A $14.3 billion bet on one hire, 70-hour workweeks, and a string of delays. That is the current state of Meta's flagship AI effort.
Meta has pushed the release of its next-generation AI model, code-named Avocado, to at least May 2026, according to a New York Times report. The model was originally slated for this month. This is not the first slip - Avocado was initially expected before the end of 2025, then moved to Q1 2026, and now pushed again.
The stated reason: training-related performance testing. Meta wants to make sure Avocado is competitive before shipping it, and right now, it apparently is not there yet. A Meta spokesperson said training efforts are "going according to plan," but internal sources describe a different picture.
The Wang Factor
Much of the pressure traces back to Alexandr Wang, the Scale AI founder who joined Meta as chief AI officer last June in a deal valued at $14.3 billion. Wang now leads Meta Superintelligence Labs (MSL), the elite unit building Avocado. He brought in former GitHub CEO Nat Friedman to run MSL's product arm, and the pair have imposed a "demo, don't memo" culture with 70-hour workweeks that clash with Meta's traditional collaboration style.
The reorganization has been brutal. Chris Cox, Meta's chief product officer of 20 years, was removed from AI oversight after Llama 4's disappointing April 2025 launch. Yann LeCun, the company's high-profile chief AI scientist, left to start his own company. In October, 600 jobs were cut across MSL.
Falling Behind the Competition
The delays matter because Meta's competitors are not waiting. OpenAI's GPT-5 and Google's Gemini 3 have both shipped to solid reviews. Nvidia CEO Jensen Huang recently listed the AI companies running on his hardware - "We run OpenAI. We run Anthropic. We run xAI" - and did not mention Meta's Llama.
That omission stings because Llama was supposed to be Meta's crown jewel. The open-source model family attracted millions of developers, but Llama 4's April 2025 release landed with a shrug. Developers found it underwhelming, and the hype Meta had built around open-source AI leadership evaporated quickly.
The Open-Source Retreat
Avocado represents a fundamental strategic shift. Unlike Llama, it will not be open-source - at least not at launch. Meta's leadership reportedly fears that releasing open model weights (the trained parameters that make a model work) lets competitors like DeepSeek clone and optimize their architecture without paying the research costs.
This is a significant reversal for a company that built its AI reputation on openness. Meta is spending $70-72 billion on AI capital expenditures in 2025 alone, and the calculation appears to be that giving away the results no longer makes strategic sense when the models cost that much to train.
The question for anyone building on Meta's AI stack is straightforward: can this team actually ship? Avocado has no public benchmarks, no release date anyone trusts, and a leadership structure that has churned through senior talent in under a year. Meta has the money and the compute. What it does not yet have is a model that works well enough to release.