun famously predicting that the platform that wins will be the open one. Yet giving away frontier technology came with a sting. Earlier this year, Chinese AI lab DeepSeek built a formidable chatbot partly leveraging Meta’s open-source architecture.

Sources noted that discussions are still preliminary and nothing is finalized without explicit approval from CEO Mark Zuckerberg. Meta could still choose a hybrid approach, keeping some open models while prioritizing closed ones to take on OpenAI, Google, and Anthropic.

Why it matters

If Meta closes the door on open source, it would completely upend the AI landscape. For developers, startups, and researchers, Meta has been the sole trillion-dollar patron funding high-end open weights. Without that lifeline, the open ecosystem risks falling an entire generation behind closed labs.

From Meta’s perspective, continuing to give away models for free makes less and less economic sense. Frontier training now commands staggering compute budgets, and rivals like DeepSeek proved that competitors can take Meta’s expensive foundations and quickly eat their lunch. Shifting toward proprietary models is a defensive move to plug that leak.

However, we think this pivot carries massive risks. Meta’s primary advantage over Google and OpenAI was developer loyalty and community-driven adoption. By closing off its models, Meta steps straight into a head-to-head enterprise and product showdown with OpenAI and Anthropic, where it lacks an obvious distribution moat like Microsoft’s corporate software or Google’s search engine.