Meta splits its AI division into two

Meta (Nasdaq:META) is restructuring its AI division into two distinct units, AI Products and AGI Foundations, marking its most significant internal overhaul as it races to compete with OpenAI and Google.

The move, detailed in an internal memo from Chief Product Officer Chris Cox and reported by Axios, appoints Connor Hayes to lead AI product integration while Ahmad Al-Dahle and Amir Frenkel will co-direct long-term AGI research.

The restructuring comes amid mounting crises, including the delayed Llama 4 Behemoth model and the departure of key Llama architects to competitors like Mistral AI. This marks Meta’s second major AI reorganization since CEO Mark Zuckerberg’s 2023 attempt to “turbocharge” generative AI efforts, which saw the company fall further behind rivals despite early promise.

“Structural changes alone won’t solve Meta’s AI challenges,” said Amandeep Singh, practice director at QKS Group. “While the new AGI Foundations unit creates clarity, retaining elite talent requires a seamless pipeline from research to real-world deployment. Meta has struggled with fragmented pipelines and unclear priorities,” added Singh.

Talent exodus and technical setbacks

The restructuring follows talent losses that have exposed fundamental weaknesses in Meta’s AI strategy. Only three authors remain from the original 14-person Llama research team, according to Business Insider. Internal surveys cited by The Information reveal plummeting morale in Meta’s AI division, where employees cite resource constraints and sluggish progress.

“Talent follows momentum, and right now, momentum lives where research decisions directly shape deployed capabilities,” noted Singh.

This talent drain coincides with technical setbacks, most notably the underperforming Llama 4 model, which has struggled with reasoning and mathematical tasks. These combined challenges have left Meta playing catch-up in the race toward artificial general intelligence, despite its early open-source advantages.

The company’s strategy, bolstered by initiatives such as Llama for Startups and the recent Llama API launch, aims to attract developers and differentiate it from competitors’ proprietary models. But analysts caution these initiatives alone may not be enough to win enterprise trust.

The enterprise adoption dilemma

While Llama’s cost advantages remain attractive to businesses, growing concerns about its governance controls and the looming copyright lawsuit over training data are giving enterprises pause.

“Companies love Llama’s affordability but expanding safety gaps and legal risks are becoming hard to ignore,” Singh said. “For mission-critical applications, many will ultimately choose more reliable, if more expensive, options like GPT or Gemini.”

These cost benefits can lose their appeal when weighed against operational risks. Meta’s reorganization attempts to mitigate these concerns through specialized teams, one deploying generative AI across products, another advancing AGI research, but analysts remain skeptical about whether structural changes alone can solve deeper issues.

Unlike Microsoft’s turnkey OpenAI integration or Google’s enterprise-ready Vertex AI platform, Meta lacks both the sales infrastructure and compliance pedigree for regulated industries. As Singh argued, “Enterprise AI adoption hinges on proven compliance frameworks, operational reliability, and mature support systems. Meta still needs to build that trust at Fortune 500 scale.”

Meta’s race to close the AGI gap

“Meta’s AGI push focused on models with reasoning, multimedia, and voice capabilities aligns with the broader industry trend toward multimodal AI as a catalyst for enterprise transformation,” said Surjyadeb Goswami, research director for AI and Automation at IDC Asia Pacific. He noted that open-source models are critical to enabling cost-effective, transparent, and customizable deployments, especially as organizations deepen their GenAI investments.

For Meta to truly capitalize on this opportunity and succeed in its AGI bid, it must rebuild trust especially for enterprise adoption. Singh highlighted the need for Linux-like community stewardship, OpenAI-level safety protocols, and robust enterprise tooling. “Balancing openness with responsibility is Meta’s real challenge, especially as models approach general-purpose cognitive capability.”

Meta now needs to show this reorganization yields meaningful improvements in model performance, talent retention, and enterprise adoption to validate its new approach. “Meta’s open-source vision is bold, but execution is everything,” Singh concluded.