From scrapping “Behemoth” to flirting with closed-source, Meta’s AI overhaul raises questions about costs, talent, and credibility.

In the high-stakes race for artificial intelligence dominance, Meta has once again redrawn its internal map. In a confidential announcement that quickly leaked into industry chatter, Mark Zuckerberg has divided Meta Superintelligence Labs into four separate units. The move is being pitched internally as a strategic realignment to accelerate the march toward "frontier models"—AI systems that can compete with, or surpass, offerings from rivals like OpenAI, Google DeepMind, and Anthropic. But beneath the glossy framing lies a more complicated reality: high turnover, mounting costs, and a company still struggling to define its place in the AI era.

The New Structure: A Bet on Specialization

The reorganization splits Meta’s AI effort into four pillars:

• FAIR (Fundamental AI Research): The long-standing research arm, with a history of publishing influential papers and pioneering open-source models like LLaMA.
• TBD Lab: A new division led by Alexandr Wang, recently elevated to chief AI officer. This lab is reportedly tasked with pursuing "moonshot" AI ideas that remain undefined but are expected to push beyond conventional architectures.
• Products Unit: Headed by ex-GitHub CEO Nat Friedman and venture capitalist Daniel Gross, this squad will focus on embedding AI into consumer-facing products.
• Infrastructure Crew: Overseen by VP Aparna Ramani, this group faces the herculean task of scaling compute and managing Meta’s expanding GPU clusters.

On paper, the model is elegant: research, experimentation, application, and infrastructure—each silo empowered to drive excellence within its domain. In practice, however, execution may prove thornier.

A Company Haunted by Its Own Behemoth

The most telling signal in this reshuffle is what Meta is leaving behind. The company is scrapping its troubled "Behemoth" model, an ambitious but faltering project intended to rival GPT-4 and Gemini Ultra. Despite billions poured into training and infrastructure, the model failed to meet internal benchmarks and created friction between research and product teams. The decision to abandon Behemoth and start afresh is both a tacit admission of missteps and a daring gamble to pivot before it’s too late.

Even more controversial is Meta’s quiet retreat from its open-source stance. For years, FAIR was the loudest evangelist for open AI research, releasing powerful models like LLaMA under permissive licenses that fueled an ecosystem of independent developers. Now, internal memos suggest the company is considering closed-source models, a move that would align Meta with more guarded competitors but could alienate the developer community it once courted.

Billions Spent, But at What Cost?

The stakes could not be higher. Meta has already poured more than $72 billion into AI capital expenditures. Yet the returns remain murky. Unlike rivals that can point to commercial successes—ChatGPT subscriptions for OpenAI, enterprise AI services for Microsoft, or Google’s integration of Gemini into its ecosystem—Meta’s AI footprint is largely confined to recommendation engines and ad targeting.

Yes, AI powers Reels and Instagram feeds, but those applications do little to position Meta as an AI innovator in the public imagination. Moreover, the relentless spending comes at a time when Wall Street is already scrutinizing Meta’s burn rate on the Metaverse. Investors may tolerate losses if a blockbuster frontier model materializes, but patience is not infinite. If the reboot flounders, even the costly Metaverse push could begin to look like a more disciplined bet by comparison.

Leadership Shifts and Cultural Friction

Adding to the uncertainty is a wave of leadership turnover. High-profile figures, including Joelle Pineau, Angela Fan, and Loredana Crisan, have departed in recent months, raising questions about institutional stability. Meanwhile, new chief AI scientist Shengjia Zhao is reportedly challenging legacy approaches and pushing for more aggressive accountability.

That may spur innovation, but it also risks exacerbating tensions in a culture already strained by reorganizations. This marks Meta’s fourth major AI shakeup in just six months. Industry veterans know that talent retention is as critical as hardware or algorithms. If defections continue, Meta could find itself with more GPUs than qualified researchers to run them.

A Wider Strategic Question

Why is Meta chasing "frontier models" at all? One could argue the company’s competitive advantage lies not in competing head-to-head with OpenAI but in embedding capable AI across its massive social and advertising platforms. Instead, Zuckerberg appears determined to secure a seat at the AI table reserved for companies producing foundational models.

The risk is obvious: Meta could spread itself too thin, trying to master both infrastructure and foundational research while also weaving AI into consumer products. Licensing third-party models is now reportedly under consideration—a tacit acknowledgment that even with tens of billions in spending, Meta cannot buy omniscience. If so, the company may have to balance pragmatism with pride, a challenge not unfamiliar to tech titans but rarely embraced openly.

Another Reset, or the Breakthrough That Sticks?

Zuckerberg’s latest AI reorganization underscores both ambition and vulnerability. The specialized unit structure could, in theory, accelerate progress by giving each team a clear mandate. But the combination of sunk costs, cultural upheaval, and shifting strategic priorities suggests that this reboot is far from a guaranteed success. If Meta finally produces a breakthrough frontier model, the company could redefine its relevance beyond Reels and digital ads. If not, the mounting churn in talent and capital may render this experiment another in a series of costly resets.

 Originally Published on LinkedIn.